<?xml version="1.0" encoding="UTF-8"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Ersin&#39;s Blog</title><description>I’m Ersin, a software engineer with 10+ years of experience across startups, product companies, and consultancies.&#xA;&#xA;I write about building software well: system design, developer tooling, AI-assisted development, and the experiments and projects I’m working on.&#xA;&#xA;I’m especially interested in how small teams can build ambitious software with less complexity and more leverage.&#xA;</description><atom:link href="https://www.ersin.nz/feed.xml" rel="self" type="application/rss+xml"></atom:link><item><title>Travel Advisory: 2026 as an engineering lead</title><link>https://www.ersin.nz/articles/travel-advisory-2026</link><description>A field report on AI adoption across two engineering organisations in 2026: shipping faster, supporting teams through change, and making time to think.</description><pubDate>25 Oct 2026 11:00 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;Kia Ora, here’s the travel advisory for anyone looking back to 2026 and looking for a slice of what life was like as an engineering leader across a big org and a bootstrapped but well established product.&lt;/p&gt;&#xA;&lt;p&gt;My position is unique in that I’ve had the pleasure of being across two organisations in 2026. First Education Perfect (EP), a 60ish person product and engineering team and a ~300 person org. Second, myWorksites (MWS) a b2g SAAS company with 10 people and bootstrapped within a services company of ~25. Both products are web applications. Each of these orgs are profitable, and have significant legacy systems to contend with.&lt;/p&gt;&#xA;&lt;p&gt;On both orgs I have been a team leader with 3-4 people reporting to me, most recently I shifted from ‘engineering lead’ to ‘product &amp;amp; engineering lead’. I swapped over from EP to MWS June 2026.&lt;/p&gt;&#xA;&lt;p&gt;The unique perspective I can bring is about the comparison of AI adoption and some of the challenges that appear to be universal across teams..&lt;/p&gt;&#xA;&lt;h2&gt;AI has brought me back to shipping production code&lt;/h2&gt;&#xA;&lt;p&gt;The IC &amp;lt;-&amp;gt; EM pendulum is something I have always vibed with, and I’ve timed my swing back to individual contributor focused roles around the acceleration of AI coding tools in 2024. If it wasn’t for the dramatic change in the job, I don’t know if I would have had the interest to stick around in a pure development role. I was considering a move into pure management. Not that coding was boring, but for building web applications I felt like my biggest impact was to be running teams and growing peoples capability as individual contributors. But coding is changing so much, and there is so much interesting stuff to learn, that it’s renewed my interest in staying in development. Now with so many new skills to learn, it feels like an important time to have hands-on experience with AI. If and when I go back to a pure management position, it will help me build that team empathy and understanding of what the actual job requires.&lt;/p&gt;&#xA;&lt;p&gt;The progression through tooling has been wild over the last two years. In 2024 I spent most of my time contributing within a tab completion or integrated chat environment like windsurf provides. Towards the end of 2025 I started experimenting with Claude Code, mostly staying in the inner loop and checking each change to keep the agent on the right path. By February I was deeply experimenting with Ralph loops, but not having much success. As of May 2026 I have been doing 95% of coding outside an IDE without a human in the loop. My time is spent talking with customers, testing, reviewing, design, grilling sessions and supporting my team.&lt;/p&gt;&#xA;&lt;h2&gt;Adoption friction is a thing&lt;/h2&gt;&#xA;&lt;p&gt;Even the most enabled large corporations will still run into policy which challenges AI adoption. EP would get access to new models within days, through the github copilot program and a google workspace subscription there was a reasonable breadth of models available to try and experiment with. Even though leadership would absolutely remove the blocks, there are just more steps to remove them in a larger org. Consider the impacts of ISO27001 and needing to change your risks and mitigations listed. In order to adopt a new practice like dropping explicit acceptance criteria and manual test recording, you must collaborate with the CISO the CTO, principle engineers and whoever happens to owns quality. If you’re unlucky this might end up at a board level decision and you are waiting a month or more to make a change.&lt;/p&gt;&#xA;&lt;p&gt;Another challenge can just be in the IT process. At EP you can’t just run your own AI IDE of choice (windsurf at the time), because it gives you access to a wider set of models, and the security team needs to lock you down to just what is approved. Software that tracked subscriptions would snitch on you if you accidentally signed up for something shadow-it-ish. I get that risks exist, but the friction is real.&lt;/p&gt;&#xA;&lt;p&gt;The best policy for adoption is one which is extremely open to tooling and models that people want to experiment with. Security sucks when you pick up this approach, I get the pain our security friends feel. It sucks. But we must take the balanced risk vs benefit approach. Security exists to enable us to take risks. Care a lot about the mitigations, but don’t worry about a perfect 0 risk solution, because you’re never getting there.&lt;/p&gt;&#xA;&lt;p&gt;Education perfect was not running any official AI training targeted at developers. Much of it was focused on adoption across the rest of the org. Within the engineering and product group AI often overtook the regular guild meetings, slack channels about AI were quite popular and many of our engineering lead chats were about how teams could pick up their AI capabilities faster.&lt;/p&gt;&#xA;&lt;p&gt;MadeCurious, the company wrapped around myWorksites, stopped all staff twice this year (with a third day planned) so far to train on AI through facilitated workshops by Joshua Vial, a trusted coach the team has held for 10+ years. This has definitely had an impact on the culture. People take it quite seriously when you stop them from doing their normal job for a couple days to learn how to use this new tool. When I started in June the adoption was pretty well embedded, 100% of people used claude code with non engineers actively pushing PR’s and features for user facing production releases.&lt;/p&gt;&#xA;&lt;p&gt;What is true everywhere is that adoption is being forced from the top down and the uncomfortable truth is that this is not unanimously received as a good thing. I have seen more than one person burn out of the industry in the last 12 months as AI automates or completely changes their day job. There is much pressure on leaders to be an evangelist and meet new ambitious targets for velocity or quality as a result of having new AI tooling. Individual contributors are needing to broaden and change their skillset. It seems like deep focused coding work has gone by the wayside; although I think deep work and thinking is still very much part of the job.&lt;/p&gt;&#xA;&lt;h2&gt;We got pretty far but we are no dark factory yet&lt;/h2&gt;&#xA;&lt;p&gt;The dark factory is what people call a software development lifecycle that involves no humans and can run all day and night with no lights on. This is not where I am at yet, although I can see us converging on something like that in the future. There is a lot of the job that won’t change though. Engineers, designers and product outcome people will get together and collaborate on designs and what to build. But the actual building might happen 24x7 in the background with minimal human input.&lt;/p&gt;&#xA;&lt;p&gt;In June, there were still people at education perfect that use copilot as a tab completion engine. I would say 75% of engineering was using Claude code.  The most ‘pilled’ of engineers would be just as enabled as anywhere I imagine. Ralph Loops, Workflows, browser use all the skills etc. The overall way of working was only just beginning to change. At EP product and design was not pushing production code yet.&lt;/p&gt;&#xA;&lt;p&gt;Comparatively at MWS a bootstrapped and established business, 100% of people at the org from founder through to designers are using claude code. Everyone has shipped production code on a complex regulatory system in the last month. Being a smaller org has its benefits here, but I would credit the progress on the dedicated training sessions and the leadership by example from the top levels. When everyone is vibe coding you feel like there is a lot of authority to do the same.&lt;/p&gt;&#xA;&lt;p&gt;Our teams are far from leaders, but I think we are slightly ahead of the curve. The hype in the industry right now is insane though and all of the ““If you’re not doing X you’re NGMI” is ridiculous and overblown. The people selling hype on the internet almost always have something to gain from the angle they are selling; you just have to follow the money to understand what is going on.&lt;/p&gt;&#xA;&lt;p&gt;The overall vibe in teams is one where you are constantly playing catch up. This is I think the first time in my career that everyone has felt like a beginner who is just exploring and experimenting and trying the next thing to see what works. As an example, we were heavily experimenting with beads but progressed on within a couple months because the harnesses and models improved to start doing a good enough job of documenting tasks and ticking them off.&lt;/p&gt;&#xA;&lt;h2&gt;Engineers are having a mixed experience&lt;/h2&gt;&#xA;&lt;p&gt;Although my experience sounds like it’s been smooth sailing, I would say the major challenge in 2026 has been in change management within dev teams. The work we are doing looks drastically different and this has proved to be a real challenge for people to adjust to. People have uncertainty that we can keep up with the acceleration in code generated. We ask ourselves, “Are we building ourselves into a corner?”. Every experienced developer shares this experience from being a newbie, they get these alarm bells ringing along with the accelerated pace of development.&lt;/p&gt;&#xA;&lt;p&gt;Being an early adopter is not a sign of capability, everyone is dabbling. Managers who are not on the coal face should not bias how someone is performing purely on how much they are adopting the new tools. AI can increase the amount of work produced, but it tends to be a multiplier for the overall quality of work one was capable of anyway, not a fix-all solution.&lt;/p&gt;&#xA;&lt;p&gt;The people who really seem to have figured AI out the most are the skill stackers. People who are just wading in and experimenting. Seeing someone experimenting in weird ways and failing is a sign that they will probably know more than you in a couple months time. Each model release also comes with a bit of adjustment to understand the capabilities so you need to have these set tasks to throw over to an AI to figure things out.&lt;/p&gt;&#xA;&lt;p&gt;Many engineers are mourning the loss of code as a craft. This hits across all levels of engineering, and particularly it hits the people who have never stepped out of the IC coding role. Managers/Leads have already managed this change in identity and realized that you can still write code by hand for fun, or write code on the side to satisfy your curiosity.&lt;/p&gt;&#xA;&lt;p&gt;There is still a huge amount of skill in not just the using of AI but also in the people and communication side of building something together. Although AI can produce a LOT of documentation, it’s barely fit for human consumption and most of the time you can say the same thing to your team in about 1/3rd the words if you summarize and write by hand. Writing english is the high leverage skill engineers should still be working on in 2026&lt;/p&gt;&#xA;&lt;h2&gt;AI just produces a ton of stuff and that’s overwhelming&lt;/h2&gt;&#xA;&lt;p&gt;This daunting realization that we are shipping so much has also come with a lot of slop-fear and maintainability pains. We constantly talk about it in the team. I would say we are 3 months from not reading a subset of PR’s. But reading code as a skill I think is still important, and I feel like pulling an engineer aside and reading some code together, expecting them to find their way around it is still mandatory.&lt;/p&gt;&#xA;&lt;p&gt;Navigating this comprehension problem is also very interesting when it comes to non engineers contributing to the codebase. We now have a variety of people building on our product that never existed before, product people and account executives are wading right in. Our approach to this for mws is to assign a buddy engineer or two to make sure that we have some understanding of the architectural approach and ensure the quality bar is met.&lt;/p&gt;&#xA;&lt;p&gt;With this increase in code being merged we have also seen a bottleneck in the PR process. In a team of 6 engineers + 3 product people writing code we see a stacking up of 20-30 PR’s in recent weeks. By the time your thing gets in front of someone for review you probably forgot how it works and can’t answer questions as you would hope. Things we have put in place to try and combat this are risk analysis skills, PR raising skills, stacking prs, smaller commits.&lt;/p&gt;&#xA;&lt;p&gt;One of the side effects of AI producing a ton of stuff is you don’t get a good feeling for the level of human effort and understanding behind the artefact. You can end up wasting time reviewing something that is an explicit YOLO project. Our team introduced a label for this as a result. Yoloing is not a goal or anti-goal, there is a time and place for it and we lean into it for prototyping or proving out the shape of a refactor. This is something you would have to really argue about before, but I love that you can go to a refactor discussion with a working proof of concept to show your idea to people now.&lt;/p&gt;&#xA;&lt;p&gt;The code bloat problem is equally met with a ‘documentation bloat’ problem. We share more AI generated artefacts and they feel like they have value at the time of generation. But it can be quite hard to discern exactly the intent by some of these documents. If the author has not fully reviewed them and shared it with you, it can really waste the reader&#39;s time. No different than the AI YOLO case above with code. Respecting the human time by being upfront with your level of self review of the doc first  is a minimum requirement.&lt;/p&gt;&#xA;&lt;p&gt;As our team ships more and engineers take on more responsibility outside of just writing code there are more skills to learn and more systems to be put in place to keep us on track. The systems we are focusing on building are things like permission frameworks / design systems / porting guidelines / risk analysis bots and feature verification agents.&lt;/p&gt;&#xA;&lt;h2&gt;Are we slipping into AI slop miasma?&lt;/h2&gt;&#xA;&lt;p&gt;Overall code quality is far greater than the legacy systems which already exist. Both EP and MWS own some legacy code. What profitable business doesn’t? As we have taken our hands off the writing of code, the amount of testing has increased. To be more concrete I have seen more tests be written and more thorough approaches to testing features pre-release in the last 12 months than ever before. Seemingly everyone is able to AI vibe a really solid senior engineers work out of an agent. Sometimes that direction is slightly off course and needs correction but I would say we are overall making a better quality product while going faster.&lt;/p&gt;&#xA;&lt;p&gt;We tightened up a lot of our CI/CD process in the last year. Projects have stricter linting rules and static analysis checks than ever before. In August I built and re-built a major project 3 times with different combinations of skills on the harness to try and come to a better 20k PR stack to merge. This was a worthwhile effort as it only added a couple days effort and burnt through several peoples claude tokens for the week.&lt;/p&gt;&#xA;&lt;p&gt;Fear of slop is high in teams. Fear of losing skills, not gaining the skills they thought they would be. People genuinely enjoyed writing out boilerplate code and solving problems by hand and they no longer get paid to do that. It’s a real concern for many people still and this was something I have seen everywhere I’ve worked and has come up in all the discussions I have had with peers.&lt;/p&gt;&#xA;&lt;p&gt;The less your team knows about the tech, the faster you go – and I’m yet to run into slop which has slowed us down. On two occasions I have worked with teams who are less strong at building UI/UX. Our largest overall velocity increase has come in that frontend work. Especially when it comes to the boilerplate, stuff that has already been done elsewhere but is 10% different. That stuff can ship extremely quickly. Are we experiencing a bit of gate keeping when it comes to proving our knowledge as engineers is still important? Or are we 6 months out of catastrophic technical debt realisations..&lt;/p&gt;&#xA;&lt;h2&gt;More people are shipping and engineers still end up owning it&lt;/h2&gt;&#xA;&lt;p&gt;Working with Product has changed. My buddies in product can just do way more and I LOVE it. I love that we can give a PM access to a data warehouse and they can come back with a PRD + queries for the metrics which support a feature, metrics for how we can track adoption and get a first signal. At EP I was fortunate to have a PM, who would go really deep on the complex data model. We would collaborate in claude code with the repo, developing queries to test theories about the system and business viability of certain solutions.  Her work &lt;em&gt;REALLY&lt;/em&gt; enabled the team to understand the problem from the businesses perspective as a theory could be accompanied by a real query that backed it up and could be argued about.&lt;/p&gt;&#xA;&lt;p&gt;People writing code is not absolutely accepted and appreciated though. I’ve seen a proliferation of build your own internal tools that have been received with disdain from engineering. People fear supporting work they didn’t create or approve of. Things like browser extensions which are super useful but no one knows how to support that. Custom CRM’s. Vibe Coded ports of legacy systems. They can result in a 20k LOC project landing on a team with little context, the team then needs to sign it off for security and support it forever. Show me a developer that would be stoked with that role!&lt;/p&gt;&#xA;&lt;p&gt;Another example of the product role transforming is that PRD’s are not all you get anymore. Great Product people are building features and prototypes, taking them to customers then coming back to the team with a couple thousand lines of code in the product which shows the idea they are hoping to experiment with.&lt;/p&gt;&#xA;&lt;p&gt;A pattern I have seen when it comes to product builders and engineering working together is that more collaboration is required and there is a significant need for engineering to follow and clean up as sub-optimal ux or architecture is shipped in order to prove a product feature or concept.&lt;/p&gt;&#xA;&lt;h2&gt;Autonomous product teams are still working great&lt;/h2&gt;&#xA;&lt;p&gt;Ways of working have not changed as much as I would have thought. Empowered product teams and a lot of authority for senior engineers to break down work is still the best. For 6+ years I have sought out teams that work this way, so I can’t compare to what a classic SCRUM team looks like in 2026, but I can only imagine that they are not seeing the acceleration that an empowered product team can get. The more you distribute capability for product decision making the faster you can go and the more likely you are to realize the benefit of AI tooling which helps drastically reduce the time to build.&lt;/p&gt;&#xA;&lt;h2&gt;The job is thinking, and that’s fundamentally still what we are doing&lt;/h2&gt;&#xA;&lt;p&gt;Building software is still thought work, even if the writing code part is now 90% automated. As a leader I expect my team to still be able to stand up and explain the decisions and architecture and options which they went through to land on a particular solution. Simply generating a solution does not mean you have done the thinking. We are being paid to do the thinking and build the long term context over years.&lt;/p&gt;&#xA;&lt;p&gt;I’m having more fun building software now than I have in years. I also don’t know exactly where we are going with this acceleration in delivery, and how badly we might regret all our building in 2026. Both of these statements are true and that is just being real about the uncertainty we are facing right now in the industry. We are learning as we go, some of our decisions right now will cost us.&lt;/p&gt;&#xA;&lt;p&gt;There is plenty of work to be done as a leader. Help build the skills your team needs to move outside pure engineering and hand writing code work. Make room to let them test and explore the new skillset that is needed. Focus the team on building a codebase and platform that is ready for non engineers to make significant product features.&lt;/p&gt;&#xA;&lt;p&gt;Shipping faster does not mean thinking less. We still get permission to take time to think through a problem and figure out the best solution .&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>AI</category><category>Engineering Leadership</category><category>Software Engineering</category><category>Agentic Coding</category><guid isPermaLink="true">https://www.ersin.nz/articles/travel-advisory-2026</guid></item><item><title>Why I am building Overpass in Go</title><link>https://www.ersin.nz/articles/overpass-go</link><description>I rewrote Overpass in Go over a dozen evening couch sessions: learning OSM, building a query and storage engine, and testing agent harnesses along the way.</description><pubDate>05 Sep 2026 12:00 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;It is now possible to ingest OSM data and query it over HTTP using Overpass QL from a pure Go service. I built &lt;a href=&#34;https://github.com/ebuckley/overpass-go&#34;&gt;overpass-go&lt;/a&gt; over about a dozen evening couch sessions, vibe coding.&lt;/p&gt;&#xA;&lt;p&gt;This was a hugely valuable learning experience, and I got to really test out the &lt;code&gt;pi&lt;/code&gt; harness along the way. What I built works end to end, and I am happy with the overall result.&lt;/p&gt;&#xA;&lt;p&gt;It is licensed under the AGPL and derived from the original C++ Overpass implementation. This was not a clean-room rewrite. I literally pointed the agents at both and told them to go ham.&lt;/p&gt;&#xA;&lt;h2&gt;Why rewrite?&lt;/h2&gt;&#xA;&lt;p&gt;It was a good way to learn Overpass QL, the OSM data model, and how the query, storage and execution engines actually work.&lt;/p&gt;&#xA;&lt;p&gt;New models and harnesses need to be tested, and this was a good fit for what I&#39;m thinking about. Through the build I used Claude Code, Codex, and a custom &lt;code&gt;pi&lt;/code&gt; harness.&lt;/p&gt;&#xA;&lt;p&gt;I also wanted a codebase that was easier to extend. The C++ project is hard to approach. A fresh Go implementation with high test coverage and purpose-built agent context seemed like a good place to start.&lt;/p&gt;&#xA;&lt;p&gt;This version hosts its own HTTP server using the Go standard library. I want a minimal Docker image that I can run anywhere, and a standalone Go service is a good fit for that.&lt;/p&gt;&#xA;&lt;h2&gt;What started as a faithful port has grown beyond&lt;/h2&gt;&#xA;&lt;p&gt;This project started with just porting the query language parser. I thought that would be the end. It grew to include a minimal set of queries executed in memory, getting the whole thing working end to end.&lt;/p&gt;&#xA;&lt;p&gt;This happened over several evening sessions on the couch. I would poke Claude to continue working on the plan, look at the tests, and tell it to keep going. Progress seemed slow, but we were getting there.&lt;/p&gt;&#xA;&lt;p&gt;Eventually I wanted to make this a more realistic test. I built tooling to query an existing Overpass instance, parse the XML into memory, and run queries against that data instead of relying on synthetic test data.&lt;/p&gt;&#xA;&lt;p&gt;I built a storage index next, an HTTP server after that, and then a process to ingest Geofabrik protobuf files. At this point it was becoming something I could actually host.&lt;/p&gt;&#xA;&lt;p&gt;The storage layer uses bbolt, a fast key-value database written in pure Go. I was being tactical about bringing in a dependency here. bbolt is &#39;complete&#39; software.&lt;/p&gt;&#xA;&lt;p&gt;The C++ storage implementation was clever, but also optimised for memory maps. That isn&#39;t something that translates easily to Go. Putting bbolt behind the common storage interface seemed like a sensible way to go.&lt;/p&gt;&#xA;&lt;p&gt;Eventually I could see this being replaced with a remote key-value database if that fits. For now, we store files locally.&lt;/p&gt;&#xA;&lt;p&gt;The first spatial index used a 0.1-degree grid, purely to make queries work end to end. The C++ implementation uses an ingenious quad tile approach, so I looked at that next.&lt;/p&gt;&#xA;&lt;p&gt;I went down a rabbit hole comparing quad tiles with an S2 index. Ultimately I decided to implement quad tiles first because it could be done in about 300 lines of code. S2 is still something I want to look at for a full-world dataset.&lt;/p&gt;&#xA;&lt;p&gt;At the moment I only query the NZ dataset: tens of millions of points and about 9 GB of memory. A full scan is not the end of the world, but I would like to make this work for whole-earth datasets eventually.&lt;/p&gt;&#xA;&lt;p&gt;On the note of making it scale, I added an on-disk merge sort so ingestion can handle more data than we have memory for. It sorts chunks of about 512 MB at a time.&lt;/p&gt;&#xA;&lt;p&gt;The HTTP server was another place where the port grew its own implementation. I wanted the behaviour to stay as close as possible to the C++ version, but it made little sense to build a CGI app in Go.&lt;/p&gt;&#xA;&lt;p&gt;The standard library already has great support for concurrency and HTTP servers. So we have a worker pool that executes queries and a very standard-looking Go HTTP handler.&lt;/p&gt;&#xA;&lt;h2&gt;The impedance mismatch between Go and C++&lt;/h2&gt;&#xA;&lt;p&gt;One of the biggest differences was error handling. The C++ project throws errors and catches them at the top level. In Go, errors are values returned from functions.&lt;/p&gt;&#xA;&lt;p&gt;I didn&#39;t translate that to the Go idiom along the way, which means I currently have 25 occurrences of the &lt;code&gt;panic&lt;/code&gt; keyword!&lt;/p&gt;&#xA;&lt;p&gt;I want to get out of this eventually, but it would be a big lift, and I wanted to prove the overall approach first. I think I&#39;ve done that now. We can start focusing on changes that make the project easier to operate and maintain.&lt;/p&gt;&#xA;&lt;h2&gt;The agentic part&lt;/h2&gt;&#xA;&lt;p&gt;I started with a very plain Claude Code + grill-me setup. Later I moved to a custom &lt;code&gt;pi&lt;/code&gt; extension built on the wayfinder skill, using herdr for sub-agent pane management.&lt;/p&gt;&#xA;&lt;p&gt;To complete the swap to &lt;code&gt;pi&lt;/code&gt; and really give it a test, I needed to swap my AI bill of choice.&lt;/p&gt;&#xA;&lt;p&gt;As the codebase grew into the tens of thousands of lines, I started thinking about quality metrics and deterministic checks. I settled on code coverage, Go linters, and static analysis tools.&lt;/p&gt;&#xA;&lt;p&gt;This is all wrapped up in a convenient &lt;code&gt;ci.sh&lt;/code&gt; that agents can call to see what needs to be improved.&lt;/p&gt;&#xA;&lt;p&gt;I also built &lt;code&gt;CONTEXT.md&lt;/code&gt;, a glossary of domain terms for the project, to help keep the language and terminology consistent as the codebase grew.&lt;/p&gt;&#xA;&lt;h2&gt;Where do I want to take this thing?&lt;/h2&gt;&#xA;&lt;p&gt;Overpass is something we depend on at myWorksites. I want a lightweight OSM service that we can host and easily extend.&lt;/p&gt;&#xA;&lt;p&gt;I also want to gradually match the rest of Overpass&#39;s capabilities. There are big things still missing from the query language, indexing, and storage layer. We can&#39;t ingest data diffs yet, and we don&#39;t support history queries.&lt;/p&gt;&#xA;&lt;p&gt;Beyond that, I want to explore vector tiles, cached layers, and new storage interfaces.&lt;/p&gt;&#xA;&lt;p&gt;I want this to become a general-purpose write-once, read-many server where you can combine OSM data with your own: filter on ingestion, merge in your datasets, and query the result.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>Go</category><category>OpenStreetMap</category><category>Overpass</category><category>Agentic Coding</category><guid isPermaLink="true">https://www.ersin.nz/articles/overpass-go</guid></item><item><title>Rewriting Spring Data JPA query methods in Go</title><link>https://www.ersin.nz/articles/repomatic</link><description>Like everyone else, I&#39;ve been experimenting with agentic coding. It took me out of my regular stack and into Spring, where I found a novel way to query a database — so I rewrote it in Go as a code generator.</description><pubDate>11 Jun 2026 08:00 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;Like everyone else, I&#39;ve been experimenting with agentic coding. Lately I&#39;ve been vibing with a bunch of frameworks to figure out what&#39;s good and what sucks stack-wise when it comes to AI code generation. On the way I ventured out of my regular stack, picking up the Spring framework for the first time and coming across a (to me) novel way of querying the database. This is the story of re-writing that idea in Go.&lt;/p&gt;&#xA;&lt;p&gt;Spring Data JPA is an ORM with a very comprehensive feature set. The bit that caught me was its &lt;a href=&#34;https://docs.spring.io/spring-data/jpa/reference/jpa/query-methods.html&#34;&gt;query methods&lt;/a&gt;: a neat piece of metadata programming where you define a method on a repository interface and the framework derives the SQL straight from the method name. Define &lt;code&gt;findByLastNameAndAgeGreaterThan&lt;/code&gt; and you get &lt;code&gt;WHERE last_name = ? AND age &amp;gt; ?&lt;/code&gt;. I was in love with the terseness and how little you actually have to write — so I re-wrote that bit in Go as a code generator. Behold &lt;a href=&#34;https://github.com/ebuckley/repomatic&#34;&gt;github.com/ebuckley/repomatic&lt;/a&gt;.&lt;/p&gt;&#xA;&lt;p&gt;It comes with far less sugar than the full JPA API does, but it holds the same basic principle: you hand it a plain interface and some plain old structs for your model layer, and it generates the implementation. So an interface like this:&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-go&#34;&gt;//go:generate go run github.com/ebuckley/repomatic/cmd/repomaticgen -type=Books,Authors&#xA;type Books interface {&#xA;    FindByAuthorLastName(db repomatic.DB, ctx context.Context, lastName string) ([]Book, error)&#xA;}&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;...turns into a query like this:&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-sql&#34;&gt;SELECT books.*&#xA;FROM books&#xA;JOIN authors ON books.author_id = authors.id&#xA;WHERE authors.last_name = $1&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Under the hood I got to build a parser and a Go SQL generator that emits the struct implementation of the defined interface. There were challenges, like trying to disambiguate field names from operator names, and joining to related fields in a sensible way.&lt;/p&gt;&#xA;&lt;p&gt;Where I landed I am happy enough with. It&#39;s probably not something you would lightly take into production yet, but it was a fun exercise and a usable library for prototyping on top of SQL.&lt;/p&gt;&#xA;&lt;h2&gt;How I built it&lt;/h2&gt;&#xA;&lt;p&gt;I wrote this with Claude Code, applying what I think of as a &#39;diffusion&#39; model of code generation: start with rough interfaces and a bit of direction, generate a proof of concept, write tests, then repeat until you&#39;re happy with the implementation. Start blurry, sharpen on each pass.&lt;/p&gt;&#xA;&lt;h2&gt;No reflection, all generation&lt;/h2&gt;&#xA;&lt;p&gt;The primary way it differs from the JPA implementation is that there is zero reflection at runtime. Instead I &lt;em&gt;generate&lt;/em&gt; the actual Go code ahead of compile time.&lt;/p&gt;&#xA;&lt;p&gt;My initial thought was that code generation would be the sensible, idiomatic Go approach — but while trying to implement it dynamically I learned that it&#39;s actually the &lt;em&gt;only&lt;/em&gt; approach. In the Java world you can reflect on the method name and then construct your implementation of the interface at runtime. This is simply impossible in Go. You can use an unsafe operation to replace pointers to functions, but you can&#39;t construct the implementation of an interface during runtime — there has to be an actual struct implementing it, and then a small type called an itable joins the interface to its implementation. So I went down a rabbit hole learning about Go&#39;s internal memory layout, and came out the other side generating real code.&lt;/p&gt;&#xA;&lt;h2&gt;Splitting identifiers from operators&lt;/h2&gt;&#xA;&lt;p&gt;Another tricky problem was parsing identifiers and separating them from operators. In a normal language you have a character to do this — &lt;code&gt;first_and_lastname and age&lt;/code&gt; splits cleanly into 3 tokens &lt;code&gt;[first_and_lastname, and, age]&lt;/code&gt;. In my little JPA-style repo method there are no separators at all; it&#39;s all one camelCase word. So &lt;code&gt;FirstAndLastName&lt;/code&gt; could be tokenized as &lt;code&gt;[first, and, last_name]&lt;/code&gt; or as a single field &lt;code&gt;[first_and_lastname]&lt;/code&gt;, and only one of those is right.&lt;/p&gt;&#xA;&lt;p&gt;To solve this I kept a memory of the field names defined on each type and searched that first. The final trick is to always search for the largest token first, then reduce to token + operator until something matches.&lt;/p&gt;&#xA;&lt;h2&gt;Landing it&lt;/h2&gt;&#xA;&lt;p&gt;I am very busy, and I never would have gotten around to this if I didn&#39;t have agents at my fingertips. It&#39;s fun to iterate and play with ideas. I might use this in a future project — dunno!&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>Go</category><category>Spring</category><category>Code Generation</category><category>ORM</category><category>Agentic Coding</category><guid isPermaLink="true">https://www.ersin.nz/articles/repomatic</guid></item><item><title>Which framework for AI first development in 2026</title><link>https://www.ersin.nz/articles/ai-first-framework</link><description>I built the same thing nine times over in completely autonomous agent mode so that you can have one empirical data point on token efficiency, quality and effectiveness of MCPs and skills.</description><pubDate>18 May 2026 07:52 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;I built the same thing nine times over in completely autonomous agent mode so that you can have one empirical data point on token efficiency, quality and effectiveness of MCPs and skills.&lt;/p&gt;&#xA;&lt;p&gt;What I found was that:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Clojure and OCaml are 10x more token expensive than Next.js, Rails and Go&lt;/li&gt;&#xA;&lt;li&gt;Framework-specific skills and MCPs are not ubiquitous, but they probably should be&lt;/li&gt;&#xA;&lt;li&gt;Being popular in the training set means a much cheaper overall implementation&lt;/li&gt;&#xA;&lt;li&gt;You can get the job done with ANY language, but your overall quality and cost may vary&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;There is no established best practice for AI assisted development. You must be experimenting constantly to understand how new tooling can be adopted. My small part to contribute here is to burn a billion tokens on building the same project over and over so that we can better understand what tools work best with AI, and what you should consider doing on your next greenfield project.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;The hypothesis&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;Functional languages would use fewer tokens — a strong type system would also push towards leaner overall solutions.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;I can confidently say this is not true. Functional languages appear to be the worst in token efficiency, although the overall solution is very line-of-code efficient. 2026 may not be the year for my beloved esoteric functional languages after all!&lt;/p&gt;&#xA;&lt;h2&gt;What was built&lt;/h2&gt;&#xA;&lt;p&gt;For this test project I built a feature-voting SaaS where product teams collect and prioritize user feedback through a public board.&lt;/p&gt;&#xA;&lt;p&gt;The app covers the full product loop end-to-end:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Visitors can&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Submit requests&lt;/li&gt;&#xA;&lt;li&gt;Vote&lt;/li&gt;&#xA;&lt;li&gt;Get email notifications on status changes&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;/li&gt;&#xA;&lt;li&gt;Operators can&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Authenticate to an admin panel&lt;/li&gt;&#xA;&lt;li&gt;Moderate posts&lt;/li&gt;&#xA;&lt;li&gt;Bulk status update&lt;/li&gt;&#xA;&lt;li&gt;View/filter requests&lt;/li&gt;&#xA;&lt;li&gt;Publish a changelog&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;/li&gt;&#xA;&lt;li&gt;Vote integrity managed with rate limiting and spike detection&lt;/li&gt;&#xA;&lt;li&gt;SSO via signed JWT or email auth flow&lt;/li&gt;&#xA;&lt;li&gt;Public board with search, sort, status filtering and voting&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;&lt;strong&gt;How it was built&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;I started with a detailed PRD that broke the work into 26 user stories across 9 epics — phased into three releases. I used the &lt;code&gt;beads&lt;/code&gt; issue tracker to manage the plan in Claude Code. The amount of nudging needed varied between 2 and 5 times to get the agent to keep working until the project was finished.&lt;/p&gt;&#xA;&lt;h2&gt;What was the result?&lt;/h2&gt;&#xA;&lt;h3&gt;Total token burn&lt;/h3&gt;&#xA;&lt;table&gt;&#xA;&lt;thead&gt;&#xA;&lt;tr&gt;&#xA;&lt;th&gt;Stack&lt;/th&gt;&#xA;&lt;th&gt;Total Tokens&lt;/th&gt;&#xA;&lt;th&gt;Tests?&lt;/th&gt;&#xA;&lt;/tr&gt;&#xA;&lt;/thead&gt;&#xA;&lt;tbody&gt;&#xA;&lt;tr&gt;&#xA;&lt;td&gt;Go&lt;/td&gt;&#xA;&lt;td&gt;~14M&lt;/td&gt;&#xA;&lt;td&gt;&lt;/td&gt;&#xA;&lt;/tr&gt;&#xA;&lt;tr&gt;&#xA;&lt;td&gt;Next.js&lt;/td&gt;&#xA;&lt;td&gt;~26M&lt;/td&gt;&#xA;&lt;td&gt;&lt;/td&gt;&#xA;&lt;/tr&gt;&#xA;&lt;tr&gt;&#xA;&lt;td&gt;Rails&lt;/td&gt;&#xA;&lt;td&gt;~36M&lt;/td&gt;&#xA;&lt;td&gt;&lt;/td&gt;&#xA;&lt;/tr&gt;&#xA;&lt;tr&gt;&#xA;&lt;td&gt;C&lt;/td&gt;&#xA;&lt;td&gt;~52M&lt;/td&gt;&#xA;&lt;td&gt;yes&lt;/td&gt;&#xA;&lt;/tr&gt;&#xA;&lt;tr&gt;&#xA;&lt;td&gt;Java Spring&lt;/td&gt;&#xA;&lt;td&gt;~88M&lt;/td&gt;&#xA;&lt;td&gt;&lt;/td&gt;&#xA;&lt;/tr&gt;&#xA;&lt;tr&gt;&#xA;&lt;td&gt;Laravel&lt;/td&gt;&#xA;&lt;td&gt;~91M&lt;/td&gt;&#xA;&lt;td&gt;yes&lt;/td&gt;&#xA;&lt;/tr&gt;&#xA;&lt;tr&gt;&#xA;&lt;td&gt;OCaml&lt;/td&gt;&#xA;&lt;td&gt;~122M&lt;/td&gt;&#xA;&lt;td&gt;yes&lt;/td&gt;&#xA;&lt;/tr&gt;&#xA;&lt;tr&gt;&#xA;&lt;td&gt;Clojure v1&lt;/td&gt;&#xA;&lt;td&gt;~133M&lt;/td&gt;&#xA;&lt;td&gt;yes&lt;/td&gt;&#xA;&lt;/tr&gt;&#xA;&lt;tr&gt;&#xA;&lt;td&gt;Clojure v2&lt;/td&gt;&#xA;&lt;td&gt;~181M&lt;/td&gt;&#xA;&lt;td&gt;yes&lt;/td&gt;&#xA;&lt;/tr&gt;&#xA;&lt;/tbody&gt;&#xA;&lt;/table&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Go, Next.js and Rails stand out as the most token efficient&lt;/li&gt;&#xA;&lt;li&gt;Writing tests forces your agent to take more turns, which naturally means more tokens burned&lt;/li&gt;&#xA;&lt;li&gt;Functional (and esoteric) languages really do burn a lot of tokens&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h3&gt;Total LOC count&lt;/h3&gt;&#xA;&lt;table&gt;&#xA;&lt;thead&gt;&#xA;&lt;tr&gt;&#xA;&lt;th&gt;Stack&lt;/th&gt;&#xA;&lt;th&gt;Primary Lang&lt;/th&gt;&#xA;&lt;th&gt;Primary LOC&lt;/th&gt;&#xA;&lt;th&gt;Total Code LOC&lt;/th&gt;&#xA;&lt;/tr&gt;&#xA;&lt;/thead&gt;&#xA;&lt;tbody&gt;&#xA;&lt;tr&gt;&#xA;&lt;td&gt;OCaml&lt;/td&gt;&#xA;&lt;td&gt;OCaml&lt;/td&gt;&#xA;&lt;td&gt;3,918 (8 files, 1,000 in test)&lt;/td&gt;&#xA;&lt;td&gt;3,928&lt;/td&gt;&#xA;&lt;/tr&gt;&#xA;&lt;tr&gt;&#xA;&lt;td&gt;Rails&lt;/td&gt;&#xA;&lt;td&gt;Ruby&lt;/td&gt;&#xA;&lt;td&gt;1,732 (88 files)&lt;/td&gt;&#xA;&lt;td&gt;4,950&lt;/td&gt;&#xA;&lt;/tr&gt;&#xA;&lt;tr&gt;&#xA;&lt;td&gt;C&lt;/td&gt;&#xA;&lt;td&gt;C&lt;/td&gt;&#xA;&lt;td&gt;2,315 (1 file)&lt;/td&gt;&#xA;&lt;td&gt;5,150&lt;/td&gt;&#xA;&lt;/tr&gt;&#xA;&lt;tr&gt;&#xA;&lt;td&gt;Clojure v2 (no REPL)&lt;/td&gt;&#xA;&lt;td&gt;Clojure&lt;/td&gt;&#xA;&lt;td&gt;3,677 (29 files, 1,241 in test)&lt;/td&gt;&#xA;&lt;td&gt;5,346&lt;/td&gt;&#xA;&lt;/tr&gt;&#xA;&lt;tr&gt;&#xA;&lt;td&gt;Clojure v1 (with REPL)&lt;/td&gt;&#xA;&lt;td&gt;Clojure&lt;/td&gt;&#xA;&lt;td&gt;2,847 (15 files, 599 in test)&lt;/td&gt;&#xA;&lt;td&gt;5,799&lt;/td&gt;&#xA;&lt;/tr&gt;&#xA;&lt;tr&gt;&#xA;&lt;td&gt;Go&lt;/td&gt;&#xA;&lt;td&gt;Go&lt;/td&gt;&#xA;&lt;td&gt;3,095 (25 files)&lt;/td&gt;&#xA;&lt;td&gt;7,013&lt;/td&gt;&#xA;&lt;/tr&gt;&#xA;&lt;tr&gt;&#xA;&lt;td&gt;Next.js&lt;/td&gt;&#xA;&lt;td&gt;TypeScript&lt;/td&gt;&#xA;&lt;td&gt;4,743 (72 files)&lt;/td&gt;&#xA;&lt;td&gt;10,230&lt;/td&gt;&#xA;&lt;/tr&gt;&#xA;&lt;tr&gt;&#xA;&lt;td&gt;Laravel&lt;/td&gt;&#xA;&lt;td&gt;PHP + TypeScript&lt;/td&gt;&#xA;&lt;td&gt;6,011 PHP + 7,530 TS (2,923 in tests)&lt;/td&gt;&#xA;&lt;td&gt;17,423&lt;/td&gt;&#xA;&lt;/tr&gt;&#xA;&lt;/tbody&gt;&#xA;&lt;/table&gt;&#xA;&lt;p&gt;Clojure, Rails and OCaml are incredibly terse — especially when you consider the significant testing component of the OCaml/Clojure codebases.&lt;/p&gt;&#xA;&lt;h3&gt;How did framework skills impact the outcome?&lt;/h3&gt;&#xA;&lt;p&gt;Next.js and Laravel have the most coherent recommended practice for AI-based development. They ship skill files, MCPs and recommended &lt;code&gt;claude.md&lt;/code&gt; structures. Within Next.js, there is an ORM-specific skill (Prisma) and a design system-specific skill (shadcn). Laravel included an MCP that can unify backend and frontend logs, and a significant suite of skill files including &#39;best practice&#39; and &#39;documentation links&#39;. Next.js was a surprisingly token-efficient implementation despite a high line count — though the shadcn components do account for a lot of those lines. The Laravel implementation wrote a full test suite and took many turns to iterate and validate through testing — as a result the token cost was higher, but this feels worthwhile in terms of the maintainability of the end solution.&lt;/p&gt;&#xA;&lt;p&gt;There is surprisingly little out there in the way of official framework-specific skill files, MCPs or other AI harness tooling. There is real opportunity here for framework developers, as many people will be approaching new projects with an LLM-first approach.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;The most interesting tooling story, though, was Clojure.&lt;/strong&gt; My experience watching Claude edit Clojure was ... painful. The agent would get in such a twist balancing parentheses when editing code that it would eventually start writing Python scripts to do the editing for it.&lt;/p&gt;&#xA;&lt;p&gt;After a small amount of research I came across &lt;a href=&#34;https://github.com/bhauman/clojure-mcp-light&#34;&gt;clojure-mcp-light&lt;/a&gt; — a recommended MCP that encourages REPL-driven development and provides a small command-line utility that helps the agent balance parentheses during edits. What made it unique was a deterministic hook that runs after every edit.&lt;/p&gt;&#xA;&lt;p&gt;The result surprised me: no marked token efficiency gain, but the agent took a fundamentally &lt;em&gt;different approach&lt;/em&gt; to the whole project. It produced 2x the tests and fewer overall lines of code than the first attempt. The right tooling didn&#39;t just make Clojure cheaper — it changed how the agent thought about the problem.&lt;/p&gt;&#xA;&lt;h2&gt;So what framework should I use?&lt;/h2&gt;&#xA;&lt;p&gt;Although AI is capable of making a solution in any language, you should consider longer-term maintainability. That comes from simplicity in the overall solution, reduced churn, and consistent use of patterns — code that a human can actually follow. Right now the bottleneck is humans understanding the system, so you really want the agent writing in a structure optimised for human consumption.&lt;/p&gt;&#xA;&lt;p&gt;It is hard to move away from mature frameworks when you want to keep a system going for 10 years and have multiple people maintaining and extending it. We should not &lt;em&gt;just&lt;/em&gt; consider lines of code, just as we should not &lt;em&gt;just&lt;/em&gt; consider token efficiency.&lt;/p&gt;&#xA;&lt;p&gt;Right now the tooling is so new that there is no clear winner in out-of-the-box skills and MCPs. The addition of framework-specific AI tooling is a nice-to-have, but not yet a game changer.&lt;/p&gt;&#xA;&lt;p&gt;I would encourage you to consider the following before committing to a framework in 2026:&lt;/p&gt;&#xA;&lt;ol&gt;&#xA;&lt;li&gt;What can I hire for reliably over the next 10 years?&lt;/li&gt;&#xA;&lt;li&gt;What will I enjoy working on?&lt;/li&gt;&#xA;&lt;li&gt;Choose your top 3 tools and vibe-code a prototype in each — see which solution fits the way your brain works.&lt;/li&gt;&#xA;&lt;/ol&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>Legacy Code</category><category>Philosophy</category><category>Lindy Effect</category><guid isPermaLink="true">https://www.ersin.nz/articles/ai-first-framework</guid></item><item><title>Who builds software that lasts 50+ years?</title><link>https://www.ersin.nz/articles/c89-for-50-year-projects</link><description>The first 10 years of my career have been focused on building things for the internet. There are other niches who focus WAY more on quality and longetivity of software, I find this strange and intersting and I&#39;m learning from their approach.</description><pubDate>01 Nov 2025 22:18 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;Over 10 years I&#39;ve used five languages, 10+ web frameworks, and probably over 500 repositories. You might be justified in calling me a fad-jumping technology enjoyer. Like many, I&#39;ve been biased towards what&#39;s new and shiny and trending on the first page of HN.&lt;/p&gt;&#xA;&lt;p&gt;But that&#39;s not representative of the industry as a whole. Did you know there are indie game developers who carefully craft their own engines and, over periods of 15+ years, diligently support the same frameworks and codebase across multiple projects? Eskil Steenberg is who I&#39;m talking about. He is a legend, and you should explore his world. He talks about being ALL IN on C89, a 36-year-old standard for the C language. The platforms he builds will compile to anything and outlast anything built for the web.&lt;/p&gt;&#xA;&lt;p&gt;Legacy software is a dirty word, but my point of view is different. Imagine your ancestors using software written 100 years ago—that sounds like a fantastic legacy, something SO useful that it can still be used in the future. This is what attracted me to C89. It will compile on computers 50 years in the future because it compiles on computers from 25 years ago. The excitement of building software that just works and will continue to work is a strange concept for someone who has spent a decade following industry trends in web development.&lt;/p&gt;&#xA;&lt;p&gt;We don&#39;t build software with a 50-year lifespan for really sensible commercial reasons. It&#39;s not worth spending 5x as long to develop a system when you&#39;re still exploring product-market fit. Once you extract the value, you rarely go back and rewrite. This commercial context is where I have focused my career. You learn a ton and explore a lot of different approaches to solving the same problems.&lt;/p&gt;&#xA;&lt;p&gt;You are more employable when you have exposure to a wider set of technologies and approaches. The cards are stacked against you if you&#39;re ALL IN on a certain tool or approach. Hiring managers care a lot about curiosity in technology as a proxy for aptitude at the job. Writing in the latest version of a tool or language is a kind of signaling to the labor market that you&#39;re going to upskill and bring fresh ideas to a team.&lt;/p&gt;&#xA;&lt;p&gt;When have you ever seen a rewrite choose the exact same stack? In my experience, it simply never happens. We choose the new shiny thing based on promises that it will last longer, be faster, or pay us more to write it. Rewriting can be in the same language and solve the old problems while achieving the needed outcomes in many cases.&lt;/p&gt;&#xA;&lt;p&gt;But what about the other side, writing software that lasts FOREVER? There are niches in the industry that take this completely different approach. Solo devs like Eskil are an  example, he grows a set of libraries, and takes them from project to project --extending them as needed. DHH is another. Yes, he does rewrite Basecamp—but it&#39;s always Rails, and it&#39;s always an incremental improvement. The same system lives on with minimal churn in technology.&lt;/p&gt;&#xA;&lt;p&gt;Low level programming for hardware often means a longer time between patching -- your code might be in space, under the ocean or deep in a mine. The systems need to last in the most harsh environments known to man: radiation, extreme hot and cold. To succeed, they MUST match the environment with an equally hardcore approach to quality and longevity. From the hardware layer through to the software layer, this permeates everything.&lt;/p&gt;&#xA;&lt;p&gt;The first ten years of my career have focused on increasing speed of shipping, because building the right thing was the biggest risk. You iterate faster to get more feedback and change your trajectory. In the hyperactive web engineering world, you often stop earlier with the realization that this code doesn&#39;t need to last longer than a couple years.  What drew me in to exploring the world of old school C, was the philosophy of it&#39;s engineers -- write 5x the code, but also get 5x the longevity with the end system. There is undoubtedly things to learn from each other when you cross domains.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>Legacy Code</category><category>Philosophy</category><category>Lindy Effect</category><guid isPermaLink="true">https://www.ersin.nz/articles/c89-for-50-year-projects</guid></item><item><title>AI copilots for the eternal tinkerer</title><link>https://www.ersin.nz/articles/ai-for-tinkerers</link><description>AI copilots won’t replace developers—they super-charge prototyping. You should keep adopting the tools, but be wary of the hype and doom.</description><pubDate>18 Sep 2025 22:38 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;If you write code, you&#39;re a programmer. If you start ventures, you&#39;re an entrepreneur. Simple, right? But what are you when you&#39;re obsessed with experimenting with every new framework that hits Hacker News?&lt;/p&gt;&#xA;&lt;p&gt;I&#39;ve spent over a decade being that person – the eternal tinkerer. Back in university, I thought chasing every trending technology would give me an edge. Now, with the rise of AI and LLMs, that same curiosity has evolved into something far more powerful.&lt;/p&gt;&#xA;&lt;p&gt;Here&#39;s the thing: While twitter is flooded with predictions about AI replacing developers, I&#39;m seeing something different. The new tools are insanely awesome for starting new projects.  After building documentation platforms, Xero tools and new backend frameworks, I&#39;ve discovered that AI isn&#39;t our replacement.&lt;/p&gt;&#xA;&lt;p&gt;Coding with AI by your side is like working with a really terrible Junior engineer. One that doesn’t learn, writes a lot of code, often goes down a rabbit hole.  Creating prototypes has never been more fun or easy as a result.&lt;/p&gt;&#xA;&lt;h2&gt;Towards higher and higher levels of code&lt;/h2&gt;&#xA;&lt;p&gt;&lt;img src=&#34;data:image/png;base64,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&#34; alt=&#34;&#34;&gt;&lt;/p&gt;&#xA;&lt;p&gt;The last 40 years of progress in computing has been all about a shift from low level abstractions to high level conceptual solutions.&lt;/p&gt;&#xA;&lt;p&gt;Solutions built for specific computer architectures were replaced with general purpose languages and better compilers that could target multiple architectures.&lt;/p&gt;&#xA;&lt;p&gt;Languages became business oriented. COBOL came with ‘human’ friendly syntax closer to stakeholder language. Computers got fast enough for dynamic languages.&lt;/p&gt;&#xA;&lt;p&gt;Platforms for massive distribution became a thing, java, the web, mobile. We saw domination by organisations that could monopolize a platform.  Big tech companies go through multi billion dollar shifts to go after speculated new platforms like virtual reality.&lt;/p&gt;&#xA;&lt;p&gt;But in October 2021 the next big thing actually happened, and that was the release of github copilot.   After a couple years with this AI sidekick, we can clearly see this as yet another shift on the continuum of focusing on the customer pain first. Prototyping has become easier than ever before. Even people who have never written a line of code can build adequate systems for testing a solution.&lt;/p&gt;&#xA;&lt;h2&gt;AI is great for exploration&lt;/h2&gt;&#xA;&lt;p&gt;&lt;img src=&#34;data:image/png;base64,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&#34; alt=&#34;&#34;&gt;&lt;/p&gt;&#xA;&lt;p&gt;Some thought leaders might tell you that AI is replacing us, that you should buy their course and change careers. The reality is that new AI tools can be great for exploring a problem space and building prototypes. When software actually becomes more valuable and legacy, then the effectiveness and usefulness of AI falls off a cliff. We still need humans in the loop to guide agents towards sustainable extension of legacy codebases.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Software runs through this life-cycle forever&lt;/strong&gt;&lt;br&gt;&#xA;All software runs through the same cycle of exploration and sunset into legacy. Agents will not change this reality.  Starting something new, the most important thing to do is get your solution in front of the people who are trying to do the job with it.  If you get it right, the application moves on, code runs for months and years. As the codebase grows, it shifts towards a state of optimization and happy legacy. Once that legacy becomes too brittle, and everyone who originally wrote it moves on – you begin the rebuild. You pull out threads from the wonderful legacy system and build little components to explore solving the problem again.&lt;/p&gt;&#xA;&lt;p&gt;Frameworks have a similar value lifecycle. In the  beginning they provide guard rails to your development. They prevent solving the boilerplate problems like authentication. Within a year, most of that value is gone. You start to build your own abstractions. After 5 years, that framework is probably completely unrecognisable from an application written with a custom framework.&lt;/p&gt;&#xA;&lt;p&gt;We should care about legacy software. The expected future value of a technology is proportional to the length of time it has existed so far. This is also known as the &lt;a href=&#34;https://en.wikipedia.org/wiki/Lindy_effect&#34;&gt;Lindy Effect&lt;/a&gt;. Agents so far have proven themselves not very capable enhancing legacy software.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Where does current AI fit into the paradigm?&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;Keep your agents focused on rapid prototyping for now. Use them to work very closely with the designers, customers and experts wherever you are applying your craft. We can learn faster than ever about the ‘job to be done’ by leveraging our new prototyping super powers.&lt;/p&gt;&#xA;&lt;p&gt;As  of September 2025 – I would say always be experimenting. You should dispatch an agent on a legacy code base, you should craft your prompts and context to bring in the right type of changes to your legacy code.&lt;/p&gt;&#xA;&lt;h2&gt;The future is exciting&lt;/h2&gt;&#xA;&lt;p&gt;This old curmudgeon of a developer is enthusiastically trying all of the AI tools. I  still keep a keen eye on hacker news and twitter for the latest and greatest. I am approaching the new tools and techniques with excitement and skepticism.&lt;/p&gt;&#xA;&lt;p&gt;If you are just entering the industry, or thinking about it. Do take up the AI tools, and experiment with the new shiny thing. But also, do not take serious account of AI grifters and doomers who might say that the end of our industry is just around the corner. We can have fulfilling and fun careers building really valuable software with copilots and agents at our side.&lt;/p&gt;&#xA;&lt;p&gt;Technological progress has not been a constant in the world, but throughout my life and yours it will be. We either ride the wave, or we take up new careers. Join me on the adventure, I really look forward to what  the future holds for our industry.&lt;/p&gt;&#xA;&lt;p&gt;As always, if you’re interested in semi-non-regular emails about my thinking, you can sign up to the RSS feed on this site or put your email in the box below for updates.&lt;/p&gt;&#xA;&lt;p&gt;Peace &amp;amp; happy coding.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>AI agents</category><category>Lindy Effect</category><category>Legacy Code</category><guid isPermaLink="true">https://www.ersin.nz/articles/ai-for-tinkerers</guid></item><item><title>Stand out be authentic: landing a job in 2025</title><link>https://www.ersin.nz/articles/apply-with-ai</link><description>Learn how to stand out in the job application process. 2025 is a tough year to be a junior, but you can pick up a few of my tricks here to stand a better chance. Everyone looks the same when they generate an application with AI. Now it is more important than ever to stand out. </description><pubDate>20 Apr 2025 14:51 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;Vibe coding killed the junior developer. Bullshit. I couldn’t disagree more. I’ve been fortunate to recently hire new developers and I can see a bunch of hungry and skilled grads who are enthusiastic and capable of picking up all the skills you need.&lt;/p&gt;&#xA;&lt;p&gt;If you’re currently searching for a job and feeling some anxiety about it, that’s OK you have good reason to be anxious, but the industry &lt;em&gt;will&lt;/em&gt; start hiring juniors again. In the interim I have some advice for you on how to stick out from a list of 100+ applications.&lt;/p&gt;&#xA;&lt;h2&gt;&lt;strong&gt;The Cover Letter Reality Check&lt;/strong&gt;&lt;/h2&gt;&#xA;&lt;p&gt;When I check your application on seek.co.nz, I’m going to spend about 30 seconds total figuring out if you’re taking a closer look. What catches my eye is anything that stands out as distinctly &lt;em&gt;NOT&lt;/em&gt; ai generated. The generated slag is super easy to spot when you have 30 instantiations of it. Apparently we are all using the same prompt. People seem to paste the job description into chat gpt and then ask for a stupid prompt like ‘write a cover letter applying for this job’.&lt;/p&gt;&#xA;&lt;p&gt;The obvious sign of AI crap writing is that it overfits exactly to the job post and every facet. The Job applications all read the same because the input is all too similar.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;What you should do instead&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;Get creative with your prompt. Use the AI for fleshing out your own thoughts, not coming up with ideas. Find what is most important and relevant from the job description, then prompt the AI to include that in the cover letter. Finally — get to the point as fast as possible. If I read one more time “I am a dynamic individual ready for the fast paced world of junior engineering” I will eat my fucking hat.&lt;/p&gt;&#xA;&lt;h2&gt;&lt;strong&gt;Standing Out in the AI Era&lt;/strong&gt;&lt;/h2&gt;&#xA;&lt;p&gt;&lt;strong&gt;Be authentic.&lt;/strong&gt; Legit, when I’m reading all these applications I look positively on the odd quirky turn of phrase or spelling/grammer mistake. Small human errors are the bit around the edge that can show you are not a robot.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Give more context&lt;/strong&gt;. When you are using AI to prompt, use it for the brain storming – but write the content yourself. Do include an extensive document of your portfolio as context and ask the AI to draw specific correlations between your own experience and the job description.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Be transparent&lt;/strong&gt; about AI usage. Seriously this will make you stick out like a sore thumb and I will probably automatically accept the application for an in person interview without a second look. Tell me about the creative ways you are using AI to apply for jobs. I want to hire people who are experts at leveraging the new tools in our industry.&lt;/p&gt;&#xA;&lt;h2&gt;&lt;strong&gt;Closing thoughts&lt;/strong&gt;&lt;/h2&gt;&#xA;&lt;p&gt;&lt;strong&gt;Don’t panic&lt;/strong&gt; there will be jobs for junior engineers in 2025 and the industry will suffer for slowing down in hiring new people. The job is awesome, and I would highly recommend it if you like solving hard problems and enjoy the craft of building systems with code. There really are some awesome opportunities out there, and you can increase your chances of finding a good one by taking a few pointers from this post.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>beginner</category><category>ai writing</category><category>hiring</category><guid isPermaLink="true">https://www.ersin.nz/articles/apply-with-ai</guid></item><item><title>The Ersin Test</title><link>https://www.ersin.nz/articles/the-ersin-test-2025</link><description>Joel Spolskey wrote an amazing piece &#39;the joel test&#39;. He gives us 12 questions that any great software team should say yes too. Today, I gift you the 2025 version with my own experience working in some pretty great teams.</description><pubDate>22 Mar 2025 14:51 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;At the turn of the century, Joel Spolskey introduced the world to the Joel Test. We got 12 quick questions to tell if a company is worth working for or if you should start running. Most companies should be 12/12 – if they answer 10/12, that’s OK too, if they answer 8/12. Get. Out. Now!&lt;/p&gt;&#xA;&lt;p&gt;Behold the ‘Ersin test’. A 2025 update of the Joel test. A lot of it has changed, but four (in bold) of his questions are perfect in their early 2000’s form.&lt;/p&gt;&#xA;&lt;ol&gt;&#xA;&lt;li&gt;Do you practice trunk base development?&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Can you build in one step?&lt;/strong&gt;&lt;/li&gt;&#xA;&lt;li&gt;Can you deploy without ops?&lt;/li&gt;&#xA;&lt;li&gt;Do you listen to customer feedback?&lt;/li&gt;&#xA;&lt;li&gt;Are you spending 20% time on quality?&lt;/li&gt;&#xA;&lt;li&gt;Do you know the quarterly goals/outcomes?&lt;/li&gt;&#xA;&lt;li&gt;Do you prototype before you build?&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Do programmers have quiet working conditions?&lt;/strong&gt;&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Do you experiment with tools?&lt;/strong&gt;&lt;/li&gt;&#xA;&lt;li&gt;Do you use the software you build?&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Do new candidates write code before joining?&lt;/strong&gt;&lt;/li&gt;&#xA;&lt;li&gt;Do you demo all the time?&lt;/li&gt;&#xA;&lt;/ol&gt;&#xA;&lt;p&gt;This list is my own, and it comes from 10+ years working on some of the best teams which have delivered great products. You might not find every answer to be yes, but I think most good teams will be 8/12 or better.&lt;/p&gt;&#xA;&lt;p&gt;What is amazing is that a lot of the vibe from the Joel test hasn’t changed much in 25 years. This stuff is going to be important in 2250 as it is in 2025.&lt;/p&gt;&#xA;&lt;h2&gt;Do you practice trunk base development?&lt;/h2&gt;&#xA;&lt;p&gt;Every team uses source control. It is ubiquitous. You &lt;em&gt;can&lt;/em&gt; find teams that do not practice trunk based development, this is a red flag 🚩. At the turn of the century, Joel observed that some teams don’t use source control. In the current age, it’s so obvious it’s not worth mentioning.&lt;/p&gt;&#xA;&lt;p&gt;No matter the way you ship code, it should pass through an automated pipeline and you should ship changes very frequently to a common branch. Tests/linters/static analysis are important in 2025. If having fun building software is a goal, you should answer yes to this question.&lt;/p&gt;&#xA;&lt;h2&gt;Can you build in one step?&lt;/h2&gt;&#xA;&lt;p&gt;True in 2000, true in 2025, true in 2250. We should work really hard to make our system buildable in one step. Has this gotten harder to achieve? My impression is that more software product teams would answer ‘yes’ to this than ever before. It is a fundamental thing that you have more disparate tools. The amount of &lt;em&gt;stuff&lt;/em&gt; you need to keep in your head as a software engineer has grown.&lt;/p&gt;&#xA;&lt;p&gt;More complexity and disparate systems to wrangle in any given product doesn’t dis-credit the importance of having a one step build.&lt;/p&gt;&#xA;&lt;p&gt;Working in a small team you have far less excuse to not have a one step build. This is a superpower you really need to go after in small product teams.&lt;/p&gt;&#xA;&lt;h2&gt;Can you deploy without ops?&lt;/h2&gt;&#xA;&lt;p&gt;A big change since the 2000’s is the shift in responsibilities between engineering and IT operations. We have much more tools and platforms that can be leveraged. Overall this has led to an environment where every startup can have an easy platform for shipping on.&lt;/p&gt;&#xA;&lt;p&gt;On top of purpose built cloud providers like old school heroku, or new school Vercel – you have a really strong group of open source tools that can be used for building a platform.&lt;/p&gt;&#xA;&lt;p&gt;You can build a platform without being an old school IT operator. I would recommend people to stick with out of the box platforms where possible, but you can engineer and operate one if you really want, and you should try and do this without separating the engineering and IT operations function.&lt;/p&gt;&#xA;&lt;h2&gt;Do you listen to customer feedback?&lt;/h2&gt;&#xA;&lt;p&gt;I recently heard this named as the ‘‘two team bifurcation &#39;’. Having a feature squad and a support squad is an anti-pattern; you should avoid this – especially in small teams and new products.&lt;/p&gt;&#xA;&lt;p&gt;The team should be close to the customer feedback at all points. Your product designers should actually be on the team. Your sales/support people should be very close to the engineers at all times.&lt;/p&gt;&#xA;&lt;p&gt;We can build a very strong engineering organisation by putting the voice of the customer as close to the delivery team as possible.&lt;/p&gt;&#xA;&lt;p&gt;Splitting the customer voice from a feature delivery team might sound amazing but in practice it leads to resentment and a higher degree of building the wrong thing.&lt;/p&gt;&#xA;&lt;h2&gt;Are you spending 20% time on quality?&lt;/h2&gt;&#xA;&lt;p&gt;There are a ton of good ways to manage this. Screwing this up is catastrophic. Quality of systems trend towards 0, flexibility and velocity grind to a halt. You have metaphorically painted your product in to a corner.&lt;/p&gt;&#xA;&lt;p&gt;A continuous approach to quality is the antidote. My favourite way to do this is just to be really explicit to the team that you can spend 20% of your time on quality. You can release a feature to customers and you will still get 20% time to improve that code base. You are allowed to tidy the internals of the product as you go.&lt;/p&gt;&#xA;&lt;p&gt;In 2025 we are about to hit a real disaster on this front, and you can see the early signs of this based on the current vibe coding trend. If you are a traveller from the future reading this blog post, it’s worth a small call out on what is happening. Novice developers are currently ‘vibe coding’ (AI LLM coding) some amazing products into the world. I love this, and I hate this trend. A massive problem with these vibers is that they inevitably hit a wall, assuming you can just cram features forever. You &lt;em&gt;must&lt;/em&gt; focus on the quality of the internal structures of a software system. The current realm of AI coding models are not good at this in 2025.&lt;/p&gt;&#xA;&lt;p&gt;The most common anti-pattern here is to let someone outside the trenches of engineering define a list of what can and can’t be worked on to improve the quality of the system. Ownership of this backlog really needs to stay with the engineers who are going to maintain the product in the long run.&lt;/p&gt;&#xA;&lt;h2&gt;Do you have quarterly goals/outcomes?&lt;/h2&gt;&#xA;&lt;p&gt;Joel reckons we should have a schedule. I agree in principle, but I don’t think his description works very well for the small/medium startup. What I love to see is a really well communicated organisational objective and quarterly goal set.  You can spot this in a high functioning team by asking a random selection of developers (what are the six things we need to achieve in the next 2 quarters). The six things should paint a pretty obvious picture on what the business is doing, they do not need to be the exact things but you should get a clear idea of the team&#39;s alignment.&lt;/p&gt;&#xA;&lt;p&gt;We have a lot more nuance to schedules in 2025. There is a spectrum of highly defined project management through to some very loosely coupled versions on project management. In my experience it is less important than it seems. However, my preference tends towards the process which allows me more freedom to change direction.&lt;/p&gt;&#xA;&lt;p&gt;More important is to understand what the hell the business outcomes are, and how your product is contributing to the success in the short term.&lt;/p&gt;&#xA;&lt;p&gt;A ‘schedule’ is not a good question to figure out if the product team is actually delivering value to the business in a way that is sustainable and going to lead to a long term stable product.&lt;/p&gt;&#xA;&lt;h2&gt;Do you prototype before you build?&lt;/h2&gt;&#xA;&lt;p&gt;You should write tracer code. You shouldn’t spend weeks on design. Building prototypes has gotten significantly easier between 2000 → 2023. It just got insane with the release of LLM’s that can write code pretty good. The 2023 → 2025 improvement in prototyping is about the same as the 2000 → 2023 improvement. I believe we are potentially in an exponential on the ability to prototype software solutions.&lt;/p&gt;&#xA;&lt;p&gt;Building prototypes is fun. It actually leads to building the right thing more often on average. Building the right thing is fun.&lt;/p&gt;&#xA;&lt;p&gt;Not building prototypes is no fun. Working for a long time really hard on a well scoped project is OK I guess, but wow, it sucks  when that really expensive product team delivers something that no one wants.&lt;/p&gt;&#xA;&lt;h2&gt;Do programmers have quiet working conditions?&lt;/h2&gt;&#xA;&lt;p&gt;Coding can be collaborative (see pairing). Design is collaborative for sure! But actually when we are doing the production of software we are getting in to flow (see Mihaly Csikszentmihalyi).&lt;/p&gt;&#xA;&lt;p&gt;The way we do this might have some subtle differences in 2025, but not as much as a lot of other factors here. We are still human and our biology changes slowly. Getting into the flow probably hasn’t changed much in 5000 years.&lt;/p&gt;&#xA;&lt;p&gt;Maybe VR will make this better? Good noise cancelling headphones are worth a note here. Flexible working is also great, I still find an occasional evening or early morning session in the editor can be amazing for getting the job done.&lt;/p&gt;&#xA;&lt;h2&gt;Do you experiment with tools?&lt;/h2&gt;&#xA;&lt;p&gt;Prev. “do you use the best tools?”. We are in a really insanely volatile stage of tool change in our industry right now. It is more important than ever to experiment with new tools.&lt;/p&gt;&#xA;&lt;p&gt;What is ‘best’ is changing on a monthly basis now. You should be open to changing and exploring new tools all the time.&lt;/p&gt;&#xA;&lt;p&gt;Now, I’m obviously alluding to AI assistants, but I think experimentation with new tools is still really important beyond the current hyper trend of large language models. Take Operating systems as an example.  In my career I have gone Windows -&amp;gt; Linux -&amp;gt; Mac -&amp;gt; Linux -&amp;gt; Windows -&amp;gt; Linux… At each time I used these, they were the best option for the job.&lt;/p&gt;&#xA;&lt;p&gt;Flexibility and open mindedness is important. I can remember starting my career and veterans being extremely hesitant to try new tools like Docker, don’t be like that! Explore new tools and choose what is best!&lt;/p&gt;&#xA;&lt;h2&gt;Do you use the software you build?&lt;/h2&gt;&#xA;&lt;p&gt;One of the trends we have seen over the last quarter century is a huge expansion of the job to be done by software engineers. More than ever, it is becoming very important to also be a ‘product engineer’. To pick up these skills you &lt;em&gt;must&lt;/em&gt; use the software that you build.&lt;/p&gt;&#xA;&lt;p&gt;It’s not like you need a perfect developer SME in the domain you are building software. If you find this specific type of unicorn you should totally hire them though. At least make sure that your users are close to the developers.&lt;/p&gt;&#xA;&lt;p&gt;We must pick up some type of ‘method acting’ paradigm when we work on products, and become the person who uses the software. Understanding deeply what they are doing is key to building amazing products.&lt;/p&gt;&#xA;&lt;h2&gt;Do new candidates write code before joining?&lt;/h2&gt;&#xA;&lt;p&gt;True in 2000, still true in 2025. Although we have invented a large number of variations on this in the preceding years. We have a lot more than fizzbuzz. We have leet code, system design interviews and a whole variety of coding tests that you might experience.&lt;/p&gt;&#xA;&lt;h2&gt;Do you demo all the time?&lt;/h2&gt;&#xA;&lt;p&gt;Release early, release often. Joel called this hallway testing. But really, you should be showing off unfinished work as early as possible. The feedback you get while you are mid feature development is 1000% more useful because you can actually immediately change the course if necessary.  Another banger of a question that really has minimally changed between 2000 and 2025.&lt;/p&gt;&#xA;&lt;h1&gt;Putting a bow on this all&lt;/h1&gt;&#xA;&lt;p&gt;There is a ton that has not changed much at all between 2000 and 2025. We work on software that looks a lot different. Your speciality is much wider and there are more disparate parts to every product. At a high level the act of writing code looks the same, the qualities of really good engineering teams haven&#39;t changed a lot either.&lt;/p&gt;&#xA;&lt;p&gt;The stuff that was true in 2000 as much as 2025 will be true in 2050. Take this forward to all your social media feeds and apply it to the current trends and you will be able to see through some of the obvious bullshit out there.&lt;/p&gt;&#xA;&lt;p&gt;Even if we don’t look at lines of code in the future, a huge amount of this is still going to apply.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>joel spolskey</category><category>advice for small teams</category><category>software engineering practice</category><guid isPermaLink="true">https://www.ersin.nz/articles/the-ersin-test-2025</guid></item><item><title>The queue theory lesson every software team needs to learn</title><link>https://www.ersin.nz/articles/queue-theory</link><description>Queue Theory helps us understand some counter intuitive things about software development. I use hardcore math and a Javascript simulation to illustrate the concepts.</description><pubDate>14 Feb 2025 14:51 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;Despite being written in 2013, the  &lt;a href=&#34;https://www.amazon.com.au/Phoenix-Project-DevOps-Helping-Business/dp/0988262592&#34;&gt;phoenix project&lt;/a&gt;  captures a timeless truth about software development: we keep making the same process mistakes. While its references to on-prem infra might feel dated in 2025, its core message about managing work flow remains crucial.&lt;/p&gt;&#xA;&lt;p&gt;Today we dive into queue theory from a theoretical and practical angle. At the end of it, you will know for sure why 100% utilization of engineers is a recipe for disaster.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Key takeaways&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;100% utilization leads to an infinite backlog and lead time.&lt;/li&gt;&#xA;&lt;li&gt;80% utilization is the sweet spot.&lt;/li&gt;&#xA;&lt;li&gt;Planned inefficiency results in longer term better outcomes and also makes your delivery anti-fragile&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h1&gt;A Nerdy Demonstration&lt;/h1&gt;&#xA;&lt;p&gt;My approach to problems is to build prototypes, and AI editors like windsurf have made this more fun than ever before. I have created &lt;a href=&#34;https://replit.com/@ebuckley2/Queuething&#34;&gt;https://replit.com/@ebuckley2/Queuething&lt;/a&gt; to prove a point about Queue theory today. My demo is a simple web application that let’s you control the job rate, and work rate – it outputs metrics on lead time and worker utilization.&lt;/p&gt;&#xA;&lt;p&gt;While you read, set the repl running with the following parameters&lt;/p&gt;&#xA;&lt;ol&gt;&#xA;&lt;li&gt;Set job creation: 500ms&lt;/li&gt;&#xA;&lt;li&gt;Max finish time: 650ms&lt;/li&gt;&#xA;&lt;li&gt;Min finish time: 450ms&lt;/li&gt;&#xA;&lt;/ol&gt;&#xA;&lt;p&gt;What you will see over time is that the worker utilization tends towards 100%, and the lead time also progressively grows. After 3000 iterations I am seeing a 14 second lead time. Imagine scaling this up to an average software project length of 2 weeks. That means that an average project would spend 12 weeks in the backlog before it can be picked up.&lt;/p&gt;&#xA;&lt;h1&gt;The math behind the result&lt;/h1&gt;&#xA;&lt;p&gt;Beyond just being intuitive, we can prove this with hard core math.&lt;/p&gt;&#xA;&lt;p&gt;The fundamental relationship in queuing theory is captured by the &lt;strong&gt;Pollaczek-Khintchine formula (1930)&lt;/strong&gt; (P-K formula), which shows how wait times relate to utilization.&lt;/p&gt;&#xA;&lt;p&gt;&lt;img src=&#34;data:image/png;base64,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&#34; alt=&#34;&#34;&gt;&lt;/p&gt;&#xA;&lt;p&gt;&lt;code&gt;Wq= ((1 + scv)/2))(p/(1-p))(1/μ)&lt;/code&gt;&lt;/p&gt;&#xA;&lt;p&gt;Where:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Wq is the average wait time&lt;/li&gt;&#xA;&lt;li&gt;SCV is service time variability&lt;/li&gt;&#xA;&lt;li&gt;ρ (rho) is utilization (between 0 and 1)&lt;/li&gt;&#xA;&lt;li&gt;μ (mu) is service rate&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;The key term here is &lt;code&gt;ρ1−ρ&lt;/code&gt;. As utilization (ρ) approaches 100%, this term approaches infinity. Here&#39;s what happens:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;At 50% utilization: 0.5 / 1−0.5 = 1&lt;/li&gt;&#xA;&lt;li&gt;At 80% utilization: 0.8 / 1−0.8 = 4&lt;/li&gt;&#xA;&lt;li&gt;At 90% utilization: 0.9 / 1−0.9 = 9&lt;/li&gt;&#xA;&lt;li&gt;At 95% utilization: 0.95 / 1−0.95 = 19&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;Service availability and service rate do impact the result. But there is a bigger factor. Utilization is the key. at 95% utilization, the lead time is a factor of 19! At 80% it is only a factor of 4, this is the sweet spot in my opinion.&lt;/p&gt;&#xA;&lt;h1&gt;Getting practical&lt;/h1&gt;&#xA;&lt;p&gt;Ok math is great, simulations have shown me the reality – Ersin, how do I take this into my day job?&lt;/p&gt;&#xA;&lt;p&gt;Idle time in a software time doesn’t mean sitting around doing nothing, it means time you get to focus on non functional things to improve the overall product. This is your time for making the product faster to build on in the future. This is the time you get to really focus on reducing complexity of the system as a whole.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Managers&lt;/strong&gt; you need to immediately build in 80% utilization or your team is doomed to be on a death march.  Your ‘work generation’ rate needs to match the team&#39;s ability to deliver. It doesn’t make sense to plan 12 weeks in the future, the competition will have already moved on! My all time favourite way to build in 80% utilization is to embrace a cycle like 4 weeks on, 1 week off. On the weeks off – your teams will focus on reducing complexity as a whole and being available for fighting fires.&lt;/p&gt;&#xA;&lt;p&gt;Not managing a team? You can embrace 80% from the trenches. As an &lt;strong&gt;individual contributor&lt;/strong&gt; you can add 20% to your normal estimation. Make sure you are investing time on reducing the overall complexity of the system. The counterintuitive outcome is that you will move faster in the long run. Managers will learn to depend on you for the ‘extra’ important projects. Your own projects will deliver with less risk. Your colleagues will notice the marked improvement in quality and reduction in complexity across the whole system.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;A balance&lt;/strong&gt; of project delivery and ‘idle’ time is a happy place for many engineers. You get to deliver valuable work – while still building a stable and resilient system. This is not the default mode of operation for most teams, so it often needs to be trained and explicitly built into a way of working.&lt;/p&gt;&#xA;&lt;h2&gt;Conclusion: Embracing &amp;quot;Inefficiency&amp;quot; as a Strategy&lt;/h2&gt;&#xA;&lt;p&gt;The drive for 100% efficiency often comes from well-meaning but misguided Tayloristic thinking.  As we’ve seen through 1930 math and 2025 react simulation, pushing for complete utilization creates exponential problems.&lt;/p&gt;&#xA;&lt;p&gt;When lead times grow to 19x the work rate you’re looking at a quarter-long lead time for new projects. No business can compete in 2025 when they need to plan 3+ months ahead of time.&lt;/p&gt;&#xA;&lt;p&gt;Make your product managers and business leaders happy by building an engineering organisation that can actually deliver the most important projects. Don’t load up on work that will take ages to even start. Don’t drive your teams to be 100% busy all the time.&lt;/p&gt;&#xA;&lt;p&gt;Efficiency is a naive goal, we need to focus on being anti-fragile and nimble. Operating at 80% is a good start on the path.&lt;/p&gt;&#xA;&lt;h1&gt;&lt;/h1&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>Software engineer</category><category>project management</category><category>product management</category><category>kanban</category><category>scrum</category><guid isPermaLink="true">https://www.ersin.nz/articles/queue-theory</guid></item><item><title>The Go Web Stack That Actually Makes Sense in 2025</title><link>https://www.ersin.nz/articles/a-great-web-stack-for-go</link><description>A minimalist&#39;s guide to server-side Go: Learn how Templ, Chi, and Air deliver speed and simplicity without sacrificing developer happiness.</description><pubDate>02 Feb 2025 14:51 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;Battery included frameworks like Laravel, Rails and Next.js have a place in enterprise teams and startups, but the golang developer really has no direct competitor. I’ve landed on what I consider to be the building blocks for a perfect Golang web development stack. My choices optimize for delivery speed, simplicity and fun.&lt;/p&gt;&#xA;&lt;p&gt;Here’s my battle-tested combo:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;a href=&#34;https://templ.guide/&#34;&gt;Templ&lt;/a&gt; for templating&lt;/li&gt;&#xA;&lt;li&gt;&lt;a href=&#34;https://github.com/go-chi/chi&#34;&gt;Chi&lt;/a&gt; for routing and http middleware&lt;/li&gt;&#xA;&lt;li&gt;&lt;a href=&#34;https://github.com/air-verse/air&#34;&gt;Air&lt;/a&gt; for smooth dev ex.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h1&gt;Why templ is a Game-Changer&lt;/h1&gt;&#xA;&lt;p&gt;Sure, Go has built-in templating, but Templ takes it to another level. You get strong typing and a language server for great dev experience. Here’s a quick taste.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-go&#34;&gt;package contact&#xA;&#xA;import (&#xA;   &amp;quot;github.com/ebuckley/mtb/components/layout&amp;quot;&#xA;   &amp;quot;github.com/ebuckley/mtb/lib/site&amp;quot;&#xA;)&#xA;&#xA;templ Thanks(sp site.Site) {&#xA;   @layout.Page(sp, ContactMeta(sp), nil, site.Contact) {&#xA;      &amp;lt;h1 class=&amp;quot;font-sans underline font-bold break-normal text-xl md:text-3xl&amp;quot;&amp;gt;Thanks for getting in touch!&amp;lt;/h1&amp;gt;&#xA;      &amp;lt;p\&amp;gt;I&#39;ll get back to you as soon as I can.&amp;lt;/p&amp;gt;&#xA;   }&#xA;}&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;The beauty here is that a template looks a lot like a go code file. You can use strongly typed parameters, and get nice code completion for the fields that are passed in. Embedding sub templates is easy using as I am here where I reference my page layout template &lt;code&gt;layout.Page&lt;/code&gt;.&lt;/p&gt;&#xA;&lt;p&gt;Templ is a code generator so you compile the &lt;code&gt;.templ&lt;/code&gt; files to plain old go functions. This means super fast performance in production and zero framework overhead.&lt;/p&gt;&#xA;&lt;h1&gt;Single Binary deployments&lt;/h1&gt;&#xA;&lt;p&gt;Deploying a go web application just means putting a binary up on a server and keeping it running. This is all thanks to the standard library &lt;code&gt;embed&lt;/code&gt; package. Two lines is all it takes:&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-go&#34;&gt;//go:embed public&#xA;var staticFiles embed.FS&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;The above takes my &lt;code&gt;public&lt;/code&gt; folder and embeds it all in the binary. You access the files using the standard file system interface. Code which works with the &lt;code&gt;FS&lt;/code&gt; or &lt;code&gt;io&lt;/code&gt; interfaces will ‘just work ™‘ with an embedded file same as a real file.&lt;/p&gt;&#xA;&lt;p&gt;My http middleware setup can use a folder or an embedded FS as simple as this:&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code&gt;lib.FileServer(mux, &amp;quot;/&amp;quot;, http.FS(staticFS))&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;This little middleware will take any path not served by an existing route and look up a public file from the embed.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;The caveat&lt;/strong&gt; is that you need to always reload to re-embed. I have a solution for you though. All you need is Air!&lt;/p&gt;&#xA;&lt;h1&gt;Supercharge Development with Air&lt;/h1&gt;&#xA;&lt;p&gt;For golang developers, &lt;a href=&#34;https://github.com/air-verse/air&#34;&gt;air&lt;/a&gt; is a secret weapon. It handles hot reloading so you can just hit save then refresh the browser. There is one gotcha; &lt;strong&gt;you need to configure around an infinite loop between air + templ&lt;/strong&gt;.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;To reload my app I use&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-toml&#34;&gt;cmd = &amp;quot;bunx tailwindcss -i styles/input.css -o cmd/server/public/output.css &amp;amp;&amp;amp; templ generate &amp;amp;&amp;amp; go build -o ./tmp/server ./cmd/server&amp;quot;&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;The compile command will compile tailwind css output and templ functions. This means that if you do not exclude certain folders and file patterns then an infinite loop of compilation and generation might happen.&lt;/p&gt;&#xA;&lt;p&gt;To work around it use the following bit of config in your &lt;code&gt;.air.toml&lt;/code&gt;&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-toml&#34;&gt;exclude_dir = [&amp;quot;assets&amp;quot;, &amp;quot;tmp&amp;quot;, &amp;quot;vendor&amp;quot;, &amp;quot;testdata&amp;quot;, &amp;quot;node_modules&amp;quot;, &amp;quot;cmd/server/public&amp;quot;]&#xA;exclude_regex = [&amp;quot;_test.go&amp;quot;, &amp;quot;.*_templ.go&amp;quot;]&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;We ignore big folders like &lt;code&gt;testdata&lt;/code&gt; and &lt;code&gt;node_modules&lt;/code&gt; and places where the content changes as a result of the build I.E &lt;code&gt;cmd/server/public&lt;/code&gt;&lt;/p&gt;&#xA;&lt;p&gt;We also ignore &lt;code&gt;_templ.go&lt;/code&gt; files because they will be generated from our previous &lt;code&gt;templ generate&lt;/code&gt; step during compilation.&lt;/p&gt;&#xA;&lt;h1&gt;Choose simplicity&lt;/h1&gt;&#xA;&lt;p&gt;Look, there are a million ways to build web apps. But after trying pretty much everything out there, this stack just hits different. &lt;strong&gt;Simplicity&lt;/strong&gt; and &lt;strong&gt;power&lt;/strong&gt; that works perfect for me.&lt;/p&gt;&#xA;&lt;p&gt;The best part about all this is that it starts small and can scale when you need it. No Over-engineering, no framework lock-in. Just standard library and small choices that can be replaced if needed.&lt;/p&gt;&#xA;&lt;p&gt;If you’re starting with Go for web development in 2025, give this mix of libraries a shot. And hey, if you’ve got questions or want to share your own Go setup, @ me on the socials or contact form – I love talking shop.&lt;/p&gt;&#xA;&lt;p&gt;Happy Coding!&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>golang web development</category><category>go templ</category><category>golang reload</category><category>web development</category><guid isPermaLink="true">https://www.ersin.nz/articles/a-great-web-stack-for-go</guid></item><item><title>Zero Drama deployments with caprover</title><link>https://www.ersin.nz/articles/zero-drama-deployments</link><description>Looking for a reliable, low-maintenance blog setup that won&#39;t break the bank? Here&#39;s how I built a robust solution that costs just $5 a month. It&#39;s enterprise-grade features without the complexity or price tag.</description><pubDate>26 Jan 2025 21:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;h1&gt;How this blog is deployed&lt;/h1&gt;&#xA;&lt;p&gt;Looking for a reliable, low-maintenance blog setup that won&#39;t break the bank? Here&#39;s how I built a robust solution that costs just $5 a month. It&#39;s enterprise-grade features without the complexity or price tag.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Quick Stats:&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;💰 Monthly Cost: $5 (Digital Ocean)&lt;/li&gt;&#xA;&lt;li&gt;🔧 Maintenance: Minimal (automated updates via cron)&lt;/li&gt;&#xA;&lt;li&gt;🚀 Deploy Time: Under 2 minutes (CapRover + Github Actions)&lt;/li&gt;&#xA;&lt;li&gt;🛠 Tech Stack: Docker + Go&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;&lt;strong&gt;What it gets me&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-text&#34;&gt;✅ Zero-downtime deployments (your content stays live, always)&#xA;✅ Automatic SSL certificate management (no more security headaches)&#xA;✅ One-click deployments (deploy new features in minutes, not hours)&#xA;✅ Simple rollbacks (instant recovery if something goes wrong)&#xA;✅ Total monthly cost less than a fancy coffee&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;&lt;strong&gt;Why this setup?&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;After years of wrestling with over-engineered solutions, I&#39;ve learned that simplicity wins in the long run. While platforms like Kubernetes or full-blown cloud services are fantastic for enterprise needs, they&#39;re often overkill for personal projects or small-to-medium size business applications. What I love about this stack is that each component does exactly one thing, and does it well. CapRover handles container management without the complexity of Kubernetes, GitHub Actions gives us reliable CI/CD without expensive build servers, and Go creates tiny and efficient binaries that deploy in seconds. The best part? If something breaks, you can actually understand and fix it because the whole stack fits in your head and you can easily understand every layer.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Perfect for:&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Small to mid-sized IT teams&lt;/li&gt;&#xA;&lt;li&gt;Companies looking to reduce cloud costs&lt;/li&gt;&#xA;&lt;li&gt;Teams that need enterprise features without enterprise complexity&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;&lt;strong&gt;Not ideal for:&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Enterprise-scale deployments (10,000+ concurrent users)&lt;/li&gt;&#xA;&lt;li&gt;Teams requiring extensive compliance certifications&lt;/li&gt;&#xA;&lt;li&gt;Organizations with complex microservices architectures&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h1&gt;The VM&lt;/h1&gt;&#xA;&lt;p&gt;I am running a very boring ubuntu LTS on a $5 digital ocean droplet. They are affordable, and you can easily replace this layer with any other cloud provider. You just need a static IP for routing your A records from the domain and you are pretty much good to go.&lt;/p&gt;&#xA;&lt;p&gt;Maintenance of the VM is pretty straight forward with cron jobs and automatic updates turned on.&lt;/p&gt;&#xA;&lt;h1&gt;The platform&lt;/h1&gt;&#xA;&lt;p&gt;CapRover is my platform kubernetes alternative. You can think of it like a self hosted single node heroku.  It is perfect for the developer who just has a few lightly used side projects to host.&lt;/p&gt;&#xA;&lt;p&gt;A cool thing you can do with CapRover is 1-click deploys. These are little blueprints you can apply to your instance which will spin up databases, or other open source self hosted software. I use this for experiments that require development databases, redis instances.&lt;/p&gt;&#xA;&lt;p&gt;To deploying your own code I recommend wrapping your app in a docker image then  leveraging the github container registry (GHCR) and the docker image deployment in CapRover. Basically, you just instruct it to pull a specific image tag and it will manage a zero downtime roll out of the new version of the service.&lt;/p&gt;&#xA;&lt;p&gt;CapRover has handy things built in like automatic https using let&#39;s encrypt, so you never need to worry about manual updates of tls certificates.&lt;/p&gt;&#xA;&lt;h1&gt;CI/CD&lt;/h1&gt;&#xA;&lt;p&gt;The Github Action for deploying is really simple, and at this point about 90% of the developers in the world can use it. Let’s go deep and look at &lt;em&gt;exactly&lt;/em&gt; how it works. It’s so simple we can describe the whole thing in this post.&lt;/p&gt;&#xA;&lt;h2&gt;The boilerplate&lt;/h2&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-yaml&#34;&gt;name: deploy&#xA;run-name: &amp;quot;${github.actor} ${github.run_number} - deploy to captain rover&amp;quot;&#xA;on:&#xA;  push:&#xA;    branches:&#xA;      - main&#xA;jobs:&#xA;  build_and_deploy:&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Our build starts with the basic instructions around creating a action, we have a name, a trigger, and one job &lt;code&gt;build_and_deploy&lt;/code&gt;.&lt;/p&gt;&#xA;&lt;h2&gt;The setup&lt;/h2&gt;&#xA;&lt;p&gt;Once we are in our build_and_deploy job, we then do all the steps required to setup an envrionment for our code and build. This means logging in to the container registry, setting up docker and checking out the code.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-yaml&#34;&gt;    runs-on: ubuntu-latest&#xA;    steps:&#xA;      - name: Check out repository&#xA;        uses: actions/checkout@v2&#xA;      - name: Set up Docker Buildx&#xA;        uses: docker/setup-buildx-action@v1&#xA;      - name: Login to Container Registry&#xA;        uses: docker/login-action@v2&#xA;        with:&#xA;          registry: ghcr.io&#xA;          username: ${{ github.repository_owner }}&#xA;          password: ${{ secrets.GITHUB_TOKEN }}&#xA;      - name: &amp;quot;setup Go&amp;quot;&#xA;        uses: actions/setup-go@v2&#xA;        with:&#xA;          go-version: &#39;&amp;gt;=1.21&#39;&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Note: There are a few variable replacements above for passing in the built in &lt;code&gt;GITHUB_TOKEN&lt;/code&gt; secret for the github container registry login.&lt;/p&gt;&#xA;&lt;h2&gt;Lint and test&lt;/h2&gt;&#xA;&lt;p&gt;So far we have been mostly boilerplating things that every golang ci/cd workflow needs. Now that the environment is ready, we can use the following two steps for running a lint process and our tests.&lt;/p&gt;&#xA;&lt;p&gt;Golangci-lint is a nice little additional golang linter which I have used at work, and really like to use on my own projects. It has sensible defaults and enforces rules for idiomatic looking golang projects.&lt;/p&gt;&#xA;&lt;p&gt;Running tests is done next using the basic and simple &lt;code&gt;go test -v ./...&lt;/code&gt; command. One of the go languages massive strengths is that you have a built in test runner that is pretty much just good enough.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-yaml&#34;&gt;      - name: golangci-lint&#xA;        uses: golangci/golangci-lint-action@v3&#xA;        with:&#xA;          version: v1.54&#xA;      - name: &amp;quot;test&amp;quot;&#xA;        run: |&#xA;          go test \-v ./...&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;h2&gt;Build&lt;/h2&gt;&#xA;&lt;p&gt;Assuming our build has made it this far, we should have a tested/linted bit of software, we can then build the binary and push the docker image.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-yaml&#34;&gt;      - name: build all the things with the current git sha&#xA;        run: |&#xA;          CGO_ENABLED=0 go build -ldflags &amp;quot;-X main.GitCommit=$(git rev-parse --short HEAD)&amp;quot; -o build/ ./...&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;For the build command, I&#39;m also assigning the GitCommit hash to a constant in the code so that we can print out a log with the git commit. Including the current commit in the code is super handy for tracking down changes between deployments. When you are searching very old logs you can then correlate older code releases with the log event. The bit which is doing this assignment below is the &lt;code&gt;-ldflags &amp;quot;-X main.GitCommit=...&lt;/code&gt; part.&lt;/p&gt;&#xA;&lt;p&gt;You can then use code like the following listing to expose that git hash&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-go&#34;&gt;package main&#xA;// GitCommit will contain the output of $(git rev-parse \--short HEAD)  &#xA;const GitCommit string&#xA;&#xA;func main() {&#xA;    println(&amp;quot;Hello!&amp;quot;, GitCommit)&#xA;}&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;h2&gt;Deploy&lt;/h2&gt;&#xA;&lt;p&gt;A three step process is taken for deploying the docker image. First is generating the image url using the repo owner, name and the first 6 characters of the hash. Second is to do the docker build and push the image with the relevant tag. Finally we make an API call to caprover for triggering a deploy.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-yaml&#34;&gt;      - name: Preset Image Name&#xA;        run: echo &amp;quot;IMAGE\_URL=$(echo ghcr.io/${{ github.repository_owner }}/${{ github.event.repository.name }}:$(echo ${{ github.sha }} | cut \-c1-7) | tr &#39;\[:upper:\]&#39; &#39;\[:lower:\]&#39;)&amp;quot; &amp;gt;&amp;gt; $GITHUB_ENV&#xA;      - name: Build and push Docker Image&#xA;        uses: docker/build-push-action@v4&#xA;        with:&#xA;          context: .&#xA;          file: ./Dockerfile&#xA;          push: true&#xA;          tags: ${{ env.IMAGE_URL }}&#xA;      - name: Deploy Image to CapRrover&#xA;        uses: caprover/deploy-from-github@4f2b50c37be9f3f325c67b16660e321395841040&#xA;        with:&#xA;          server: &amp;quot;${{ vars.CAPROVA_INSTANCE }}&amp;quot;&#xA;          app: &amp;quot;${{ vars.APP_NAME }}&amp;quot;&#xA;&#xA;          token: &amp;quot;${{ secrets.DEPLOY_TOKEN }}&amp;quot;&#xA;&#xA;          image: &amp;quot;${{ env.IMAGE_URL }}&amp;quot;&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;&lt;strong&gt;Generating the image name&lt;/strong&gt; When you append to the &lt;code&gt;$GITHUB_ENV file&lt;/code&gt;, then the environment variables in there will be available to subsequent steps in the file. This is how we set the &lt;code&gt;IMAGE_URL&lt;/code&gt; above, by using &lt;code&gt;echo&lt;/code&gt; and &lt;code&gt;&amp;gt;&amp;gt;&lt;/code&gt;. A bit of bash foo is used to turn the github.sha into a shorter version and converting the upper case letters to lowercase.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Building the docker image&lt;/strong&gt;&#xA;The dockerfile is incredibly simple for this blog, it simply copies the built binary into a distroless base. Behold the crazy simple 5 line docker file.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-Dockerfile&#34;&gt;FROM gcr.io/distroless/static&#xA;&#xA;WORKDIR /&#xA;&#xA;COPY build/server /&#xA;&#xA;EXPOSE 80&#xA;&#xA;ENTRYPOINT [&amp;quot;./server&amp;quot;]&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;The private docker image is then pushed to CapRover, and our deployment is ready to roll out on the server.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Pushing the image to prod&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;CapRover has an API which can be used to trigger new deployments. Using a helpful github action I can call this API to kick things off. All going well, the new deployment should happen within seconds of it being triggered.&lt;/p&gt;&#xA;&lt;h1&gt;What’s good&lt;/h1&gt;&#xA;&lt;p&gt;It’s simple and fast. Being simple isn’t always easy when you want some of the features like docker deployment and automatic https and zero downtime deployments. The approach is fast enough for my current stage and scale. Sure – you wouldn’t want this if you are a growth start up with millions of customers. This is a simple AND fast solution that works really well for a personal site.&lt;/p&gt;&#xA;&lt;p&gt;Zero maintenance, it just runs in the background. I have been very pleased with the amount of admin overhead on this solution. Once it is all setup it’s very easy to keep running, and the deployment pattern is simple enough that I know I can get any dockerized solution released under a domain in less than 30 minutes effort.&lt;/p&gt;&#xA;&lt;p&gt;I can understand how it all works because I built it from the operating system up. People severely underrate the ability to understand how software gets in to production. In big kubernetes deployments you often have whole teams responsible for the operations. If you are using something like heroku or vercel, good luck understanding exactly how your software is running and the environment it is deployed in.&lt;/p&gt;&#xA;&lt;h1&gt;What’s next&lt;/h1&gt;&#xA;&lt;p&gt;Go, templ and htmx have been 0 regret choices for running this application and I plan to share a bit more about how it works soon. It is because of these choices I can have teeny tiny docker images and super fast cache-able websites. I think this platform will last me for the next “&lt;a href=&#34;https://www.ersin.nz/articles/10-year-blog-platform&#34;&gt;100 years&lt;/a&gt;” still, and I’m excited to see it evolve as a place where I experiment and build something cool&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><guid isPermaLink="true">https://www.ersin.nz/articles/zero-drama-deployments</guid></item><item><title>Build a local first sync engine with AI</title><link>https://www.ersin.nz/articles/rust-crdt-sync-engine</link><description>I just wrapped up an interesting project: building a CRDT (Conflict-free Replicated Data Type) backend in Rust with a frontend sync engine. The cool part? I did most of it through Windsurfer editor&#39;s compose feature. Here&#39;s how it went.</description><pubDate>11 Jan 2025 21:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;I created a dead simple Conflict Free Replicated Datatype (CRDT) backend and sync engine using plain old Rust and React. This turned out to be a fun experiment in using the windsurf IDE and chat oriented programming (CHOP).&lt;/p&gt;&#xA;&lt;p&gt;&lt;a href=&#34;https://github.com/ebuckley/rust-local&#34;&gt;Check out the code here&lt;/a&gt;&lt;/p&gt;&#xA;&lt;p&gt;Not familiar with &lt;strong&gt;CHOP&lt;/strong&gt;? It’s a modern way to put together an application through a combination of prompting and coding. The basic flow is to use your code base as context, and then ask an LLM what to implement. You spend your time writing prompts, a prompt is like a requirement you would write for a junior developer. Reading and integrating code is still needed because it hardly ever works perfectly the first time. There are new IDE’s and plugins that lean into this workflow, and for today’s blog I use the &lt;a href=&#34;https://codeium.com/windsurf&#34;&gt;Windsurf IDE&lt;/a&gt;.&lt;/p&gt;&#xA;&lt;p&gt;The big goal with this project is twofold. First, I wanted to test out CHOP coding using windsurf. Second, I wanted to create a sync engine which allows me to quickly build single page apps that share state between multiple clients.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;What is a CRDT backend and sync engine? A sync engine is a core piece of technology in creating &lt;a href=&#34;https://www.ersin.nz/articles/local-first-software&#34;&gt;local first&lt;/a&gt; software. Previously I looked in to using sqlite compiled to wasm as a state store, but I found this to be quite complex. So I’ve taken a completely different approach with this implementation and built the most simple but useful API possible.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;h1&gt;Keeping the backend Simple&lt;/h1&gt;&#xA;&lt;p&gt;To synchronize any type of state I just need the following three API methods&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Get the initial state from &lt;code&gt;api/bootstrap&lt;/code&gt; this returns ALL the models.&lt;/li&gt;&#xA;&lt;li&gt;Update the state on the backend with some transactions by posting them to &lt;code&gt;api/transactions&lt;/code&gt; . After this API it will increment a transaction counter on the server which indicates other clients should update their own state.&lt;/li&gt;&#xA;&lt;li&gt;Get all the transactions required to update my local state from &lt;code&gt;api/transactions?from=99&lt;/code&gt;. A regular interval will poll the backend to get the latest set of transactions.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;I chose Rust for this backend to better test out CHOP programming. I am no professional rust developer, just a mid coder who has gone through the rustlings and tutorial a couple times. The value in using an LLM based workflow here is that it can help me learn the frameworks and language by building something from scratch.&lt;/p&gt;&#xA;&lt;p&gt;For keeping the state on the backend I used Sqlite, and the LLM chose actix and tokio as the web framework du jour.&lt;/p&gt;&#xA;&lt;p&gt;The experience of chopping my way through creating this project was fairly seamless. It took a few tries to get things compiling and I had to do a bit of manual testing with &lt;code&gt;curl&lt;/code&gt; to verify things are working as expected. The overall implementation is so simple that I don’t feel like this is any great feat of engineering. The coolest part really is just what I was able to achieve in relatively short amounts of time just by prompting and having a tutorial level knowledge of the language.&lt;/p&gt;&#xA;&lt;h2&gt;Keeping the schema simple&lt;/h2&gt;&#xA;&lt;p&gt;I am using sqlite as the database but the data is not relational. My data layer is designed to work with any type of model, so I’m using JSON columns. The schema consists of just two tables, &lt;code&gt;sync_history&lt;/code&gt; and &lt;code&gt;model_data&lt;/code&gt;. We only do simple queries on this database like getting all models of a type, or getting a list of transactions for a client to apply.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-rust&#34;&gt;fn new(path : \&amp;amp;str) \-\&amp;gt; Result\&amp;lt;Self\&amp;gt; {&#xA;    &#x9;let conn \= Connection::open(path)?;&#xA;&#xA;    &#x9;// Create tables&#xA;    &#x9;conn.execute(&#xA;        &#x9;&amp;quot;CREATE TABLE IF NOT EXISTS sync\_history (&#xA;            &#x9;id INTEGER PRIMARY KEY AUTOINCREMENT,&#xA;            &#x9;actions TEXT NOT NULL&#xA;        &#x9;)&amp;quot;,&#xA;        &#x9;\[\],&#xA;    &#x9;)?;&#xA;&#xA;    &#x9;conn.execute(&#xA;        &#x9;&amp;quot;CREATE TABLE IF NOT EXISTS model\_data (&#xA;            &#x9;id TEXT PRIMARY KEY,&#xA;            &#x9;model\_name TEXT NOT NULL,&#xA;            &#x9;data TEXT NOT NULL,&#xA;            &#x9;created\_at INTEGER NOT NULL,&#xA;            &#x9;updated\_at INTEGER NOT NULL&#xA;        &#x9;)&amp;quot;,&#xA;        &#x9;\[\],&#xA;    &#x9;)?;&#xA;&#xA;    &#x9;Ok(Database { conn: Mutex::new(conn) })&#xA;&#x9;}&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;h1&gt;The easiest possible frontend&lt;/h1&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-javascript&#34;&gt;const { items, createItem, updateItem, deleteItem } \= useItems(&#39;Todo&#39;);&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;A component can work with state through the handy &lt;code&gt;useItems&lt;/code&gt; API. You just provide it a name for the type of data you are mutating, and it will seamlessly figure out all the conflict resolution and synchronisation in the background automatically.&lt;/p&gt;&#xA;&lt;p&gt;A simple hook to return the latest data and create, update, delete methods is the best way to interface with a CRDT.&lt;/p&gt;&#xA;&lt;p&gt;The best libraries expose a narrow interface and a deep amount of usefulness. The above API does this extremely well. Under the hood we have a polling sync engine, optimistic updates and a local database implementation.&lt;/p&gt;&#xA;&lt;h2&gt;Synchronizing multiple frontends&lt;/h2&gt;&#xA;&lt;p&gt;The sync engine will poll for updates from the backend. When you apply a change it immediately gets updated in the local database and then queued for sending to the backend when it is online.&lt;/p&gt;&#xA;&lt;p&gt;One common challenge in CRDT algorithms is how you manage two updates to the same model. We take the simple solution here and just choose the last write as the winner. Yea, it’s possible that this isn’t great in the long run, but I reckon it’s more of an application specific concern if you care about that or not. Everyone knows by now that you shouldn’t be representing bank transactions in CRDT’s anyway!&lt;/p&gt;&#xA;&lt;h1&gt;Deploying it&lt;/h1&gt;&#xA;&lt;p&gt;You can access this little bit of software from my caprover instance at &lt;a href=&#34;https://todo-rust.zealot.ersin.nz/&#34;&gt;https://todo-rust.zealot.ersin.nz/&lt;/a&gt; .&lt;/p&gt;&#xA;&lt;p&gt;Caprover is my boring choice in deployment for all my little side hacks. It’s a cool platform that I will write about later in the year, but for now – just know it’s my preferred way to spin up a docker container behind https and nginx.&lt;/p&gt;&#xA;&lt;p&gt;I am very familiar with CapRover/Docker and github actions, but this is my first time  building and deploying a rust program from a github action. For completeness I CHOP’d my way through it. CI/CD is something i love to use, but hate to build so it’s a welcome proposition to have AI do it for me. The overall structure is something pretty standard, a build image plus a deployment image and pushing it all up to GHCR at the end.&lt;/p&gt;&#xA;&lt;h1&gt;Did I learn anything?&lt;/h1&gt;&#xA;&lt;h2&gt;About CHOP?&lt;/h2&gt;&#xA;&lt;p&gt;Are we learning when we use AI Chat to generate all our code? Yeah — nahh – yea?. I ended up sending a lot of my time reading code. You learn when you read code, but I also think there is benefit to building muscle memory by actually writing the code. I’m more familiar with rust/actix than before. But the fundamentals of Rust are probably not as advanced as spending the same time on writing code from scratch.&lt;/p&gt;&#xA;&lt;p&gt;The future of learning to code is not just writing prompts for an AI. The fundamentals of syntax, and language semantics are still going to be important going forward. Yet we still need to be pragmatic and deliver value if we are to be professionals. What this means for people learning new languages is that you should balance your time between CHOPing big projects and also still focusing on the low level fundamentals. Remember you can also use the AI as a guide to explain what’s going on semantically, then experiment on your own terms with the compiler.&lt;/p&gt;&#xA;&lt;h2&gt;About CRDT’s&lt;/h2&gt;&#xA;&lt;p&gt;The sync engine seems solid, but it needs testing under the fire of production use. I am more convinced of the idea that small interfaces and deep implementations are the most useful shape of abstractions. I am not sold on the conflict resolution or polling approach, but it really just needs more applications to be built out to test this.&lt;/p&gt;&#xA;&lt;p&gt;Some things might not be great with this type of sync engine. A lot of relational data is one. A high number of clients and edits would also be quite challenging for this approach. Needing to be extremely real time like a game engine would also fall over.&lt;/p&gt;&#xA;&lt;p&gt;But overall, this was a good experiment and I’m already thinking of&lt;/p&gt;&#xA;&lt;h2&gt;What to take away from this&lt;/h2&gt;&#xA;&lt;h2&gt;Use CHOP, but don’t ignore the fundamentals&lt;/h2&gt;&#xA;&lt;p&gt;As a senior+ engineer, LLMs are powerful tools for prototyping in unfamiliar languages.  The new tools are powerful, but not an alternative for learning the actual language and semantics.&lt;/p&gt;&#xA;&lt;h2&gt;CRDTs can be simple&lt;/h2&gt;&#xA;&lt;p&gt;You can build a functional CRDT without much fuss if you keep it simple. This might be pragmatic and easy to achieve for your application. Getting rid of the complex parts like conflict algorithms and relational data models allowed us to do this.&lt;/p&gt;&#xA;&lt;h2&gt;Tiny interfaces with deep implementations&lt;/h2&gt;&#xA;&lt;p&gt;The best abstractions are small, with deep and complex implementations. These kind of abstractions are the ones you can use across more places, and replace more easily when they are highly depended on. In our simple program here, we have a simple backend interface and a simple react API interface. As a result, it’s something that can have bits swapped out easily.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>rust</category><category>crdt</category><category>sqlite</category><category>llm</category><guid isPermaLink="true">https://www.ersin.nz/articles/rust-crdt-sync-engine</guid></item><item><title>My blog grew by 12% in 2024</title><link>https://www.ersin.nz/articles/2024-blog-review</link><description>In 2024, I proved a theory: Yes, you can grow a tech blog through search traffic alone. With zero active marketing, I saw a 4,509% increase in organic search volume and a 12% increase in active user volume.</description><pubDate>26 Dec 2024 21:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;In 2024, I proved a theory: Yes, you can grow a tech blog through search traffic alone. With zero active marketing, I saw a 4,509% increase in organic search volume and a 12% increase in active user volume.&lt;/p&gt;&#xA;&lt;h2&gt;The numbers say..&lt;/h2&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;12% increase in active users (622)&lt;/li&gt;&#xA;&lt;li&gt;4,509% organic search growth&lt;/li&gt;&#xA;&lt;li&gt;14 posts published&lt;/li&gt;&#xA;&lt;li&gt;🏆  &lt;a href=&#34;https://www.ersin.nz/articles/reflection-is-slow&#34;&gt;reflection is slow&lt;/a&gt;. Is my winning post for 2024.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;&lt;img src=&#34;data:image/png;base64,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&#34; alt=&#34;&#34;&gt;&lt;/p&gt;&#xA;&lt;p&gt;Above, we compare the steady growth of 2024 to the spiky growth in 2023. My ‘small viral’ post about a &lt;a href=&#34;https://www.ersin.nz/articles/markdown-editor-with-wails-react-tailwind&#34;&gt;markdown editor&lt;/a&gt; is the culprit for that blue dotted spike in November 23.&lt;/p&gt;&#xA;&lt;p&gt;Contrast that to this year: you see a steady growth.  I am unreasonably proud of this. I set out to make this blog after 10+ years of not publishing because I wanted to test a theory. What if I published without explicitly marketing my writing, would I get any traction?&lt;/p&gt;&#xA;&lt;p&gt;At the end of 2024, I can report that back. Yes! If you write something useful on the internet, the search engine will index it and serve it to people, who will actually spend time reading your words.&lt;/p&gt;&#xA;&lt;p&gt;Despite not having a post shared on X like in 2023, I still achieved a modest 12% bump in active users.&lt;/p&gt;&#xA;&lt;h3&gt;How do people find my blog?&lt;/h3&gt;&#xA;&lt;p&gt;Behold a ludicrous 4,509% increase in organic search traffic increase in 2024.&lt;/p&gt;&#xA;&lt;p&gt;&lt;img src=&#34;data:image/png;base64,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&#34; alt=&#34;&#34;&gt;&lt;/p&gt;&#xA;&lt;p&gt;What posts got the majority of organic search results?&lt;/p&gt;&#xA;&lt;p&gt;&lt;img src=&#34;data:image/png;base64,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&#34; alt=&#34;&#34;&gt;&lt;/p&gt;&#xA;&lt;p&gt;Talking about how &lt;a href=&#34;https://www.ersin.nz/articles/reflection-is-slow&#34;&gt;reflection is slow&lt;/a&gt; in golang was the main source of new active users. I am quite proud of this post and the underlying work, it’s gratifying that it also seems to be good for the metrics&lt;/p&gt;&#xA;&lt;h1&gt;Key Insights&lt;/h1&gt;&#xA;&lt;p&gt;The data tells a clear story: while I write about various technologies, and I’ve been spending a lot of my side-project elsewhere, Golang content consistently drives my organic traffic. My niche is at the intersection of Go and emerging technologies.&lt;/p&gt;&#xA;&lt;p&gt;People are still using search engines for finding content. You can publish your writing on the internet and it &lt;em&gt;will&lt;/em&gt; eventually be picked up if it is useful. 2024 is still a year where organic search reigns supreme. The future may be uncertain as LLM’s increasingly become our entrypoint to the internet – but for now we don’t need to do Chatbot engine optimization on our content!&lt;/p&gt;&#xA;&lt;p&gt;Good content still has longevity. My post about crafting a markdown editor is number two on my organic search results, even though it was published in October 2025.  I had ‘small viral’ effects from this post but it has paid off on a longer term. In the future I hope to write more things that can hit the short term and long term pay off models of viral vs organic search acquisition.&lt;/p&gt;&#xA;&lt;h1&gt;Goals for 2025 and this blog?&lt;/h1&gt;&#xA;&lt;p&gt;What worked for me in 2024 was simple: write something useful consistently. Building on this my 2025 goals are:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Ship 25 blog posts in 2025&lt;/li&gt;&#xA;&lt;li&gt;Do at least one marketing &lt;em&gt;activity&lt;/em&gt; per blog post.&lt;/li&gt;&#xA;&lt;li&gt;Increase my active Users by 25% compared to the previous year.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;2025 is about doubling down on what is working and experimenting with marketing. If I have written 25 posts, and don’t see the increase in active users, I won’t regret it.  This is a decision with no downside.&lt;/p&gt;&#xA;&lt;p&gt;I am a writer, and this is where I am honing the craft. My theory is that the long term benefits of working on writing will pay off in the long run.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>blog</category><category>review</category><category>2024</category><category>seo</category><guid isPermaLink="true">https://www.ersin.nz/articles/2024-blog-review</guid></item><item><title>An AI copilot review for 2024</title><link>https://www.ersin.nz/articles/2024-ai-copilot-review</link><description>I had a fun afternoon building this little LLM tool which lets me chat with gemini-1.5 from the command line. You should be checking this out if you are a developer. The context window is massive and you can process so much more than you could with the first iteration of LLM apis.</description><pubDate>24 Nov 2024 21:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;If you write code, you&#39;re a programmer. If you start ventures, you&#39;re an entrepreneur. Simple, right? But what are you when you&#39;re obsessed with experimenting with every new framework that hits Hacker News?&lt;/p&gt;&#xA;&lt;p&gt;I&#39;ve spent over a decade being that person – the eternal tinkerer. Back in university, I thought chasing every trending technology would give me an edge. Now, with the rise of AI and LLMs, that same curiosity has evolved into something far more powerful.&lt;/p&gt;&#xA;&lt;p&gt;Here&#39;s the thing: While Xitter is flooded with grifter predictions about AI replacing developers, I&#39;m seeing something different. The new tools are insanely awesome for starting new projects.  After building documentation platforms, Xero tools and new backend frameworks, I&#39;ve discovered that AI isn&#39;t our replacement – it&#39;s our pocket junior developer. One that never learns, often surprises, and completely changed my approach to building prototypes&lt;/p&gt;&#xA;&lt;h2&gt;Towards higher and higher levels of code&lt;/h2&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-plaintext&#34;&gt;Bits in memory                                       Customer pain&#xA;|----------------------------------------------------------------|&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Solving problems with computers has been shifting right from low level abstractions over hardware to high level conceptual solutions. Hardware solutions were replaced with general purpose languages. Languages became business oriented (COBOL). Computers got fast enough for dynamic languages. Platforms for massive distribution became a thing (web, phones).&lt;/p&gt;&#xA;&lt;p&gt;In October 2021 the next big thing happened, the release of github copilot.  After a couple years with this AI sidekick, I think we can clearly see this as yet another shift from lower to higher levels of abstractions.&lt;/p&gt;&#xA;&lt;p&gt;What does this mean? More than ever, it’s important to take on a senior mindset. Coding with AI by your side is like working with a really terrible Junior engineer that never learns. Sure, you can type a lot less lines of code, but get ready for some pretty weird interpretations of requirements. Your challenge becomes structuring your system in such a way that you need less code. Code is so easy to generate you need to push against this influence of the coding copilot. Less code to solve a problem is always better than more code.&lt;/p&gt;&#xA;&lt;p&gt;Really understanding the problem you are solving has always been important. Now understanding the problem might be the only thing left for us to do.&lt;/p&gt;&#xA;&lt;h2&gt;AI is great for exploration&lt;/h2&gt;&#xA;&lt;pre&gt;&lt;code&gt;Exploration                                                Legacy&#xA;|---------------------------------------------------------------|&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Grifters on the internet will tell you that AI is replacing us, that you should buy their course and change careers. New AI tools can be great for exploring a problem space and building prototypes. The grifters are wrong in general.. When software actually becomes more valuable and legacy, then the effectiveness and usefulness of AI falls off a cliff.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Software runs through this lifecycle forever&lt;/strong&gt;&lt;br&gt;&#xA;Starting something new, the most important thing to do is get your solution in front of the people who are trying to do the job with it. Things move on, and code runs for months and years, it shifts towards a state of optimization and happy legacy. Once that legacy becomes too difficult to maintain, and everyone who worked on it is dead a slow rebuild happens. You pull out threads from the wonderful legacy system and build little components to explore solving the problem again.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Where does current AI fit into the paradigm?&lt;/strong&gt;&lt;br&gt;&#xA;Copilots and agents for coding are insanely biassed towards the getting started side of the spectrum.  If you point an AI agent at a legacy codebase you are in for a surprise. You get a confident tone and a completely wrong solution&lt;/p&gt;&#xA;&lt;p&gt;Frameworks have a similar value lifecycle. In the  beginning they provide guard rails to your development. They prevent solving unimportant problems like authentication. Within a year, most of that value is gone. You start to build your own abstractions. After 5 years, that framework is probably completely unrecognisable from an application written with a custom framework.&lt;/p&gt;&#xA;&lt;p&gt;Software in the exploration phase may have great potential for value, but this will never get close to the value of a legacy system. A back office cobol program that clears and processes transactions between banks will be more impactful on a bottom line than another react native app for showing your balance.&lt;/p&gt;&#xA;&lt;h2&gt;The future is exciting&lt;/h2&gt;&#xA;&lt;p&gt;This old curmudgeon of a developer is enthusiastically trying all of the AI tools. Just like 10 years ago, I keep a keen eye on hacker news for the latest and greatest. My scepticism levels are higher than normal, and I think the most important problem to solve for AI tool builders is to convince me they do not suck at legacy software.&lt;/p&gt;&#xA;&lt;p&gt;If you are just entering the industry, or thinking about it.. Heed my advice. Take up the AI tools, and experiment with the new. But also, do not take serious account of AI grifters and doomers who might say that the end of our industry is just around the corner. We can have fulfilling and fun careers building really valuable software with copilots and agents at our side.&lt;/p&gt;&#xA;&lt;p&gt;Technological progress has not been a constant in the world, but throughout my life and yours it will be. We either ride the wave, or we take up new careers. Join me on the adventure, I really look forward to what 2025 and beyond holds for our industry.&lt;/p&gt;&#xA;&lt;p&gt;As always, if you’re interested in semi-non-regular emails about my thinking, you can sign up to the RSS feed on this site or put your email in the box below for updates.&lt;/p&gt;&#xA;&lt;p&gt;Peace &amp;amp; happy coding.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>beginner</category><category>ai</category><category>llm</category><category>golang</category><guid isPermaLink="true">https://www.ersin.nz/articles/2024-ai-copilot-review</guid></item><item><title>BFF is a tool for making user interfaces</title><link>https://www.ersin.nz/articles/bff</link><description>Tired of wrestling with HTML, CSS, and JavaScript just to create a simple user interface? Meet BFF, the Go developer&#39;s new best friend for backend flows.</description><pubDate>21 Oct 2024 21:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;h1&gt;Bff is a tool for making user interfaces&lt;/h1&gt;&#xA;&lt;p&gt;Tired of wrestling with HTML, CSS, and JavaScript just to create a simple user interface? Meet &lt;a href=&#34;http://github.com/ebuckley/bff&#34;&gt;github.com/ebuckley/bff&lt;/a&gt; , the Go developer&#39;s new best friend for backend flows.&lt;/p&gt;&#xA;&lt;p&gt;BFF is a powerful library that lets you build functional UIs and forms without writing a single line of frontend code. It&#39;s designed for Gophers who need quick, efficient internal tools without the hassle of learning new frameworks.&lt;/p&gt;&#xA;&lt;p&gt;Key features:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Create multi-step forms and data pages with pure Go&lt;/li&gt;&#xA;&lt;li&gt;Built-in websocket protocol for real-time updates&lt;/li&gt;&#xA;&lt;li&gt;Suite of react components for display and input capturing&lt;/li&gt;&#xA;&lt;li&gt;Focus on your logic, not on vite webpack and fotm JS libraries.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;Let&#39;s dive into how BFF can streamline your development process and save you valuable time.&lt;/p&gt;&#xA;&lt;h1&gt;Write Go, get a free user interface&lt;/h1&gt;&#xA;&lt;p&gt;Our example will be a process that updates &lt;a href=&#34;https://docs.google.com/spreadsheets/d/1Uol7hBRk65_FVs5Z13WeZ2rJ2yv5z-OHKhi33fo2yIw/edit?gid=0#gid=0&#34;&gt;this&lt;/a&gt; Google Sheet. You can try out this exact code which I have integrated into the site &lt;a href=&#34;/backend/a/update-a-google-sheet&#34;&gt;here&lt;/a&gt;.&lt;/p&gt;&#xA;&lt;h2&gt;Creating the handler&lt;/h2&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-go&#34;&gt;func(ctx context.Context, io *bff.Io) error {&#xA;&#x9;&#x9;io.Display.Markdown(`&#xA;# Update a Google Sheet&#xA;&#xA;This little action exists to show you how you can update a Google Sheet from a BFF action.`)&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Here, we set up our handler which simply takes an &lt;code&gt;bff.Io&lt;/code&gt;. The Io struct is our portal into the UI, it creates components on the screen and lets us capture input from the person on the website. The interface for BFF  is inspired by the &lt;code&gt;http.Handler&lt;/code&gt; we know and love, but better because you can just return an error if you want to end execution.&lt;/p&gt;&#xA;&lt;h2&gt;Capturing input&lt;/h2&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-go&#34;&gt;    row, err := io.Input.Number(&amp;quot;What Row would you like to update?&amp;quot;)&#xA;    if err != nil {&#xA;        return err&#xA;    }&#xA;    col, err := io.Input.Text(&amp;quot;What Column would you like to update?&amp;quot;)&#xA;    if err != nil {&#xA;        return err&#xA;    }&#xA;    val, err := io.Input.Text(&amp;quot;What value would you like to set?&amp;quot;)&#xA;    if err != nil {&#xA;        return err&#xA;    }&#xA;    io.Display.Markdown(fmt.Sprintf(&amp;quot;Updating row %d, column %s with value %s&amp;quot;, row, col, val))&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;The &lt;code&gt;Input&lt;/code&gt;  struct encapsulates a bunch of handy form components. Right now you can choose from &lt;code&gt;text&lt;/code&gt;,&lt;code&gt;textarea&lt;/code&gt;,&lt;code&gt;number&lt;/code&gt;,&lt;code&gt;email&lt;/code&gt;,&lt;code&gt;datetime&lt;/code&gt; and &lt;code&gt;slider&lt;/code&gt; variants.&lt;/p&gt;&#xA;&lt;h2&gt;All the ceremony&lt;/h2&gt;&#xA;&lt;pre&gt;&lt;code&gt;      keyPath := os.Getenv(&amp;quot;GOOGLE_API_KEY_PATH&amp;quot;)&#xA;      if len(keyPath) == 0 {&#xA;          return fmt.Errorf(&amp;quot;GOOGLE_API_KEY_PATH is not set&amp;quot;)&#xA;      }&#xA;      keyJSON, err := os.ReadFile(keyPath)&#xA;      if err != nil {&#xA;          return err&#xA;      }&#xA;&#xA;      srv, err := sheets.NewService(ctx, option.WithCredentialsJSON(keyJSON))&#xA;      if err != nil {&#xA;          return err&#xA;      }&#xA;      spreadsheetID := &amp;quot;1Uol7hBRk65_FVs5Z13WeZ2rJ2yv5z-OHKhi33fo2yIw&amp;quot;&#xA;      writeRange := fmt.Sprintf(&amp;quot;Sheet1!%s%d&amp;quot;, col, row)&#xA;      vr := &amp;amp;sheets.ValueRange{&#xA;          Values: [][]interface{}{{val}},&#xA;      }&#xA;&#xA;      // Perform the update&#xA;      _, err = srv.Spreadsheets.Values.Update(spreadsheetID, writeRange, vr).&#xA;          ValueInputOption(&amp;quot;RAW&amp;quot;).Do()&#xA;      if err != nil {&#xA;          return err&#xA;      }&#xA;&#xA;      io.Display.Markdown(&amp;quot;# 🍬 Updated the Google Sheet 🍬&amp;quot;)&#xA;      io.Display.Link(&amp;quot;Do another update?&amp;quot;, &amp;quot;/backend/a/update-a-google-sheet&amp;quot;)&#xA;&#xA;      return nil&#xA;   }&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;All the rest of this is about setting up and making the API call for updating the google sheet. This is a good thing. We want our handlers to not be about UI framework but about doing the work that’s actually delivering value. Finally we let the user know that the process is complete and invite them to do it again by showing a link.&lt;/p&gt;&#xA;&lt;h2&gt;All wrapped up in a websocket&lt;/h2&gt;&#xA;&lt;p&gt;The run of this entire handler function is within a single websocket connection, as soon as you call a function on the &lt;code&gt;IO&lt;/code&gt; it will update the user interface. If you need to run a slow query – that’s totally fine the display function will be waiting to show your content as soon as it is ready for the user.&lt;/p&gt;&#xA;&lt;h1&gt;Do and Do not&lt;/h1&gt;&#xA;&lt;p&gt;Do use this for quickly standing up UI for your internal team to run processes and automations. Don’t run it for the public like I’m doing in this example. This thing is designed to be the retool alternative that doesn’t require any learning and clicking in a new UI.&lt;/p&gt;&#xA;&lt;p&gt;Here’s a quick list of ideas you could use bff for today.&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Build quick and helpful tools for your sales team to manage trials and feature flags.&lt;/li&gt;&#xA;&lt;li&gt;Let the CMO run that big dirty SQL script with custom parameters which spits out a customer report from the database.&lt;/li&gt;&#xA;&lt;li&gt;Quickly ban and unban users.&lt;/li&gt;&#xA;&lt;li&gt;Publish notes/banners on the frontend&lt;/li&gt;&#xA;&lt;li&gt;Send a quick email to a set of the existing customer base.&lt;/li&gt;&#xA;&lt;li&gt;Run a script to spin up a new cloud environment.&lt;/li&gt;&#xA;&lt;li&gt;Break-in-case-of-emergency button to call all the developers at once.&lt;/li&gt;&#xA;&lt;li&gt;Process a customer refund&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h1&gt;What’s next?&lt;/h1&gt;&#xA;&lt;p&gt;It is early days for this little library inspired by interval. I’m really interested in helping people like you get started with it.&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Get the code &lt;a href=&#34;http://github.com/ebuckley/bff&#34;&gt;github.com/ebuckley/bff&lt;/a&gt; .  Feel free to create issues or ask questions through github.&lt;/li&gt;&#xA;&lt;li&gt;Sign up to my list vv I email infrequently, but would love to keep in touch with the people that read and appreciate this blog.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>golang</category><category>frontend</category><category>js</category><category>beginners</category><guid isPermaLink="true">https://www.ersin.nz/articles/bff</guid></item><item><title>Having fun with LLM tools</title><link>https://www.ersin.nz/articles/llm-tools</link><description>I had a fun afternoon building this little LLM tool which lets me chat with gemini-1.5 from the command line. You should be checking this out if you are a developer. The context window is massive and you can process so much more than you could with the first iteration of LLM apis.</description><pubDate>11 Oct 2024 21:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;pre&gt;&lt;code&gt;cat somebook.txt | llm \-q “Write a summary of each chapter, 3 bullet points per chapter”&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Context windows are crazy long these days. I wrote a little utility that pipes stdin plus your question into an LLM. The result is a tool that can query 50 minute transcript files or 132 page books. I was planning on making a genius RAG system, but it turned out to be all in vain. What I learned is to just chill out and enjoy the long context window.&lt;/p&gt;&#xA;&lt;h1&gt;How I was planning to RAG this problem&lt;/h1&gt;&#xA;&lt;p&gt;Retrieval augmented generation (RAG) is a basic technique employed for bringing in context to an LLM Query. Basically, you build a search system that can return chunks of relevant information, then give the LLM the ability to call a function which can search for and add context to its query. As a result you will have references in your response and the AI system will give better answers and be able to pull from a very large set of sources, more than it would be able to understand directly in its context window.&lt;/p&gt;&#xA;&lt;p&gt;I wanted to start asking questions of a 48 minute video transcript. Naturally I assumed this will be way too much content for an LLM to provide decent results for. As a fun exercise I decided to experiment with an algorithm for chunking up the data for a RAG system to pull from.&lt;/p&gt;&#xA;&lt;p&gt;My Idea was:&lt;/p&gt;&#xA;&lt;ol&gt;&#xA;&lt;li&gt;Read the next 20 sentences&lt;/li&gt;&#xA;&lt;li&gt;Ask the LLM to pull out a few bullet points&lt;/li&gt;&#xA;&lt;li&gt;Read the next 20 sentences&lt;/li&gt;&#xA;&lt;li&gt;Ask the LLM to summarise the next 20 sentences, merging concepts with the existing bullet point list.&lt;/li&gt;&#xA;&lt;li&gt;???&lt;/li&gt;&#xA;&lt;li&gt;Profit&lt;/li&gt;&#xA;&lt;/ol&gt;&#xA;&lt;p&gt;After looping through steps 1-4 about 100 times, we did actually get a set of bullet points out of the LLM. You could look through the output and see a full history of points. The algorithm &lt;em&gt;kind of&lt;/em&gt; did an ok job of summarising the content.&lt;/p&gt;&#xA;&lt;p&gt;The problem was that the &lt;strong&gt;meaning&lt;/strong&gt; really got lost along the way. Points would get repeated for a number of iterations, then all of a sudden be dropped – even though they should have never been dropped. The LLM was losing context and not weighting the important points correctly.&lt;/p&gt;&#xA;&lt;p&gt;Would this approach eventually work? Incremental improvements, yep. My prompt was far from optimal. If you looked at the unique points in the final list (maybe 50-100 sentences), you could get the meaning of the text, but it was again way too duplicate. A secondary layer of vector indexing (search for context) on the key bullet points might have been a cool experiment to enhance the result.&lt;/p&gt;&#xA;&lt;h1&gt;I started to love the context window&lt;/h1&gt;&#xA;&lt;p&gt;It was not long after the first dumb attempt of using RAG that I came up with an even dumber idea. Just give the entire transcript with no preprocessing to the LLM and ask the question directly.&lt;/p&gt;&#xA;&lt;p&gt;I decided to do this after looking at the spec for &lt;code&gt;gemini-1.5-flash&lt;/code&gt; and seeing that it supports an input count of just over 1 million tokens. WOW! Surprisingly you can just ignore fancy splitting and summarising techniques to just give it the whole content, and that amount of content can be larger than you think if you’ve spent the last 6 months not following LLM development.&lt;/p&gt;&#xA;&lt;p&gt;The 48 minute transcript I passed to the model turned out to be about 400k input tokens, so there was plenty of space for more. I was able to ask simple questions about the content and get an answer back.&lt;/p&gt;&#xA;&lt;h1&gt;My tiny LLM tool&lt;/h1&gt;&#xA;&lt;p&gt;I created a command line &lt;code&gt;llm&lt;/code&gt; binary, that can take input from stdin with a parameter to provide additional question input. No fancy prompcraft or engineering was needed.  With just a simple input + question. I was able to chat with video transcripts, 132 page books. The whole thing has been an amazing little experiment. Now I can quickly ask a question from the terminal, or give it a code file to tell me something about. Given the huge context window you can just chill out and not worry about clever chunking or RAG algorithms.&lt;/p&gt;&#xA;&lt;p&gt;Here is the entire code listing:&#xA;package main&lt;/p&gt;&#xA;&lt;p&gt;import (&#xA;&amp;quot;context&amp;quot;&#xA;&amp;quot;flag&amp;quot;&#xA;&amp;quot;fmt&amp;quot;&#xA;&amp;quot;go-agent/llm&amp;quot;&#xA;&amp;quot;io&amp;quot;&#xA;&amp;quot;log/slog&amp;quot;&#xA;&amp;quot;os&amp;quot;&#xA;&amp;quot;strings&amp;quot;&#xA;)&lt;/p&gt;&#xA;&lt;p&gt;var (&#xA;keyPath  = flag.String(&amp;quot;key&amp;quot;, mustStr(os.UserHomeDir)+&amp;quot;/key&amp;quot;, &amp;quot;Path to the API key file&amp;quot;)&#xA;question = flag.String(&amp;quot;q&amp;quot;, &amp;quot;&amp;quot;, &amp;quot;Question to ask the model&amp;quot;)&#xA;)&lt;/p&gt;&#xA;&lt;p&gt;func main() {&#xA;flag.Parse()&lt;/p&gt;&#xA;&lt;p&gt;// read input from stdin&#xA;ctx := context.Background()&#xA;b, err := os.ReadFile(*keyPath)&#xA;if err != nil {&#xA;slog.Error(&amp;quot;Error reading key file&amp;quot;, &amp;quot;err&amp;quot;, err)&#xA;os.Exit(1)&#xA;}&#xA;chat, err := llm.NewClientFromKeyfile(ctx, string(b))&#xA;if err != nil {&#xA;slog.Error(&amp;quot;Error creating genai client&amp;quot;, &amp;quot;err&amp;quot;, err)&#xA;os.Exit(1)&#xA;}&#xA;stat, err := os.Stdin.Stat()&#xA;if err != nil {&#xA;slog.Error(&amp;quot;Error getting stdin stat&amp;quot;, &amp;quot;err&amp;quot;, err)&#xA;os.Exit(1)&#xA;}&#xA;isPipe := (stat.Mode() &amp;amp; os.&lt;em&gt;ModeCharDevice&lt;/em&gt;) == 0&lt;/p&gt;&#xA;&lt;p&gt;var input []byte&lt;/p&gt;&#xA;&lt;p&gt;if isPipe {&#xA;input, err = io.ReadAll(os.Stdin)&#xA;if err != nil {&#xA;slog.Error(&amp;quot;Error reading stdin&amp;quot;, &amp;quot;err&amp;quot;, err)&#xA;os.Exit(1)&#xA;}&lt;/p&gt;&#xA;&lt;p&gt;}&#xA;fullQuestion := string(input) + &amp;quot;\n&amp;quot; + *question&#xA;slog.Debug(&amp;quot;Read input&amp;quot;, &amp;quot;input&amp;quot;, fullQuestion)&#xA;slog.Info(&amp;quot;Sending message to model&amp;quot;)&#xA;msg, err := chat.Send(ctx, fullQuestion)&#xA;if err != nil {&#xA;slog.Error(&amp;quot;Error sending message&amp;quot;, &amp;quot;err&amp;quot;, err)&#xA;os.Exit(1)&#xA;}&lt;/p&gt;&#xA;&lt;p&gt;fmt.Println(msg.Text())&#xA;slog.Info(&amp;quot;Tokens used:&amp;quot;, &amp;quot;input&amp;quot;, chat.Tokens.Input, &amp;quot;output&amp;quot;, chat.Tokens.Output)&#xA;}&#xA;func mustStr(s func() (string, error)) string {&#xA;v, err := s()&#xA;if err != nil {&#xA;slog.Error(&amp;quot;Error getting string&amp;quot;, &amp;quot;err&amp;quot;, err)&#xA;os.Exit(1)&#xA;}&#xA;return v&#xA;}&lt;/p&gt;&#xA;&lt;h1&gt;What you should do&lt;/h1&gt;&#xA;&lt;p&gt;If you are a developer, you should definitely be exploring this stuff. You can get a lot of usage out of small command line utilities and the capabilities of LLM’s are rapidly evolving. The development experience with this kind of project is also pretty fun. You will get a really enjoyable non-deterministic output that is a bit surprising sometimes. It reminds me a lot of game development where you sort of play with the program a bit to see how it feels before tapping out your next iteration.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>beginner</category><category>ai</category><category>llm</category><category>golang</category><guid isPermaLink="true">https://www.ersin.nz/articles/llm-tools</guid></item><item><title>What do I want from your new API?</title><link>https://www.ersin.nz/articles/api-design</link><description>Writing a new API product? Here&#39;s a list of what you should care about and the must-have features. Following all this will help you decrease churn, increase my hapiness and decrease support load.</description><pubDate>21 Sep 2024 21:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;blockquote&gt;&#xA;&lt;p&gt;My hands are sweaty as they race across the keyboard. My eyes dart through the stream of log messages going past. There it is! The reason I woke up at 2:30am on a Sunday. A broken API integration caused by an update I didn’t notice was supposed to happen 12 months ago.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;So, you are writing a new API product, and I am going to be leading an effort to integrate with it. Read on for my list of what you should be caring about and the must-have features.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;TLDR?&lt;/strong&gt; The most important parts of your api design are:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Never break backwards compatibility&lt;/li&gt;&#xA;&lt;li&gt;Be respectful of performance and stability&lt;/li&gt;&#xA;&lt;li&gt;Give me tools for security and access management&lt;/li&gt;&#xA;&lt;li&gt;Break the APi with care&lt;/li&gt;&#xA;&lt;li&gt;Give me a great onboarding and documentation experience&lt;/li&gt;&#xA;&lt;li&gt;Make it gorgeous and easy to understand&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h2&gt;Stability and Backwards compatibility&lt;/h2&gt;&#xA;&lt;p&gt;The number one thing that I want from your API is that it never breaks, and my integration just works forever! Be gentle if you are going to break it. Once I have written the code, don’t make me go back and look at it 6 months, 1 year, 5 years later.&lt;/p&gt;&#xA;&lt;p&gt;10% of effort is spent on implementing successful software and 90% is spent on operation and support. I call this the 10-90 rule. Most blogs, products and new developers focus on making the initial implementation easy. The reality is that business will spend 90% of the effort on operating and supporting a successful integration. An API really explodes this factor because so many external developers are also involved for each partner or customer that uses your product. A significant break in the integration could mean 100’s of developers need to stop and fix something. This level of distraction is churn just waiting to happen.&lt;/p&gt;&#xA;&lt;p&gt;A surprising example of &lt;strong&gt;not&lt;/strong&gt; breaking the API is Windows. A decades old binary will still run, because of the stellar focus on backwards compatibility. Keeping the ‘classic’ software running is a huge factor in the ongoing enterprise success that windows has. Many small and medium size businesses depend on these niche databases and applications. A huge sector of IT professionals are employed to operate and support them.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;How do you avoid breaking backwards changes?&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Use versioning to make sure that deprecated functions are never removed but newer implementations are available.&lt;/li&gt;&#xA;&lt;li&gt;Never remove a field from the returned data type.&lt;/li&gt;&#xA;&lt;li&gt;Do not rename a field.&lt;/li&gt;&#xA;&lt;li&gt;Never re-type a field.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;Sticking to these basic rules are the most important thing you can do to make me happy with your API product.&lt;/p&gt;&#xA;&lt;h2&gt;Performance and Scalability&lt;/h2&gt;&#xA;&lt;p&gt;Your API should never cause me a butt-clenching text message or phone call at 1:30am in the morning. To do this, you need to shift my focus to a better implementation which can handle an unexpected increase in the scale or performance of your API.&lt;/p&gt;&#xA;&lt;p&gt;If a request is slow it should timeout. If it is &lt;em&gt;always&lt;/em&gt; slow, it should be asynchronous. Timeouts are super simple to implement, and you can usually just add some configuration to your load balancer for it. To really handle long running requests though you need to have a ‘start the work’ endpoint and a ‘get the work’ endpoint.  This is an idempotent approach that allows really easy caching of requests. It will result in less work done by the servers, as holding multiple long running requests open is more likely to result in timeouts and re-running the request.&lt;/p&gt;&#xA;&lt;p&gt;Rate limiting is another must have for your API. It is important because I will be forced to think about retry logic during implementation, and this will yet again save me from an unplanned late evening wake-up call.&lt;/p&gt;&#xA;&lt;p&gt;When you are developing the webhook feature, please be kind and provide me options for queuing or load shedding behaviour. I have never implemented a push integration which hasn’t eventually gone haywire at some point. This could mean your product has gone viral (good), or the vendor has gone wild (bad). In either situation I have been extremely grateful for systems put in place ahead of time to handle this load.&lt;/p&gt;&#xA;&lt;h2&gt;How to version and break changes&lt;/h2&gt;&#xA;&lt;p&gt;Never remove or break an API that is still being used. You need to wait at least a year, really! I have actually worked with customers and partners that do a once a year API call. When the yearly call breaks, you can probably figure out what that person thinks about your product when they come to fix it.&lt;/p&gt;&#xA;&lt;p&gt;Instead of changing the meaning of an existing API, you need to introduce new versions. One easy way to do this is to introduce a new name for the API method. Or if you have a whole suite of breaking changes to endpoints you might mint a major version upgrade in the URL path.&lt;/p&gt;&#xA;&lt;p&gt;But before you go blasting ahead on your API, I want you to ask yourself.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;Do I really need to remove/break that api?&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;The best way to break it is to get on a call and help every customer using the API to upgrade their integration. If you did this you will be 1000% better than every other company that has an API product. This is something that I think is a very cool way to do something which doesn’t scale, and can delight your customer.&lt;/p&gt;&#xA;&lt;h2&gt;Security and access management&lt;/h2&gt;&#xA;&lt;p&gt;Give me tools to audit, manage and control how API keys are used.&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Adding new keys or revoking old keys&lt;/li&gt;&#xA;&lt;li&gt;Setting expiry times for temporary keys&lt;/li&gt;&#xA;&lt;li&gt;Setting granular permission for the key&lt;/li&gt;&#xA;&lt;li&gt;Seeing the usage of the key&lt;/li&gt;&#xA;&lt;li&gt;Setting billing limits&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;You should care about these features because they will reduce your customer support workload. They help customers when their integration goes wild. If something bad happens, all the important audit data is available through the usage history for the API.&lt;/p&gt;&#xA;&lt;h2&gt;Clear Documentation and onboarding&lt;/h2&gt;&#xA;&lt;p&gt;Every commit that changes the API should come with a changelog update and a new documentation site. Tools like openapi help with the generation of docs and you should consider something similar for your platform. Include all the things, like generated types and field by field data model comments. Offering client libraries I can &lt;code&gt;npm install&lt;/code&gt; or &lt;code&gt;go get&lt;/code&gt; are nice to have, but not necessary in my opinion.&lt;/p&gt;&#xA;&lt;p&gt;When I get round to integration of your API I’m going to start with the documentation. This is your opportunity to set me on the right path and make something that is high quality and won’t break on xmas day when I’m enjoying time with my family.&lt;/p&gt;&#xA;&lt;p&gt;Another amazing thing you could consider here is ‘white glove’ level of service, give me access to a developer who knows your API and can help me integrate. Another opportunity for standing out amongst a field of mediocre at best API products.&lt;/p&gt;&#xA;&lt;h2&gt;Style doesn’t matter but you asked!&lt;/h2&gt;&#xA;&lt;p&gt;Least of all and last in the list is on the API style. HTTP, REST-ish and json formatted is just fine thanks!&lt;/p&gt;&#xA;&lt;p&gt;I like plain old HTTP because I can call it from just about anything from &lt;code&gt;curl&lt;/code&gt; through to any old other programming language.&lt;/p&gt;&#xA;&lt;p&gt;Using roughly REST (no need to be zealots) style to me means using clear URL paths, and GET/POST/DELETE methods. It is incredibly nice to have cache headers in the response so that I can check if the rest of the body needs to be read or if I can simply wait and query again later for new data.&lt;/p&gt;&#xA;&lt;p&gt;There are a lot of options for API style, graphQL, protoubf, xml, soap. &lt;strong&gt;Style is much less important than any of the other points here&lt;/strong&gt;. Reason for being chill on API style is that it is more of an &lt;em&gt;implementation&lt;/em&gt; factor than a &lt;em&gt;support&lt;/em&gt; factor. One style from another doesn’t represent too much difference in the ongoing support of the integration.&lt;/p&gt;&#xA;&lt;h2&gt;Too long, did read&lt;/h2&gt;&#xA;&lt;p&gt;Thank you for taking the time to implement an API in your product, I really appreciate it and I love the ability to integrate and automate on top of the platform. It is important, and will be ever more important to the decision makers and buyers of software to have API functionality.&lt;/p&gt;&#xA;&lt;p&gt;Most important in your implementation is to respect the 90% of time spent on supporting the API. You should be thinking about this with every increment on the product. We underestimate the pain and time that goes into the operation of software. If you follow my suggestions here your support people will be happy and your customers unburdened.&lt;/p&gt;&#xA;&lt;p&gt;My favourite type of API integration is the one I do once and think about very little for the next few years.&lt;/p&gt;&#xA;&lt;p&gt;If you are putting the effort into making an API product you should care about all this stuff like I do! If you don’t care about it but you&#39;re reading this anyway maybe we should chat and I can figure out how to help you on your journey.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>api</category><category>saas</category><category>architecture</category><category>developer ux</category><category>rest apis</category><guid isPermaLink="true">https://www.ersin.nz/articles/api-design</guid></item><item><title>Is Kubernetes a Meme in 2024?</title><link>https://www.ersin.nz/articles/should-i-use-kubernetes-2024</link><description>I&#39;m diving in to my experience and recommendations on who should use Kubernetes at work or for fun in 2024.</description><pubDate>31 Aug 2024 21:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;h1&gt;The good the bad and the YAML&lt;/h1&gt;&#xA;&lt;p&gt;Is kubernetes (K8s) a meme, or should you stop what you are doing right now and pick it up? Read on for my recommendations a couple years into using K8s in production.&lt;/p&gt;&#xA;&lt;p&gt;Over the last decade I have deployed software in just about every PAAS possible. In the beginning, it was good old SFTP and phpmyadmin, then it moved on to administering individual linux servers with shell script glue. The PAAS phenomena of heroku I experienced during its glory day and then I have experienced the second coming of low config PAAS services like vercel and netlify. But I never got to really use kubernetes beyond tutorials and short books.&lt;/p&gt;&#xA;&lt;p&gt;In my current role I work across a couple k8s clusters, we haven’t needed to venture beyond the GKE platform. Today we’re going to dive into that experience, what you should expect, what to do (and not do).&lt;/p&gt;&#xA;&lt;h2&gt;You’re going to need more than one cluster.&lt;/h2&gt;&#xA;&lt;p&gt;Any serious deployment of kubernetes is going to involve multiple clusters. I believe this to be an absolute certainty. You need isolation between a dev and a prod environment, and you may even run a local environment although that’s something we have avoided thus far. Given the reality that you will be running multiple nodes, you should make sure to introduce some kind of infrastructure tool. I have found terraform to do a good job, and I prefer HCL to YAML, but YMMV!&lt;/p&gt;&#xA;&lt;h2&gt;It makes it really easy to deploy services&lt;/h2&gt;&#xA;&lt;p&gt;There is relatively low friction to deploying new services. I would say it is equivalent to the ease of CapRova or Dokku. Beyond that, it works very well for adhoc things, I.E you just want to temporarily run a pod so that you can get a shell and test some things in the infrastructure.&lt;/p&gt;&#xA;&lt;h2&gt;Cron jobs work, but they kind-of-suck&lt;/h2&gt;&#xA;&lt;p&gt;Kind-of-suck is a mild way to put it. Coordinating jobs is non-existent in the platform, there is no facility for automated alerting. Monitoring gives you simple health/not-health status messages. At SAAS platform, we introduced dagster as an orchestration layer for jobs, but there are a few other good options that could be considered. If you’re raw-dogging the k8s CronJob, then you need to do stuff like have application code that logs specific messages so you can setup a metric on that log and alert on it.&lt;/p&gt;&#xA;&lt;h2&gt;You are literally swimming in yaml&lt;/h2&gt;&#xA;&lt;p&gt;Managing yaml is a weird sisyphean job. It is by far my least favourite language that I need to work with, and it makes up 90% of the lines in the infrastructure repos. I would &lt;strong&gt;LOVE&lt;/strong&gt; to see better tools than pure yaml or kustomize here. There are so many footguns with YAML that I could riff on here, but I will keep it short and spare you the rant.&lt;/p&gt;&#xA;&lt;h2&gt;Easy to integrate with CI/CD&lt;/h2&gt;&#xA;&lt;p&gt;A declarative deployment manifest is obviously good for deployment, and it’s been really easy to manage this in CI/CD. At work we used github actions. These days there are just a huge plethora of good CI/CD tools though. Not Google cloud build though, I found that to be functional but not very good for sharing common patterns and actions.&lt;/p&gt;&#xA;&lt;h2&gt;Where is the workload?&lt;/h2&gt;&#xA;&lt;p&gt;Kubernetes can burn through all your cloud credit if you don’t take a cautious approach.Spinning up new nodes is so easy to do, and they can sit underutilised burning the runway.  Workloads go wild with no memory limits, the controller helpfully kills another important process mid way through a distributed transaction. Managing this is a &lt;strong&gt;NEW&lt;/strong&gt; job that exists in the kubernetes world. You get good reporting out of GKE to manage it, but a human is unfortunately still needed to check on things on a weekly basis. Being responsible for the platform still requires a good knowledge of what workloads run when so that you can ensure your cluster is efficient and optimised.&lt;/p&gt;&#xA;&lt;h2&gt;It would be HARD to build k8s on your own&lt;/h2&gt;&#xA;&lt;p&gt;There is so much that you just don’t need to think about when you’re bought in to a good k8s platform like GKE. Things that are insanely easy and hardly need to be thought about:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Storage provisioning&lt;/li&gt;&#xA;&lt;li&gt;Auto Scaling groups&lt;/li&gt;&#xA;&lt;li&gt;Local Service to Service domain system&lt;/li&gt;&#xA;&lt;li&gt;Logging&lt;/li&gt;&#xA;&lt;li&gt;Load balancing in a deployment&lt;/li&gt;&#xA;&lt;li&gt;Patching the host OS&lt;/li&gt;&#xA;&lt;li&gt;Restarting services &amp;amp;&amp;amp; failing over to another node&lt;/li&gt;&#xA;&lt;li&gt;Killing broken nodes&lt;/li&gt;&#xA;&lt;li&gt;Monitoring and alerting&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h2&gt;Custom operators have hardly come up&lt;/h2&gt;&#xA;&lt;p&gt;We haven’t needed to implement our own custom operator, although you could totally take that path if you were a larger organisation, at our current size (3ish teams), it’s just not needed. We have integrated a few off the shelf components in to our cluster like Dagster and Minio, which provide fast redundant object storage and an orchestration platform respectively. I wouldn’t strictly recommend these unless you are doing a lot of data engineering and AI model work in your business.&lt;/p&gt;&#xA;&lt;h2&gt;It’s Easy to experiment&lt;/h2&gt;&#xA;&lt;p&gt;I can’t under-represent how awesome it is to have an environment where you can easily temporarily spin up a postgres, a deployment and start messing with it in the infrastructure. This has been a huge boon for proof of concept work. Ease of experimentation is a virtue of infrastructure that every platform team should be focusing on.&lt;/p&gt;&#xA;&lt;h2&gt;Using a cluster != Operating a cluster != Making a cluster&lt;/h2&gt;&#xA;&lt;p&gt;There are three levels you can work at with k8s, and I think it’s worth considering that as a needed complexity. Using a cluster as a developer you can ignore a lot of the things in this list. Most of this list in fact is relevant to people in the middle category. People who are operating a cluster and also using that cluster to driver solutions and supporting a about 10 other devs in using the cluster.&lt;/p&gt;&#xA;&lt;p&gt;Dropping down a level to making a cluster requires bringing a few other things, Auto scaling groups, VM images with k8s, configuration management. At the current SAAS we haven’t needed to drop to this level and in fact I think we get huge benefit from just sticking to the GKE platform. Giving up the ‘make a cluster’ flexibility is an intentional decision to depend on best practice and out of the box solutions.&lt;/p&gt;&#xA;&lt;h2&gt;Actually, you can run a database on k8s&lt;/h2&gt;&#xA;&lt;p&gt;You probably shouldn’t do this for critical workloads, but I’ve found it fantastic for development and prototype purposes. It’s super easy to destroy it when the experiment is done and you haven’t needed to faaf around with any terraform to get it done, you can do it all within a dev environment without involving any synchronised process and hat is a huge benefit (see “It’s easy to experiment”).&lt;/p&gt;&#xA;&lt;h2&gt;Containers Pods, Deployments and StatefulSets&lt;/h2&gt;&#xA;&lt;p&gt;Working with k8s has introduced all kinds of weird new terminology, but I think it’s a &lt;em&gt;good&lt;/em&gt; abstraction to have. See past rants about cattles vs pet. It’s a lot more fruitful to talk about the components of your system in terms of the Deployments/StatefulSets/Jobs. The learning curve is steep as you need to grok a lot, but it &lt;em&gt;does&lt;/em&gt; make sense. Stick it through the painful start, it’s worth it!&lt;/p&gt;&#xA;&lt;h2&gt;Containers are great&lt;/h2&gt;&#xA;&lt;p&gt;For development, it’s really nice to be able to just run locally in a container, or pull something from the registry to get it working alongside your code. I would always reach for containers on any deployment even a single $5 digital ocean droplet project ( a note to the reader, that’s actually what your reading this website from right now).&lt;/p&gt;&#xA;&lt;h2&gt;A platform with many options&lt;/h2&gt;&#xA;&lt;p&gt;Kubernetes is like a buffet with every tasty option available. You know you need to fill yourself up, but there are way too many options, and it’s not always obvious which of the many solutions is ‘right’. Many solutions are often recommended but overrated, ‘service mesh’ being the biggest culprit. We have managed to get a long way on our AI SAAS without bringing a service mesh. Maybe your milage is different here, but we found the existing abstractions over the network acceptable out of the box. Working with k8s, you can feel how long the platform has been around, some of the API’s are starting to show their age and multiple ways to do the same thing might exist, it’s worthwhile coming up with an ‘example’ service in your system so that devs know the right configuration to apply and don’t need to duplicate the work of deciding.&lt;/p&gt;&#xA;&lt;h1&gt;So who should use K8s?&lt;/h1&gt;&#xA;&lt;p&gt;If you have a couple teams of developers, and you want to be deploying a lot of services, keeping a fairly independent release schedule then kubernetes might be a really good option for your team.  If you are already on the google cloud platform and just spinning things up then it’s kind of a no brainer. The amount of built in services is really nice to have and will save you a huge amount of non-product work time at the start, allowing you to focus on delivering value. When your workload gets massive, you know you can scale and migrate to better services that make sense, and you are not locked in to a PAAS software that holds your business by the neck with an invoice.&lt;/p&gt;&#xA;&lt;p&gt;It’s not to say k8s has no warts or pain. Don’t use kubernetes if you are a solo founder or a bootstrapper, or even working in an organisation that will only afford one or two teams. I also wouldn’t recommend agencies or consultancies to focus on k8s unless they are going to double down on the managed service and selling kubernetes operators as a business. It’s a very advanced, complicated tool to run in production and you really need to focus a bit of time on operating the system even if you get a lot of features out of the GKE platform anyway.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>kubernetes</category><category>saas</category><category>paas</category><category>sysadmin</category><guid isPermaLink="true">https://www.ersin.nz/articles/should-i-use-kubernetes-2024</guid></item><item><title>Reflection is slow in Golang</title><link>https://www.ersin.nz/articles/reflection-is-slow</link><description>How to get a 29x performance improvement by removing reflection from a go codebase.</description><pubDate>25 May 2024 21:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;Recently I’ve been processing huge amounts of AIS (Automatic Identification System)  data at my day job. By improving performance of the data processing, we enable prompt alerts and facilitate fast feedback loops, empowering platform users with real-time insight. Removing reflection from the hot loop resulted in a &lt;strong&gt;29x performance improvement&lt;/strong&gt;. Read on if you want to dive head first into what reflection is and why it’s not a good idea to call it thousands of times per second.&lt;/p&gt;&#xA;&lt;p&gt;Generally speaking, the Go language is considered a performant and efficient language compared to its peers, but there are certain foot guns when it comes to using the reflection package. If you are using the package, it’s definitely worthwhile building out some simple benchmark tests to ensure you aren’t introducing a performance issue.&lt;/p&gt;&#xA;&lt;h2&gt;What is reflection anyway?&lt;/h2&gt;&#xA;&lt;p&gt;Reflection is a standard library package in go that lets you introspect the type system to dynamically figure out what the types and values of an &lt;code&gt;any&lt;/code&gt; type are at runtime. This is the perfect solution if you want to write some code that iterates through the fields of a struct, or you want to write some code that reads the &lt;code&gt;struct tag&lt;/code&gt; comment and does something smart with it.&lt;/p&gt;&#xA;&lt;p&gt;In short, it’s the way to ‘meta program’ and write code that works on your own programs state at runtime. Let’s check out an example.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code&gt;package main&#xA;&#xA;import (&#xA;    &amp;quot;fmt&amp;quot;&#xA;    &amp;quot;reflect&amp;quot;&#xA;)&#xA;&#xA;// A generic function to print struct details&#xA;func PrintStructDetails(data any) {&#xA;    val := reflect.ValueOf(data)&#xA;    typ := reflect.TypeOf(data)&#xA;&#xA;    // Check if we&#39;re dealing with a struct&#xA;    if val.Kind() != reflect.Struct {&#xA;        fmt.Println(&amp;quot;Not a struct!&amp;quot;)&#xA;        return&#xA;    }&#xA;&#xA;    fmt.Println(&amp;quot;Type:&amp;quot;, typ.Name())&#xA;    fmt.Println(&amp;quot;Fields:&amp;quot;)&#xA;&#xA;    // Iterate over struct fields&#xA;    for i := 0; i &amp;lt; val.NumField(); i++ {&#xA;        field := typ.Field(i)&#xA;        value := val.Field(i)&#xA;&#xA;        fmt.Printf(&amp;quot;\tName: %s Type: %s Value: %v Tag: %s\n&amp;quot;, field.Name, field.Type, value, field.Tag)&#xA;    }&#xA;}&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Try it: &lt;a href=&#34;https://play.golang.com/p/TVx-vokhiyY&#34;&gt;https://play.golang.com/p/TVx-vokhiyY&lt;/a&gt;&lt;/p&gt;&#xA;&lt;p&gt;In the above example, we have created a function PrintStructDetail, that will print out the names and fields of a struct programmatically. It takes an ‘&lt;code&gt;any&#39;&lt;/code&gt; type as the parameter, and then if it is a struct type it iterates through the fields and prints out the name and value of the field.&lt;/p&gt;&#xA;&lt;h2&gt;Why use reflection?&lt;/h2&gt;&#xA;&lt;p&gt;You probably have already used reflection in golang without realising it. For example, json encoding and decoding in the standard library is reflection based, it does something like the above to map parsed json values into a struct type based on the values of the field names.&lt;/p&gt;&#xA;&lt;p&gt;Albeit you are unlikely to need to use reflection in your day to day coding, it can be used to make some magical things happen in Go, it’s a great way to extend the language. Advanced Go programmers should definitely add &lt;code&gt;reflection&lt;/code&gt; to their tool belt.&lt;/p&gt;&#xA;&lt;p&gt;Today however, we won’t be going in depth on how and why to use reflection, we’re going to show you how we can make parsing massively more performant by not using it.&lt;/p&gt;&#xA;&lt;h2&gt;Why is reflection slow?&lt;/h2&gt;&#xA;&lt;p&gt;To keep it dense and too the point, reflection is slow because of missed compiler optimization and extra memory allocation. When we call reflection on a type, it means the go runtime needs to allocate more memory to store the types, fields, names and tagblocks. All of this extra runtime stuff provides indirection in the generated assembly code. Processors are extremely efficient at doing the same operation many times in a row, because they can ‘pipeline’ multiple bits of data into the same operation. When reflection is used the pipeline optimization is not possible, furthermore – you need to do slow allocations and lookups on the heap to figure out what the type is before you can do the actual operation or work on the data&lt;/p&gt;&#xA;&lt;p&gt;You don’t need to know exactly the theory on &lt;em&gt;why&lt;/em&gt; this is slow to get the benefits of not using reflection though. To prove that it makes sense, just benchmark the code in practice, and use production telemetry to show how pulling out reflection can result in a performance improvement.&lt;/p&gt;&#xA;&lt;p&gt;TLDR: if your types are massive or you are frequently invoking reflection, then you should look at pulling out the need for reflection from the code.&lt;/p&gt;&#xA;&lt;h2&gt;A deep dive on AIS parsing&lt;/h2&gt;&#xA;&lt;p&gt;AIS (Automatic Identification System) is a radio based protocol for helping ships and anything on the water avoid collision. My team is concerned with ingesting a global dataset of the stuff. Speed of ingesting the data is paramount. The first step of all our work involves decoding &lt;code&gt;[]bytes &lt;/code&gt;into a struct so that we can perform analytics and build datasets for training machine learning models.&lt;/p&gt;&#xA;&lt;p&gt;There are many AIS message types (27 of them), most commonly they are a bag of fields related to a ship’s position, heading, course, callsign and mmsi (a kind of unique vessel identifier).&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-golang&#34;&gt;type PositionReport struct {&#xA;&#x9;Header                    `aisWidth:&amp;quot;38&amp;quot;`&#xA;&#x9;Valid                     bool            `aisEncodeMaxLen:&amp;quot;168&amp;quot;`&#xA;&#x9;NavigationalStatus        uint8           `aisWidth:&amp;quot;4&amp;quot;`&#xA;&#x9;RateOfTurn                int16           `aisWidth:&amp;quot;8&amp;quot;`&#xA;&#x9;Sog                       Field10         `aisWidth:&amp;quot;10&amp;quot;`&#xA;&#x9;PositionAccuracy          bool            `aisWidth:&amp;quot;1&amp;quot;`&#xA;&#x9;Longitude                 FieldLatLonFine `aisWidth:&amp;quot;28&amp;quot;`&#xA;&#x9;Latitude                  FieldLatLonFine `aisWidth:&amp;quot;27&amp;quot;`&#xA;&#x9;Cog                       Field10         `aisWidth:&amp;quot;12&amp;quot;`&#xA;&#x9;TrueHeading               uint16          `aisWidth:&amp;quot;9&amp;quot;`&#xA;&#x9;Timestamp                 uint8           `aisWidth:&amp;quot;6&amp;quot;`&#xA;&#x9;SpecialManoeuvreIndicator uint8           `aisWidth:&amp;quot;2&amp;quot;`&#xA;&#x9;Spare                     uint8           `aisWidth:&amp;quot;3&amp;quot; aisEncodeAs:&amp;quot;0&amp;quot;`&#xA;&#x9;Raim                      bool            `aisWidth:&amp;quot;1&amp;quot;`&#xA;&#x9;CommunicationStateNoItdma `aisWidth:&amp;quot;19&amp;quot;`&#xA;}&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;&lt;a href=&#34;https://github.com/BertoldVdb/go-ais/blob/master/messages.go#L32-L50&#34;&gt;https://github.com/BertoldVdb/go-ais/blob/master/messages.go#L32-L50&lt;/a&gt;&lt;/p&gt;&#xA;&lt;p&gt;In the AIS parsing library, it’s represented by the above struct. The struct tag contains a &lt;code&gt;aisWidth&lt;/code&gt; parameter that defines the number of bytes to ingest for that part of the message. A message is simply a sequence of bytes, where certain chunks of bytes represent different fields of data.&lt;/p&gt;&#xA;&lt;p&gt;You can probably imagine how parsing works already. The parser will iterate through the fields of the struct, reading the ‘width’ number of bytes, reflecting on the type of the field, then calling the write conversion method to turn that data into something like a uint8, or a bool, or a string.&lt;/p&gt;&#xA;&lt;p&gt;OK, so how does this all relate to the slowness of reflection? This approach works perfectly fine for lower volumes of data, but for larger datasets, you have repeated calls to &lt;code&gt;reflect.TypeOf reflect.ValueOf&lt;/code&gt;.  All this really adds up to the amount of time it takes for parsing a message.&lt;/p&gt;&#xA;&lt;h2&gt;How did I speed it up?&lt;/h2&gt;&#xA;&lt;p&gt;One day, while getting to the bottom of another issue in the codebase, I had just run the heap profiler on the decoder component. What I saw was a huge number of allocations in the &lt;code&gt;reflect&lt;/code&gt; package. Specifically, the thing that iterates over struct fields, and figures out the type of these fields. Because of what I know about reflection being a ‘slower’ operation than usual, I had the confidence to dive in and perform some tests .&lt;/p&gt;&#xA;&lt;p&gt;Ultimately, this led me to creating a small custom code generator that removes reflection from the parser, in favour of a more simple parser. I was able to use the existing metadata about the types provided in the struct tags, but generate a parser that didn’t call &lt;code&gt;reflect&lt;/code&gt; at runtime at all.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;Removing reflection resulted in a 29x speed improvement&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;Fortunately I work for an organisation that values open source, and I was able to contribute this change back to the source of the parser..&lt;/p&gt;&#xA;&lt;p&gt;&lt;a href=&#34;https://github.com/BertoldVdb/go-ais/tree/master/parser_generator&#34;&gt;https://github.com/BertoldVdb/go-ais/tree/master/parser_generator&lt;/a&gt;&lt;/p&gt;&#xA;&lt;p&gt;It’s a dense 500 lines, but it gets the job done. The code works by knowing which struct represents a message that can be parsed, and the code integer that maps to it.  It was important to reuse as much of the existing parser as possible, because it is known to be ‘complete’ and correct. After running both implementations side by side over massive datasets of AIS, I was able to confirm this, and with confidence move the change into production.&lt;/p&gt;&#xA;&lt;h2&gt;Would I recommend this to others?&lt;/h2&gt;&#xA;&lt;p&gt;The way parsing works in this ais library is an interesting one. It works well as an expressive way to show how the AIS format works. However, I think it’s not a very idiomatic approach for a go codebase. I’m not sure I would recommend the code generation approach for building your own new parser either.  Parsers are a really important advanced topic to learn. For anyone who hasn’t dabbled with parsers I would recommend a deep dive on the topic.&lt;/p&gt;&#xA;&lt;p&gt;A good small hand rolled parser is worthwhile maintaining and developing, and I would avoid formal grammars and intermediate languages where possible. The problem with these parser generator approaches is that you often end up with a custom DSL that is difficult for others to pick up and learn.&lt;/p&gt;&#xA;&lt;p&gt;There is a good amount of public knowledge that backs up this approach to building parsers, and you can see that most languages do not rely on formal or custom grammars like BNF, instead deciding to focus on a hand rolled approach&lt;/p&gt;&#xA;&lt;p&gt;One recommendation I will make before wrapping this up, is that you should check out this &lt;a href=&#34;https://interpreterbook.com/&#34;&gt;really great book&lt;/a&gt;. In the parsers and compilers book, you go through a step by step test first approach to building a tokeniser, lexer and interpreter for a tiny language called ‘monkey’. I found it to be a very awesome intermediate -&amp;gt; advanced level book, and it uses the go language (even better!).&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>golang</category><category>ais</category><category>reflection</category><category>advanced</category><guid isPermaLink="true">https://www.ersin.nz/articles/reflection-is-slow</guid></item><item><title>The Best Go API Framework for RESTful services</title><link>https://www.ersin.nz/articles/better-api-framework-for-golang</link><description>Tired of clunky Go API frameworks? This week I explore a framework that leverages Go&#39;s type system for clean RESTful APIs, streamlined development, and superior documentation.</description><pubDate>02 Apr 2024 21:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;The built in http server library in Golang is an ‘S’ tier masterpiece that works fantastically well, even for large monolithic applications. But it’s not quite right. What I miss is something that allows me to abstract away some of the http details and just focus on a simple service definition. The HTTP API it generates should be idiomatic REST, but the implementation should be largely focused around simple type interfaces. This dream framework should give me an implementation experience that looks similar to GRPC, but exposes a REST api instead of protobuf.&lt;/p&gt;&#xA;&lt;p&gt;Interested in my vision for this? Read On!&lt;/p&gt;&#xA;&lt;h2&gt;It should generate the best documentation&lt;/h2&gt;&#xA;&lt;p&gt;The best developer experience is always underpinned by fantastic documentation. OpenAPI is a good start. Being explorable, and being able to try out the API is key. The Documentation should be easy to extend and seamlessly update based on your type definitions and code changes.&lt;/p&gt;&#xA;&lt;p&gt;Beyond just documentation being generated, Open API is the right choice because it&#39;s the lingua-franca of the API world. Tools like Postman make it easy to test and import API’s. Smart API gateways plug on to an open api spec with additional functionality like authentication, authorization, load balancing, rate limiting and caching.&lt;/p&gt;&#xA;&lt;h2&gt;It should be REST-ish&lt;/h2&gt;&#xA;&lt;p&gt;Sorry we won’t be diving into the details of the REST warfare of the early noughties today. But this perfect API framework for go should make it dead simple to make a json based API, that fully supports the HTTP platform it builds on. We want to be as standard as possible so that caches can respect e-tags and we can use PUT/POST/DELETE/GET to their full extent.&lt;/p&gt;&#xA;&lt;p&gt;The &lt;a href=&#34;https://github.com/grpc-ecosystem/grpc-gateway&#34;&gt;grpc gateway&lt;/a&gt; project is an example of what not to do. It provides a very simple reverse proxy for your GRPC service that also generates an Open API schema for your service. The problem is that it railroads you into a POST call for every RPC. Thankfully there are escape hatches, but it makes things unnecessarily complex. Our dream API framework for go should simply support making an api that can leverage all the good parts of the http platform.&lt;/p&gt;&#xA;&lt;h2&gt;Let the developer focus on the business logic&lt;/h2&gt;&#xA;&lt;p&gt;Our days are already filled with writing and managing config in a number of horrible languages (HCL, yaml, json, ini etc). The dream framework keeps this config to an absolute minimum.&lt;/p&gt;&#xA;&lt;p&gt;I frame this as &#39;complexity points&#39;. We only have so much capacity for fitting new tools and lanugages in to our head. We should be optimising on this to provide the most value possible. Is memorizing yet another config schema really the way? Why not leverage what we already know well, the type system of the language. This is why I think the best solution is to eschew config in favour of using what the language provides us.&lt;/p&gt;&#xA;&lt;p&gt;To make it painfully clear. Let me spend my complexity points on solving the problem, doing the valuable thing, not memorizing another DSL for your framework. This means not writing the schema first in protobuf, or open api YAML (or something else).&lt;/p&gt;&#xA;&lt;p&gt;A better way in my opinion is to build your framework on top of the go languages simple and easy to use type system. Let’s just keep it simple and build it on top of interfaces, structs and error types.&lt;/p&gt;&#xA;&lt;h2&gt;Provide simple escape hatches when needed&lt;/h2&gt;&#xA;&lt;p&gt;Leaky abstractions don’t always suck. They are actually essential for web frameworks. Our dream API framework should allow you to drop back to the raw material we are working on. You should always be able to turn a request handler into a simple &lt;code&gt;http.HandlerFunc&lt;/code&gt;. Too often, frameworks will pile so much abstraction that it makes it really hard to find your way to the familiar. This is a common problem that is not often thought about when evaluating what tools to use. What this often looks like is happy sailing, churning out features until you inevitably get stuck and need to fork your way to a solution. You then get the joy of supporting an old fork, or merging in all the upstream changes. This is yet another unnecessary spend of the &#39;complexity points&#39; that your team can handle.&lt;/p&gt;&#xA;&lt;p&gt;Escape hatches can also mean being able to &lt;em&gt;exit&lt;/em&gt; the framework. For our fictional case here, that would mean falling back to the excellent built in http server library provided by golang. Need to ditch the framework and go old_school? Not a problem!&lt;/p&gt;&#xA;&lt;h2&gt;What I looked at, what sucks and what is &lt;em&gt;almost&lt;/em&gt; OK&lt;/h2&gt;&#xA;&lt;p&gt;There is no API web framework for golang that is &lt;em&gt;Quite&lt;/em&gt; right. They either go too far in my opinion building all these abstractions that hide the fact you are working in the web platform, or they simply don’t go far enough. When it comes to the higher level frameworks, it’s sometimes hard to even know you are working on the web, and there’s layers and layers of abstractions between your handlers and the actual requests/responses. Then other minimalist libraries can be a bit like the wild west. They invite you to shoot yourself in the foot or require you to introduce reams of boilerplate. This opens up opprotunities for inconsistent code. To work effectively in the minimalist framework a team needs huge amounts of discipline.&lt;/p&gt;&#xA;&lt;p&gt;Let&#39;s look at two fo the best candidates I found in a recent review of golang web frameworks that focus on open api definitions.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Go swagger&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;&lt;a href=&#34;https://goswagger.io/&#34;&gt;https://goswagger.io/&lt;/a&gt; is a promising candidate, in that you can do all the rest-ish features you need. But they went &lt;em&gt;too&lt;/em&gt; far with the abstraction in my opinion.  It generates a huge amount of unneccessary code, and the request handler interface is fundamentally flawed. Less is more, and this framework proves it. For example, consider the addition of jsonschema and full openapi spec to the code generation in goswagger. I admit, it’s a cool feat of engineering. However, the complexity of the types generated being forced into the golang language is not worth the cost.&lt;/p&gt;&#xA;&lt;p&gt;Another example of this negative impact feature is the interface chosen for request handlers. The way you can escape to a raw http middleware is great, but the problem is that you lose all the type safety support from the framework. My thesis is that it is a net positive if you catch &lt;strong&gt;MORE&lt;/strong&gt; errors at compile time. The interface type chosen here is an example of the opposite, you can easily write a handler that compiles, but returns the wrong data type!&lt;/p&gt;&#xA;&lt;p&gt;I also despise needing to start with writing json schema and YAML first 😿&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Swag-go&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;&lt;a href=&#34;https://github.com/swaggo/swag&#34;&gt;https://github.com/swaggo/swag&lt;/a&gt; also has some nice things to be said about it. It is a simpler and less abstract option. You can easily mix/match and plug just about any web framework into it, and that’s very cool, but it misses some features I would love to have around isolating the business logic. For one, it doesn’t expose an interface that allows your code to not compile if the schema is incorrect. This is thanks to its ingenious design of code generating an open api schema based on the code comment. It’s just a shame that this doesn’t hook in to forcing some type safety on the request handler.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;GRPC-gateway&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;&lt;a href=&#34;https://github.com/grpc-ecosystem/grpc-gateway&#34;&gt;https://github.com/grpc-ecosystem/grpc-gateway&lt;/a&gt; If you are a GRPC enjoyer, this might be the choice for you. It wraps a service definition, generating an http 1.1 server and open api spec. You run a tiny little reverse proxy that forwards requests to your backing grpc service. It’s quite easy to wrap your own http middleware and just write your business logic as a simple GRPC service.&lt;/p&gt;&#xA;&lt;p&gt;Did you think I was going to only say nice things about this project? Sorry, I love it, but I can’t not poke a few holes.&lt;/p&gt;&#xA;&lt;p&gt;First, it requires a side-car to deploy. This is probably adding a significant amount of drain on your non value generating faffery quota.&lt;/p&gt;&#xA;&lt;p&gt;There is also a quality issue with the generated openapi schema. All rpc calls will turn in to POST requests. This is not ideal if you are planning on supporting a big distributed caching front end. It&#39;s a problem that can be solved via supported annotations in the schema, but again -- you&#39;re swimming against the current here.&lt;/p&gt;&#xA;&lt;h2&gt;What does better look like?&lt;/h2&gt;&#xA;&lt;p&gt;My vision for the best framework to write an API in golang is a little different to what exists at the moment. We take inspiration and build on the great work done prior. Overall, the framework should look pretty familiar. It is about sane defaults, and a massive respect for developer productivity over unnecessary faffery.&lt;/p&gt;&#xA;&lt;p&gt;The &#39;best framework&#39; will never focus on forcing you in to any of the following:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Orms&lt;/li&gt;&#xA;&lt;li&gt;Mapping to database schemas&lt;/li&gt;&#xA;&lt;li&gt;Non JSON interfaces&lt;/li&gt;&#xA;&lt;li&gt;HTML&lt;/li&gt;&#xA;&lt;li&gt;Hypermedia&lt;/li&gt;&#xA;&lt;li&gt;Novel types beyond the Golang standards.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;Instead it will focus on:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Generating great documentation&lt;/li&gt;&#xA;&lt;li&gt;Escape hatches for plain old &lt;code&gt;http.Server&lt;/code&gt; interface&lt;/li&gt;&#xA;&lt;li&gt;Your service interface defines the API schema&lt;/li&gt;&#xA;&lt;li&gt;A simple subset of open API 2.0&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;&lt;strong&gt;Behold -- my tiny Dumb Version half page diagram&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;&lt;img src=&#34;/public/golang-framework-idea.png&#34; alt=&#34;&#34;&gt;&lt;/p&gt;&#xA;&lt;p&gt;Let&#39;s dive in to the above 5 minutes exalidraw-job. The request mapper, response mapper is provided automatically and generated on your behalf.&lt;/p&gt;&#xA;&lt;p&gt;This includes routing, and choosing the right method to call. In this example &lt;code&gt;service.GetArticles&lt;/code&gt; . All of the code you work on for implementing your new handler is encapsulated in your simple handler function&lt;/p&gt;&#xA;&lt;p&gt;&lt;code&gt;func (s *Service) GetArticle(ctx context.Context, articleID int) (Article, error)&lt;/code&gt;&lt;/p&gt;&#xA;&lt;p&gt;The framework is clever enough to figure out the route, the response type, the request parameter is articleID, and the response type is shaped like an Article. If an error is returned, it should handle that and set the appropriate response code depending on the type of error.&lt;/p&gt;&#xA;&lt;p&gt;If you need more granular control on your request/response mapper, just copy over the generated function to your own code package. The code generation mechanism is the smart part, it can scan your own package to see if you are implementing your own direct handler for the path. I.E if you have a `func (s *Service) GetArticle(r *http.Request, w http.ResponseWriter) in your own code, it will not generate the handler for you, allowing you to take control of the request/response mapping.&lt;/p&gt;&#xA;&lt;p&gt;You can start with the sane defaults, and if eventually you need fine grained control -- it&#39;s all their ready for you to take over.&lt;/p&gt;&#xA;&lt;h2&gt;A Golang API Framework That Respects You the developer&lt;/h2&gt;&#xA;&lt;p&gt;This framework would be one that puts the developer experience first. Think Rails, but with more explainable and obvious magic. We want the simplicity of the go language but the full strength of HTTP as an engine of application state. The framework would be as simple as defining a service interface, but let you out-grow the framework and support a code base for 10+ years.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;The core pillars&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;Simplicity&lt;/strong&gt;: Built on top of interfaces, structs, and error types, making it easy to understand and use.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Flexibility&lt;/strong&gt;: Escape hatches to raw HTTP handlers, allowing developers to drop back to the underlying Go HTTP library when needed.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Type safety:&lt;/strong&gt; Catch as many errors at compile time as possible by enforcing return types from request handlers.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Developer Friendly&lt;/strong&gt;: Generates great documentation. Doesn’t force you to write schema first in painful config languages.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Lightweight&lt;/strong&gt;: The framework is lightweight, with no external dependencies, making it easy to deploy and maintain.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;If you are interested in collaborating on this project, or you can point me to one that satisfies what I’m looking for ( even in other languages ), let me know!&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>web</category><category>golang</category><category>web frameworks</category><category>openapi</category><guid isPermaLink="true">https://www.ersin.nz/articles/better-api-framework-for-golang</guid></item><item><title>Python is still a GREAT choice in 2024 -- here&#39;s why</title><link>https://www.ersin.nz/articles/wrangling-your-llm</link><description>Choosing some old boring technology reignited my passion for building. HTMX, Flask and SqlAlchemy are a great choice for a project in 2024.</description><pubDate>07 Mar 2024 22:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;Listen, I&#39;m gonna be honest with you -- I&#39;ve been slacking. Writing about CRDT&#39;s, AI tools, emailing the list... It was not my jam the past two weeks. The itch to create hit hard, and I started a new project. The fire to build from scratch, going from 0-1 with no roadmap or guidance. So I&#39;ve ditched the boring stuff and started a new project. Guess what? It&#39;s been the most productive and fun few weeks I&#39;ve had in ages!&lt;/p&gt;&#xA;&lt;p&gt;The new project is written in good old python. It has been an exercise in nostalgia as I pick up a few old tools which have laid neglected, gaining cobwebs. But it&#39;s amazing how fast this stuff comes back, and I think python still works great for this type of project, a web app doing basic CRUD with a few background workers driven from a queue.&lt;/p&gt;&#xA;&lt;p&gt;The last time I used python in any significant way was in 2018, while working for a consultancy on various CRUD and geospatial apps that had a flask or django backend. Included was a fairly large CMS, that had stood the test of time and seen a migration between three frameworks (pylons -&amp;gt; pyramid -&amp;gt; flask).&lt;/p&gt;&#xA;&lt;p&gt;Since then, I have been mostly using the Go language, and PHP for this kind of work. PHP is one I had revisited after several years away when I picked it up in 2019, and a lot had changed for the better. Is python the same? Read on reader, to find out :)&lt;/p&gt;&#xA;&lt;h2&gt;Python packaging is still a nightmare, but it&#39;s getting slightly better&lt;/h2&gt;&#xA;&lt;p&gt;By this stage it&#39;s a bit of a known fact that python packaging suffers from the &#39;npm problem&#39;, too many ways to do the same thing. It&#39;s not all bad news in 2024 though. There has been some unification of standards since 2018. Out of the box support for a &lt;a href=&#34;https://peps.python.org/pep-0621/&#34;&gt;pyproject.toml&lt;/a&gt; file is a big win. This file is a configuration file that can be used to specify the build system, dependencies, and other options that you want to use when building your project. It&#39;s a bit like a &lt;code&gt;package.json&lt;/code&gt; for python. Before this, you would have inconsistent implementations of various config files (pipenv/poetry et al). Now tools can standardize on one format for project metadata.&lt;/p&gt;&#xA;&lt;p&gt;The python packaging story is trending in the right way. I am hopeful that in six years time, there are fewer packaging tools than in 2024.&lt;/p&gt;&#xA;&lt;h2&gt;Python has types now!&lt;/h2&gt;&#xA;&lt;p&gt;Like many dynamic languages, you now have the ability to gradually add types to your code base, and then get warnings and errors at &#39;compile time&#39;. The python implementation is not one I have dug very deep into, but it has cool things like &#39;union&#39;, which allow you to specify one or the other type. There is also a new &lt;code&gt;match&lt;/code&gt; language structure that looks like something you would see in an ocaml style ADT. I find these gradual type systems &#39;interesting&#39;, but they lack the simple ruleset of languages that are designed this way from the start. So, you get a few errors in your editor warning you of typo mistakes, that&#39;s a big win. But; you still won&#39;t get the exact type safety that languages like rust/go/java will, you&#39;re still going to have to write tests to cover your butts and include defensive code to avoid unexpected runtime errors.&lt;/p&gt;&#xA;&lt;p&gt;Gradual typing remains super nice for library usage, and it tells my editor a lot about what parameters/arguments are as I type them. This is the most impactful change with the introduction of types.&lt;/p&gt;&#xA;&lt;h2&gt;I no longer worry about python 2 -&amp;gt; 3&lt;/h2&gt;&#xA;&lt;p&gt;We did it! Hooray, the problems are solved, every distro has python 3, and no more python 2 dependencies exist! Not. Of course python 2 exists, and I&#39;m sure in the millions of pip packages you will run in to some in your travels, but over the last three weeks I have not come across any problems with python 2 dependencies appearing. The story for legacy projects I&#39;m sure is very different, but you can easily get what you want done in python 3.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;Sarcasm alert: the 2 -&amp;gt; 3 problem is probably still very real for legacy projects, but in green field or new things -- you shouldn&#39;t really worry about the two languages at all!&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;h2&gt;Async python is a thing now&lt;/h2&gt;&#xA;&lt;p&gt;You can now plug in various runtimes to allow you to &lt;code&gt;async&lt;/code&gt; and &lt;code&gt;await&lt;/code&gt; in python. I haven&#39;t explored this in much detail, but there are a suite of new frameworks released to work with this new runtime. I really don&#39;t have experience here so I can&#39;t call out if it sucks or not. Looking forward to seeing how it develops, because the GIL and running background jobs has never been a strong point of python. Where-as, my work with the go language has allowed significant parallel process and background processing in almost every project.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;Async python is something I would love to play with, but I&#39;m skeptical that I would ever choose python over go for async runtimes.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;h2&gt;Flask &amp;amp;&amp;amp; HTMX &amp;amp;&amp;amp; Sqlalchemy are brilliant&lt;/h2&gt;&#xA;&lt;p&gt;When it comes to ranking python web frameworks. The Flask framework is second only to django for being battle tested and still widely used. In 2024, it works pretty much the same as in 2018. It first hit public changelogs with a &lt;code&gt;0.1&lt;/code&gt; release in 2010. So I think if you&#39;re choosing Flask as a web framework now, you are making a pragmatic and uncontroversial decision. One which you will unlikely regret in 10 years time (unless you want async). The fact that this framework is turning 14 years old in 2024 is enough to say that it will probably make it to it&#39;s 28th birthday in 2038.&lt;/p&gt;&#xA;&lt;p&gt;The one pain point I have come across with flask, is that there are simply so many plugins with it, you need to be a bit discerning on choosing plugins that will stand the test of time. For example, in the past I have been tasked with maintaining legacy Flask code bases, and you end up taking over -- forking and maintaining these old niche projects, that could have just been an internal module in your codebase in the first place, and made to fit your exact purposes with much less unnecessary concerns to worry about.&lt;/p&gt;&#xA;&lt;p&gt;My new project uses my &#39;default&#39; frontend of HTMX + tailwind. I&#39;ve been extremely happy with this. It fits gorgeously with flask, and the bundled template engine &#39;jinja&#39;. Everything you want to do has been done for the last 14 years using jinja + flask + forms, so you can easily find &#39;prior art&#39;.  &lt;a href=&#34;/articles/htmx-is-old-but-gold-4i9i&#34;&gt;HTMX just lets you do everything you could do 25 years ago on the web&lt;/a&gt;, but bundled into a tiny and easy to use package.&lt;/p&gt;&#xA;&lt;p&gt;Data access for my web application is managed using sqlalchemy. Another fantastically old and mature ORM library. It&#39;s weird for me to put &#39;fantastic&#39; and &#39;orm&#39; in the same sentence. Over the last few years I have eschewed them for plain SQL. That&#39;s a bit of a side-effect of golang not being the greatest language for dynamic style ORMs (a.k.a ActiveRecord). But something has changed with my thinking and this project. First of all, python has gradual typing support and sqlalchemy picks this up nicely, so you get a bit of fancy editor checking on your field names. Second, you can drive your database migrations with sqlalchemy and the &lt;code&gt;alembic&lt;/code&gt; tool. Thirdly, the orm doesn&#39;t hide too much of the sql factor from you, it is fairly easy to escape the orm and write plain sql when you need to.  So all of this combined has resulted in me choosing an ORM for the first time in 4ish years, and after two weeks it&#39;s going good! At this stage it&#39;s too early to say if I&#39;m converted to the ORM life, but I think I remain open to it, and I can say that it works well for this &lt;em&gt;prototype&lt;/em&gt; phase of the project I am currently in.&lt;/p&gt;&#xA;&lt;h2&gt;Python is &lt;strong&gt;S L O W&lt;/strong&gt;&lt;/h2&gt;&#xA;&lt;p&gt;I didn&#39;t expect to run in to this, but after slurping my way through a couple 100 kb of yaml, I found that python was actually taking some decent time to run. It&#39;s a bit annoying really, but not the end of the world. I know I can easily write bits in a better language if I need to. Beyond this, performance hasn&#39;t been such a problem that I have looked in to optimization. There is certainly scope for writing python better to not have such slow code.&lt;/p&gt;&#xA;&lt;h2&gt;the REPL and jupyter notebook experience&lt;/h2&gt;&#xA;&lt;p&gt;My side hacks with clojure and python over the last few years have just embeded the fact that having a REPL (read eval print loop) is an awesome and productive tool for getting shit done. exploring your code, easily copying functions into the repl and testing them, doing random prototyping in the db using the repl are all amazing. I yearn for this experience when I am writing go. The go runtime doesn&#39;t let you just inject code and test it out, but with python you can really just debug from any point in the code and start writing new things to change the outcome. It&#39;s fantastic and magical!&lt;/p&gt;&#xA;&lt;p&gt;Tools like jupyter notebook are also very cool. My day job, we have some excellent data scientists that write a lot of python, and a huge part of their work is shared via jupyter notebooks. These handy files give you an interactive python session that can render graphs, documentation and your dynamic code all at the same time. This is like the fanciest repl in the world, and a good way to share a clever algorithm.&lt;/p&gt;&#xA;&lt;h1&gt;That old stuff is good try it out!&lt;/h1&gt;&#xA;&lt;p&gt;Python might be a classic, a language with more history than some new-hotness framework, but let me tell you -- it&#39;s still excellent! It&#39;s better than it was in 2018, and it&#39;s still a great choice for a classic style web 2.0 application. I&#39;m a believer in choosing less technology if you can, and focusing on being innovative in just that small place where it really counts. The flask stack is perfect for getting an idea off the ground and building something that can go from prototype to mature scalable project.&lt;/p&gt;&#xA;&lt;p&gt;And that&#39;s the key here, right? Sometimes you need tools that let you focus on what instead of how. The stack I described here let me do that. I got to focus on being in the flow of hacking on my project. So if you&#39;ve written some python in the past, don&#39;t just write it off as a non choice for new projects. Taking a step back in time can help you realize and compare that the new flavour of the month framework might not be all it&#39;s marketing page claimed it to be.&lt;/p&gt;&#xA;&lt;p&gt;Now, I&#39;m off to keep coding. Buh bye!&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>flask</category><category>golang</category><category>beginners</category><category>python</category><guid isPermaLink="true">https://www.ersin.nz/articles/wrangling-your-llm</guid></item><item><title>Learning better testing using BDD</title><link>https://www.ersin.nz/articles/learning-better-testing-using-bdd</link><description>I reflect on a fun evening hack, and the future of authentic human writing on the web.</description><pubDate>24 Feb 2024 20:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;As a developer traversing time and space over the last 10 years, I&#39;ve seen some trends come and go. Around the early 2010&#39;s, as TDD was reaching it&#39;s plateau of productivity, I came across a software tool called &#39;cucumber&#39;. It was a magical tool that let you write bits of code that automate your frontend. A technical product manager, would then compose bits of this to code their own acceptance criteria and tests for the app. Now in 2024, I&#39;m seeing less dialogue about this type of tool. Is it because it&#39;s shit? Or are there better options? Now, I think there is merit to the idea, but we can leave behind some of the complexity and heavy weight tools. It is also important to consider how many end to end tests should be supported.&lt;/p&gt;&#xA;&lt;p&gt;This blog post delves into &lt;strong&gt;Behaviour-Driven Development (BDD)&lt;/strong&gt;, exploring how we can embrace its core principles for clear communication and collaboration without succumbing to the allure of overly complicated frameworks.&lt;/p&gt;&#xA;&lt;h2&gt;Bridging the Communication Gap:&lt;/h2&gt;&#xA;&lt;p&gt;The value of testing is undeniable. However, traditional unit testing, integration testing, while essential, can sometimes feel undervalued by the non-programmers in your team. Maybe at some stage you&#39;ve struggled to communicate with stakeholders, project managers or customers the value of these tests.  A typical test, is a simple function with a number of actions taken, ultimately a crash of the test, or a &#39;assert&#39; that fails will result in a failed test. You might see the output of your tests as a series of red/green lines in a text file report. Tools like junit will show a structure that has a class and method names in the report. Although these tests may be hugely valuable for the developers, they may as well be complicated hieroglyphs to others.&lt;/p&gt;&#xA;&lt;h2&gt;The BDD way&lt;/h2&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;BDD bridges the comms gap by establishing a &lt;strong&gt;ubiquitous language&lt;/strong&gt; between devs and stakeholders.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;The most extreme version of BDD involves the use of a domain specific language to agree usecases with the stakeholders. We then code up exactly this test in the app, and can point to a green line in a report to communicate that we have exactly developed this feature, and it&#39;s working as intended! It builds up a valuable asset for understanding what the system does too. New people will understand the breadth of a feature. You will be able to see how features break each other. Furthermore, if you use a tool like (gherkin/cucumber) then the test may even be able to be composed mostly from reused chunks of domain language.&lt;/p&gt;&#xA;&lt;p&gt;Here&#39;s what an example test might look like in the cucumber BDD style DSL&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-gherkin&#34;&gt;Given a user is logged in&#xA;When they click the &#39;add to cart&#39; button&#xA;Then the cart should contain the item&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;The test above is broken in to three parts, given/when/then. The first line is our test setup, which provides all the preconditions of what the test does. Following, we have a &#39;When&#39; statement, which represents some kind of action that the user takes. Finally, we have a &#39;Then&#39; statement, which represents the expected outcome of the test. This simple and clear description of the feature has some underlying code that will run the test, and report back to the user if the test passed or failed. Each of the lines in this test can be composed and remixed to create a new test.&lt;/p&gt;&#xA;&lt;p&gt;For example, we can add a new test case by implementing just the &#39;then the car should be empty&#39; part of the test. This is a powerful way to build up a suite of tests that describe the functionality of the app.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-gherkin&#34;&gt;Given a user is logged in&#xA;When they click the &#39;add to cart&#39; button&#xA;When they click the &#39;checkout&#39; button&#xA;Then the cart should be empty&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;In my experience, this approach can work really well with the right team. However, it&#39;s not without problems. Often, the underlying technology of driving your app is a janky nightmare. In web apps this will be a browser driving tool like playwright or selenium. In desktop applications you might have nightmarish OCR tool like sikuli. Over time, you build up many many tests to describe the functionality. Because you are driving the tests end to end, they can be slow to run, involve complex setup and be flaky due to underlying changes in the user interface. The process of adding new tests is hyper optimized and easy thanks to the domain language, but the act of maintaining the tests gets harder and harder. It makes your momentum on changing the app slow down, and the tests become a major time sink. What compounds with this problem is that the tests become a contract with the business, where you need to be dogmatic about having new coverage with each feature, and never breaking backwards compatibility.&lt;/p&gt;&#xA;&lt;h2&gt;What we can learn and take in to our own projects&lt;/h2&gt;&#xA;&lt;p&gt;Despite all this pain, there are good parts of BDD. I don&#39;t recommend the heavyweight frameworks like gherkin or cucumber. I do think you should write &lt;em&gt;some&lt;/em&gt; end to end tests with playwright/selenium. So what should we take from BDD in to our every day practice?&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;Shared language&lt;/strong&gt;: We should get credit for delivering tests as part of feature development. Use the shared domain language to communicate clearly what the software will do, and build this in to the tests so that we have a solid way of communicating the behaviour of the software when it&#39;s 10 months down the track and no one remembers how the feature works.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Single purpose tests&lt;/strong&gt;: Each test focuses on a single specific scenario, improving test clarity and simplifying maintenance.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Given/When/Then structure&lt;/strong&gt;: This structure provides a clear and concise way to define the test scenario, ensuring everyone understands the setup, action, and expected outcome.&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;strong&gt;Given&lt;/strong&gt;: Sets the initial state of the system.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;When&lt;/strong&gt;: Describes the action which we are testing.&lt;/li&gt;&#xA;&lt;li&gt;&lt;strong&gt;Then&lt;/strong&gt;: Defines the expected outcome of the action.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;These seemingly simple principles can significantly enhance your existing testing practices. By embracing the shared language and focusing on collaboration, you can build better software faster. A consistent approach to these tests are really helpful for yourself when you are coming back to the feature 12 months down the line. They can be helpful for new developers to see what a feature actually does. If you have technical thinking non-programmers in the team, they will love the ability to specify and understand functionality by looking at the way it is tested.&lt;/p&gt;&#xA;&lt;p&gt;BDD isn&#39;t about forcing rigid automation onto non-technical stakeholders. Nor is it about making your developers work with brittle and annoying test frameworks.  It&#39;s about fostering clear communication and shared understanding throughout the development process. By embracing the core principles of BDD and adapting them to your workflow, you can build stronger relationships with stakeholders, improve collaboration, and ultimately, build better software faster.&lt;/p&gt;&#xA;&lt;p&gt;Remember, the key is to find the approach that works best for you and your team. Happy coding!&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>BDD</category><category>testing</category><category>beginners</category><category>software engineering</category><guid isPermaLink="true">https://www.ersin.nz/articles/learning-better-testing-using-bdd</guid></item><item><title>Wrangling chat GPT for writing</title><link>https://www.ersin.nz/articles/wrangling-your-llm</link><description>I reflect on a fun evening hack, and the future of authentic human writing on the web.</description><pubDate>10 Feb 2024 22:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;Depending on where you get your feed, we might be headed for certain doom with what I&#39;m about to describe. I&#39;m not sure if that&#39;s the case. People are saying that the internet will be awash with LLM generated content, and our last ever authentic human thought will be published. Read on, to hear about my evening adventure attempting to generate a blog post with a handful of dozen lines of python. Stay until the end, to find out the problems with this approach, and how I might solve them.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-python&#34;&gt;def write_a_blog_the_hard_way(prompt: str) -&amp;gt; str:&#xA;    &amp;quot;&amp;quot;&amp;quot;&#xA;    Talk to a chat gpt model, using the prompy to develop a blog outline then deliver each part of the outline iteratively&#xA;&#xA;    :param prompt:&#xA;    :return: the post&#xA;    &amp;quot;&amp;quot;&amp;quot;&#xA;    total_post = &amp;quot;&amp;quot;&#xA;    chats = [make_blog_summary(prompt)]&#xA;    for i in range(5):&#xA;        chats = chat(chats + [f&amp;quot;&amp;quot;&amp;quot;&#xA;        {chats[-1]}&#xA;        ---&#xA;        Instructions: Write a paragraph for section {i + 1}&#xA;        &amp;quot;&amp;quot;&amp;quot;])&#xA;        total_post += chats[-1] + &amp;quot;\n----\n&amp;quot;&#xA;    return total_post&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Behold the above 5ish lines of actual python code to write a blog post using chat gpt. Now I&#39;m not saying this is going to change the world and destroy journalism and blogging. But it&#39;s interesting to see a naive approach at implementing a blog generator on top of the chat gpt api. Let&#39;s dive in and see how it holds together.&lt;/p&gt;&#xA;&lt;h2&gt;Getting output from chat gpt&lt;/h2&gt;&#xA;&lt;p&gt;I&#39;m using the &lt;code&gt;openai&lt;/code&gt; pip module to interact with the chat gpt api. It&#39;s a simple matter of getting an API key, instantiating the class and getting started. You can essentially look at the first page of their documentation to see how to interact with the API. I&#39;ve created an even simpler abstraction on top of it with my chat function &lt;code&gt;def chat(msgs: List[str]) -&amp;gt; List[str]:&lt;/code&gt; which takes a list of messages and returns a list of messages. This is a simple way to keep the conversation going, and to keep the state of the conversation in memory. The conversation is alternating messages between you and the agent, starting with your first prompt message.&lt;/p&gt;&#xA;&lt;h2&gt;Setup prompt&lt;/h2&gt;&#xA;&lt;p&gt;In late 2022, I first came across &#39;prompt engineering&#39; as a term. What it means, is to &#39;engineer&#39; your input in to a text model to get the expected output and solve the problem. Now, I&#39;m no expert here, and I&#39;m not even sure if we should be calling it prompt engineering. My Take is to have user provided input, mashed together with a few instructions to the model on how to create an outline to plan the blog.&lt;/p&gt;&#xA;&lt;p&gt;For example, the user might input something like this:&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-text&#34;&gt;Write me a blog in the first person perspective about the benifits of using RSS feeds in 2024&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;And in my code I add the following:&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-text&#34;&gt;Your task is to create a five part outline for a comprehensive blog post.&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;We simply send this combined output to the chat agent, and await a response, which should be a set of headings and bullet points for the post.&lt;/p&gt;&#xA;&lt;h2&gt;The loop&lt;/h2&gt;&#xA;&lt;p&gt;Now that we have an outline for the blog post, lets loop through and ask the agent to give us output for the different paragraphs of the post. Our outline is provided to the LLm, along with the instruction &lt;code&gt;write part {N} of the blog post&lt;/code&gt;. Using this technique I have been able to build larger sets of content from the LLM, nearing a reasonable quality of writing.&lt;/p&gt;&#xA;&lt;h2&gt;But there are problems..&lt;/h2&gt;&#xA;&lt;p&gt;Is this post you are reading now generated by the above code? No, it&#39;s not. I found that the quality of the writing was not up to the standard I would like to publish on my blog. I found that the LLM would often repeat itself. Plus the writing just wasn&#39;t quite up to the standard of what I would be happy putting out in to the world. Hallucination is the other big problem here, your generated blog will lie and vary from the plan you set out. A human is definitely needed to boost quality and make something that is actually usable.&lt;/p&gt;&#xA;&lt;p&gt;I have some untested ideas on how we can solve these problems. Although I think we are still early days on finagling the output from LLM&#39;s, and there are scare resources around on how to solve the issues in making longer form content with them.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Repetitive sections of the post&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;More context on the sections/parts of the post that have been already covered might help here. Or maybe a more granular approach to building the content. One idea would be to take the prompt to an outline, but maybe it should expand on the outline to a next step. Simple sentences. Simple paragraphs. This might be a way to get a bit more control over the output, and get rid of the often repeated paragraphs/sentences.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Low quality content&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;Just not happy with the quality of the content. The voicing never seems quite right. It&#39;s just a bit cringe-worthy! We might be able to improve things with a more refined prompt. Maybe ingest your current content, extract the &#39;tone&#39; and &#39;style&#39; aspects to then generate a prompt which makes similar content.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Hallucination&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;As we all know by now, the LLM&#39;s are prone to hallucination. Having a human in the loop is our only way to ensure this doesn&#39;t happen. At a certain degree you might get some solid results from prompt engineering, but I suspect this to be a minor improvement on results.&lt;/p&gt;&#xA;&lt;h2&gt;Wrapping up&lt;/h2&gt;&#xA;&lt;p&gt;Was this a fun evening hack? You betcha. Will I be using LLM&#39;s in my workflow, yep! But we are far from the dystopian world of an internet filled with pure LLM garbage. There will continue to be real humans, posting for the short term, and I&#39;m still on my journey of regular publishing and writing.&lt;/p&gt;&#xA;&lt;p&gt;Thanks for staying tuned!&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>openai</category><category>chatgpt</category><category>beginners</category><category>python</category><guid isPermaLink="true">https://www.ersin.nz/articles/wrangling-your-llm</guid></item><item><title>My Blog now supports RSS</title><link>https://www.ersin.nz/articles/rss-in-2024</link><description>This week we add an RSS feed to my blog using golang. Here&#39;s how</description><pubDate>01 Feb 2024 22:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;Behold, the latest and greatest feature on my tiny blog platform.  &lt;a href=&#34;/feed.xml&#34;&gt;RSS feeds&lt;/a&gt;! Yes reader, we now support the uber trending web 2.0 tech from 2005.&lt;/p&gt;&#xA;&lt;p&gt;Now, is it relevant to have rss in 2024? I&#39;m not sure the answer to that, and I&#39;m not going to spend time on it today. Stay tuned if you want to get straight in to the nitty gritty details of how to add a rss feed to any site built using golang.&lt;/p&gt;&#xA;&lt;h2&gt;What is RSS?&lt;/h2&gt;&#xA;&lt;p&gt;RSS is a way to subscribe to a website, and get updates when new content is published. It&#39;s a simple format, and a simple protocol. The protocol is based on http, and exposes itself via a url with an XML format.&lt;/p&gt;&#xA;&lt;h2&gt;How to tell the web about your RSS feed&lt;/h2&gt;&#xA;&lt;p&gt;The first step is to tell the world about your RSS feed. This is done by adding a link tag to the head of your html page. This is a simple html tag that looks like this:&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-html&#34;&gt;&amp;lt;link rel=&amp;quot;alternate&amp;quot; type=&amp;quot;application/rss+xml&amp;quot; title=&amp;quot;RSS Feed&amp;quot; href=&amp;quot;/rss.xml&amp;quot; /&amp;gt;&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;h2&gt;How to generate the RSS feed&lt;/h2&gt;&#xA;&lt;p&gt;Here&#39;s what we are trying to generate, a simple XML document that looks like the following:&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-xml&#34;&gt;&amp;lt;rss version=&amp;quot;2.0&amp;quot;&amp;gt;&#xA;    &amp;lt;channel&amp;gt;&#xA;        &amp;lt;title&amp;gt;Hacker News&amp;lt;/title&amp;gt;&#xA;        &amp;lt;link&amp;gt;https://news.ycombinator.com/&amp;lt;/link&amp;gt;&#xA;        &amp;lt;description&amp;gt;Links for the intellectually curious, ranked by readers.&amp;lt;/description&amp;gt;&#xA;        &amp;lt;item&amp;gt;&#xA;            &amp;lt;title&amp;gt;Making a PDF that&amp;amp;#x27;s larger than Germany&amp;lt;/title&amp;gt;&#xA;            &amp;lt;link&amp;gt;https://alexwlchan.net/2024/big-pdf/&amp;lt;/link&amp;gt;&#xA;            &amp;lt;pubDate&amp;gt;Wed, 31 Jan 2024 22:47:47 +0000&amp;lt;/pubDate&amp;gt;&#xA;            &amp;lt;comments&amp;gt;https://news.ycombinator.com/item?id=39210507&amp;lt;/comments&amp;gt;&#xA;            &amp;lt;description&amp;gt;&amp;lt;![CDATA[&amp;lt;a href=&amp;quot;https://news.ycombinator.com/item?id=39210507&amp;quot;&amp;gt;Comments&amp;lt;/a&amp;gt;]]&amp;gt;&amp;lt;/description&amp;gt;&#xA;        &amp;lt;/item&amp;gt;&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;The key parts that should be standing out are &lt;code&gt;title&lt;/code&gt;, &lt;code&gt;link&lt;/code&gt;, &lt;code&gt;description&lt;/code&gt; and &lt;code&gt;pubDate&lt;/code&gt;. These are what we will use to build up our own feed.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;Building all that XML by hand is a pain, so we are going to do this the easy way. Behold &lt;a href=&#34;https://xml-to-go.github.io/&#34;&gt;xml -&amp;gt; golang generation&lt;/a&gt;. This online tool is&#xA;the easy way to build up your type structs, simply grab that whole document and paste it in to the tool. It will generate the golang structs for you automagically.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;The generator will split out a big nested type. The code will contain all the necessary annotations and incantations to make your rss feed, we simply need to fill in the blanks with our own feeds parameters.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-golang&#34;&gt;rs := Rss {&#xA; //... a whole bunch of fields here&#xA;}&#xA;err := xml.NewEncoder(w).Encode(rs)&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Actual encoding can then be done using the familiar and easy to use &lt;code&gt;Encoder&lt;/code&gt; interface.&lt;/p&gt;&#xA;&lt;h2&gt;A gotcha with dates&lt;/h2&gt;&#xA;&lt;p&gt;There is one catch with the generated type, it expects the &lt;code&gt;pubDate&lt;/code&gt; to be a string, we need to serialize a date string for this field. This is all well and fine, you probably assume that a good all ISO8601 date will work.&#xA;You would be wrong to think this. It&#39;s closer to &lt;code&gt;RFC822&lt;/code&gt;. So you might think the built in RFC822 formatter from Go&#39;s great standard library. Well, almost, but not quite! The RSS feed is particularly janky, and requires us to use a 4 digit year.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;To save you a bunch of time just use &lt;code&gt;&amp;quot;02 Jan 2006 15:04 MST&amp;quot;&lt;/code&gt;. Oh, and make sure to include the timezone it was posted in. This is a requirement of the RSS spec. Check out the below example&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-golang&#34;&gt;    gmtTimeLoc := time.FixedZone(&amp;quot;GMT&amp;quot;, 0)&#xA;&#xA;&#x9;for _, param := range fp.Posts {&#xA;&#x9;&#x9;itm = append(itm, Item{&#xA;&#x9;&#x9;&#x9;Title:       param.Title,&#xA;&#x9;&#x9;&#x9;Link:        &amp;quot;https://&amp;quot; + param.URL,&#xA;&#x9;&#x9;&#x9;Description: param.Description,&#xA;&#x9;&#x9;&#x9;PubDate:     param.PubDate.In(gmtTimeLoc).Format(&amp;quot;02 Jan 2006 15:04 MST&amp;quot;),&#xA;&#x9;&#x9;&#x9;GUID: struct {&#xA;&#x9;&#x9;&#x9;&#x9;Text        string `xml:&amp;quot;,chardata&amp;quot;`&#xA;&#x9;&#x9;&#x9;&#x9;IsPermaLink string `xml:&amp;quot;isPermaLink,attr&amp;quot;`&#xA;&#x9;&#x9;&#x9;}{&#xA;&#x9;&#x9;&#x9;&#x9;Text:        &amp;quot;https://&amp;quot; + param.URL,&#xA;&#x9;&#x9;&#x9;&#x9;IsPermaLink: &amp;quot;true&amp;quot;,&#xA;&#x9;&#x9;&#x9;},&#xA;&#x9;&#x9;})&#xA;&#x9;}&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;The above is an example of how I mush my regular &lt;code&gt;Post&lt;/code&gt; type into the rss type. As you can see I&#39;m making the date field be in the gmtTimeLocation, and using the expected time format string.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-golang&#34;&gt;&#xA;## How to make the rss endpoint&#xA;&#xA;The final step is to make the endpoint that will serve the rss feed. This is a simple `http.Handler` that will serve the rss feed.&#xA;&#xA;```golang&#xA;func rssHandler(w http.ResponseWriter, r *http.Request) {&#xA;  //... generate the rss feed, posts, fill out the Rss type.&#xA;  w.Header().Set(&amp;quot;Content-Type&amp;quot;, &amp;quot;application/rss+xml&amp;quot;) // always include this&#xA;  w.WriteHeader(200) // always include this&#xA;  //... write the rss feed to the response&#xA;  err := xml.NewEncoder(w).Encode(rs)&#xA;  if err != nil {&#xA;      http.Error(w, err.Error(), 500)&#xA;  }&#xA;}&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;As you can see, it&#39;s all straight forward and familiar. That&#39;s one of the gorgeous things about RSS feeds and the go language. Simple, easy to understand and easy to implement. A great evening project to make your blog just a tiny bit more awesome.&#xA;The important part in the above example is to have a &lt;code&gt;Content-Type&lt;/code&gt; header. After all this&lt;/p&gt;&#xA;&lt;h2&gt;Wrapping up&lt;/h2&gt;&#xA;&lt;p&gt;The RSS feed is a cool little evening project I implemented to make this tiny blog platform more awesome. Golang as a language makes it easy to do all this with just the standard library. There&#39;s a few gotcha&#39;s but nothing too crazy. I hope you enjoyed this little write up, and I hope you enjoy the RSS feed. If you&#39;re reading this from a RSS feed, cool! Didn&#39;t know people actually still used these things!&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>rss</category><category>coding</category><category>beginners</category><category>golang</category><guid isPermaLink="true">https://www.ersin.nz/articles/rss-in-2024</guid></item><item><title>Screens First, Data Model Second</title><link>https://www.ersin.nz/articles/screens-first</link><description>This week we are reflecting on a recent blog post from a developer of 30 years who asserts that you should never do screens first. I disagree, and here is why.</description><pubDate>22 Jan 2024 22:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;Read a great little post from a dev reflecting on what they have learned over the last 30 years building systems. They say.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;Doing the screens first and persistence last is a common top-down approach and it is a very bad mistake&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;I respectfully disagree. But we also shouldn&#39;t just do screens first and not consider the persistence (modelling) of the data. Let&#39;s talk about the problem and a better way to build high quality software that lasts.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;The big problem&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;Like many things in life, the sure thing is change and death. Software is no different. Projects start off at a great pace, with enthusiastic engineers and users, but over time they accrue cruft. Before you realize it, the project is a nightmare to change. Velocity crawls to a halt. People leave, new engineers take forever to learn the software. It&#39;s a &#39;giant ball of mud&#39;.&lt;/p&gt;&#xA;&lt;p&gt;Is &#39;screen first&#39; or &#39;data model first&#39; going to get you to your inevitable demise faster? I think this is the wrong question!  Taking an absolute view is absolutely the wrong way, and will lead you to many WTFs on the project.&lt;/p&gt;&#xA;&lt;p&gt;A &#39;data model first&#39; approach wil harden and become brittle to new and unexpected ways of querying the data. It will accrue cruft related to features that are never needed.  In my experience I&#39;ve seen teams sink big amounts of time in to work that is never needed, which just clutters the way forward and makes the software harder to work with.&lt;/p&gt;&#xA;&lt;p&gt;&#39;Screens first&#39; as a mandate is no better. Projects that do this end up in what I call &#39;firebase hell&#39;. There is often no data model, confusing denormalisation. Sometimes, if you are in a special hellish place, you might end up with a nosql model when a relational model is desperately needed. This is an avoidable nightmare if you consider the data model from day one.&lt;/p&gt;&#xA;&lt;p&gt;Why do we take these hard positions on one side or the other? Is it a skill issue? I say no, I think it&#39;s natural -- it happens, don&#39;t beat yourself up reader, we all make mistakes. How do we fix it?&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;The perfect team&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;We need the three musketeers of software development. Product, backend and design. We need to make sure the solution actually solves the problem by (product). A designer to make the screens to solve the problem, and a data modeller to give us a model that scales beyond the life of this one application. Including this diverse set of mindsets is critical. It allows a team to succeed in the short term, while laying a foundation that can be built on in to the future.&lt;/p&gt;&#xA;&lt;p&gt;If you&#39;re working on your own, it&#39;s important to wear all the hats. Don&#39;t worry about being balanced or perfect at all of these, it&#39;s ok to be specialized in one area. Just respect that these points of view are of equal importance!&lt;/p&gt;&#xA;&lt;p&gt;The designer is going to make screens which you can put in front of customers, validating that they solve the problem. Do not worry about making pixel perfect figma designs. My favorite prototype tool is a screenshot or simple powerpoint. Use what works best and fastest, some developers I know like to do this with plain html/css if that  is there jam. There are interesting AI tools that might even work in this space such as v0.dev. The important part about this is to ask the right questions when you are talking to the customer. Ask &amp;quot;What is your biggest problem with the existing solution?&amp;quot; &amp;quot;Show me how you solve the problem now?&amp;quot;. Listen very hard to their answers, build many iterations of the screens.&lt;/p&gt;&#xA;&lt;p&gt;A good data modeller is going to be able to sit in on these interviews with customers, and their mind will spin with the ways you could persist and model this data. A graph, a document store or a relational database? The modeller will ask questions to tease out the best data model, validating or invalidating the mental model.&lt;/p&gt;&#xA;&lt;p&gt;In an ideal world, this happy trio of people (or alternate personalities), is going to be able to quickly iterate and test ideas and models of the system. I can&#39;t stress more how important it is to have everyone along for the design journey.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Wrapping this all up&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;So if you&#39;ve made it this far, I hope you are convinced that the absolute position of &#39;screens first&#39; or &#39;data model first&#39; comes from a place of pain, and will end in pain. It&#39;s a natural reaction to a bad experience. Instead, try and think about the problem you are solving, and the people you are solving it for. Try and get a good team together (or adopt multiple personalities), and iterate on the design and data model. You will be rewarded with a better product, and a better experience building the software.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>design</category><category>engineering</category><category>user testing</category><category>coding</category><category>agile</category><guid isPermaLink="true">https://www.ersin.nz/articles/screens-first</guid></item><item><title>Sailing through 2023: a reflection on Remote Work, SaaS Ventures, and Writing in public</title><link>https://www.ersin.nz/articles/2023-in-review</link><description>Join me as I recap 2023. A year of balancing remote work, a new SaaS role, creating content and becoming a dad. I&#39;ll share what went well, what didn&#39;t go so well, and what I&#39;m looking forward to in 2024.</description><pubDate>12 Jan 2024 22:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;Hello future me, if you&#39;re reading this. I&#39;m posting to answer the following questions:&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;How was 2023?&#xA;What did you learn?&#xA;What could have gone better?&#xA;What are you looking forward to in 2024?&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;Twenty three was a big year no matter how you look at it.  Starting a new role, becoming a dad and committing to a new schedule of writing in public.&lt;/p&gt;&#xA;&lt;h2&gt;What went well&lt;/h2&gt;&#xA;&lt;p&gt;&lt;strong&gt;Working from home&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;Working remote full time is something I started in the pre-covid years of 2017. In 2023, I spent my full time working remote other than 3 meet ups in person with colleagues. Back in 2017 it was a lucky chance that I happened to get an interview at a budding New Zealand based agency that was hiring exclusively remotely. This was something I picked up as an experiment, loving the idea, but also fearing failure and thinking about how it might not work out or might be tough. Instead, of my worst fears, I found a thriving pace where I used the flexibility of remote work to hit this insanely awesome life/work balance. Delivering what needed to be done, but also allowing epic adventures. Surfing, working in other countries, traveling the country in my Van. In 2023, I&#39;ve returned to remote work, but this time with 100x less epic surf adventures and about 1000x &lt;em&gt;MORE&lt;/em&gt; epic dad adventures.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;The motivation on &#39;why remote&#39; works has changed, but the core remains the same. You should choose remote work only if you can find the extrinsic motivation to make it work.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;&lt;strong&gt;New interesting saas product&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;With the end of 2022, I started a new role.  Still working on a SAAS product, in a small team, but one with many interesting diferrences. I traded backup software for managed service providers to maritime domain awareness tools. The customers couldn&#39;t be more different.  Technology wise, I kept Go, and added a suit of geospatial tech I had used in prior roles. It has been an exciting ride to say the least!&lt;/p&gt;&#xA;&lt;p&gt;My new role is focused on ingesting new data sources in to the app and working with machine learning and data science people to develop features and train models on that data. This &#39;data engineering&#39; focus is a new one for me, and one where I&#39;m enjoying being able to focus on just one aspect of the stack. Making the platform, cheaper, faster and more resillient to operate remains a really rewarding part of the role. This year I have been able to deliver a number of scale-up improvemnts to the platform, added interesting infrastructure as code and introduced new SLO&#39;s to track user experience.&lt;/p&gt;&#xA;&lt;p&gt;There have been new technologies to pick up and learn in 2023. Highlighting just a few, I have picked up Kubernetes, Terraform and google cloud for the first time. These are new flavours of technology I am already familiar with. You get a nice comparison between the platforms and tools when swapping around between them, and I&#39;m hoping this improves that tacit knowledge of when and what tool to use for the job.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;A trend I am noticing is that the technology at work is becoming more and more complex, and more optimized for multiple software teams. I get to play with technology at work that makes &lt;em&gt;zero&lt;/em&gt; sense to use on my own side projects. The problems being solved are differrent, and the constraints are different. It&#39;s interesting to think about what does scale between the side project and the team aspect, and that&#39;s an area where we (developers) should all be putting a bit more focus on.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;&lt;strong&gt;Shipping content instead of code&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;A big change at the end of 2023 was to start shifting my writing to public forums. Like most developers, I started this the wrong way, by &lt;a href=&#34;/articles/10-year-blog-platform&#34;&gt;writing a static site generator&lt;/a&gt;. My goal here has a few facets, but the most important part is to improve my writing, and start playing in the content marketing and SEO space. Beyond the learning goals, I just want to put something useful back in to the internet. I have been a blog enjoyer for most of my life, and it&#39;s time to put useful content back in to the internet.&lt;/p&gt;&#xA;&lt;p&gt;I built a few apps using the Wails framework. Not only as a fun and new way to build light weight desktop applications, but also as a good avenue for creating content. It might seem tiny, but getting a retweet from the creators was a huge personal win for me! I&#39;m hoping to continue making content on Wails in 2024.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;Making a solid push at the end of 2023 on posting blog content was one of my greatest personal achievements. Look out for more in 2024. I will aim to publish 20 blog posts between now and the end of 2024.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;h2&gt;What didn&#39;t go so great&lt;/h2&gt;&#xA;&lt;p&gt;It&#39;s hard to list things here, not just because it&#39;s the internet and I don&#39;t want to encourage even more ridicule. You see, I am reasonably optimistic, and 2023 has been a pretty great year. But let&#39;s push a bit and see what we can do to give this exercise a real chance.&lt;/p&gt;&#xA;&lt;p&gt;My CRDT work is a bit lost. I&#39;m stuck on a tricky bug with the frontend, that impacts the synchronization algorithm. This will be fixed eventually if I just do the work, but I&#39;m now keeping myself to a higher standard on publishing my journey with the blog, and sometimes the journey doesn&#39;t feel meaningful when you come to the end of your short available window of time to work on it, then it feels silly to write about such small gains.&lt;/p&gt;&#xA;&lt;p&gt;I picked up and flipped a number of tiny things that never saw the light of day. Learning Zig, Evaluating &#39;Rustic&#39;. Picking up and writing about rust.&lt;/p&gt;&#xA;&lt;p&gt;There is always that background itch that I could ship things faster, better stronger. Something that&#39;s been with me from when I first started coding at university 13 years ago.&lt;/p&gt;&#xA;&lt;p&gt;I wish I started this whole process earlier of writing and sharing earlier. Imagine how much better I would be today?&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;The main thing I&#39;m learning from all this is that when I&#39;m writing about a tech side project, I should really complete the project before starting to write about it.  This is going to lead to less pressure to deliver content + technical progress. As a result, my blog might be a bit more focused on topics that fit this model better, more story time and question/answer time, and less &#39;here is a new thing I&#39;m working on&#39;.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;h2&gt;Looking forward to 2024&lt;/h2&gt;&#xA;&lt;p&gt;My year ahead is going to be focused on three pillars; professional work, personal work, family.&lt;/p&gt;&#xA;&lt;p&gt;I&#39;m excited for working on Starboard (maritime domain awarenes). We have a lot on the road map. Technically, there are some tricky problems to solve involving large datasets and high volumes of events. Business wise, I&#39;m very excited to be more involved in sales conversations and partnerships with vendors. It is still at that sweet spot of stretching my technical skills, and giving me a lot of access to work alongside fantastic engineering/marketing/sales/product people.&lt;/p&gt;&#xA;&lt;p&gt;Publishing my work to the internet is going to be a focus of &#39;24 for my personal work. Getting better at writing, publishing code, telling the world about it in more places. The specific technical areas I&#39;m keen on doubling down in are &#39;local first&#39; stack. My writing will likely to continue being a mix of practice related chat, and deep technical posts about what I&#39;m working on.&lt;/p&gt;&#xA;&lt;p&gt;Growing as a dad. The mother of my daughter will be returning to work, and I&#39;m hoping to get the balance just right, where I can put dedicated dad time in, and get to be a part of my daughters foundational first year. If I&#39;m winning, I think I will look back and be happy with the amount of time I&#39;m not working and just dadding. Being a good husband, and father are core to who I am now, and I&#39;m going to be focusing on this pilar of life deeply. How I&#39;m going to do it is to engage fully with this double working parent of a 1 year old life, making sure the pipeline is always full with a next great family adventure involving surfing or camping or hiking.&lt;/p&gt;&#xA;&lt;p&gt;If I am killing it in &#39;24, then I will look back and say I focused on these areas authentically and earnestly. Not comprimising too much on one or the other.&lt;/p&gt;&#xA;&lt;p&gt;Twenty twenty three was a big year, I&#39;m excited for what&#39;s coming. Expecting heaps of growth and challenge as a family, in work and side-hacks on the weekend. If you&#39;ve made it this far, thanks! If you are ersin reading this from the future, hope the year went well and that this post has been useful in reflecting and planning on the next year after!&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>meta</category><category>work</category><category>life</category><category>coding</category><category>review</category><guid isPermaLink="true">https://www.ersin.nz/articles/2023-in-review</guid></item><item><title>Creating the Local First Stack</title><link>https://www.ersin.nz/articles/creating-the-local-first-stack</link><description>We continue</description><pubDate>17 Dec 2023 22:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;This week I&#39;ve been busy leading up to the holiday season working on my CRDT (conflict free replicated data type) project. My goal is to test out some of the assumptions I made about building a really effective stack for offline and local first software. Stay tuned to read more about how it went on my journey to implement a sync-server for &lt;a href=&#34;https://vlcn.io/docs/cr-sqlite/intro&#34;&gt;CRSQLite&lt;/a&gt; using Go, and a frontend using React + wasm SQLite. If you want to know a bit more about why I&#39;m doing this, check out last weeks article &lt;a href=&#34;/articles/local-first-software&#34;&gt;Local First Software&lt;/a&gt;.&lt;/p&gt;&#xA;&lt;p&gt;The foundational technology I want to build on for this project is CR-Sqlite. You can think of it as a Git layer on top of SQLite. What it allows you to do is create a normal relational schema, then mark tables as replicated, you will then able to track the changes to these tables through a simple query. On the other side of the synchronization picture, you can handle merges by inserting changes in to the database and the CR-Sqlite plugin will handle updating the right values in the database all in the background.&lt;/p&gt;&#xA;&lt;h2&gt;Keeping two databases in sync&lt;/h2&gt;&#xA;&lt;p&gt;The greatest thing about CRSQLite is that it just works on top of an existing database. We start with a schema like this:&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-sql&#34;&gt;create table foo (a primary key not null, b);&#xA;create table baz (a primary key not null, b, c, d);&#xA;select crsql_as_crr(&#39;foo&#39;);&#xA;select crsql_as_crr(&#39;baz&#39;);&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Then we can get a stream of update events made to the database with a query like this!&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code&gt;&#xA;result, err := conn.Query(&amp;quot;SELECT * FROM crsql_changes WHERE db_version &amp;gt; ? &amp;quot;, currentVersion)&#xA;&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;The special relation &lt;code&gt;crsql_changes&lt;/code&gt; contains the column, row and database version for that column and row. Using this data we can &#39;materialize&#39; a view of the table at a certain version. &lt;a href=&#34;https://github.com/ebuckley/cr-sqlite-go/blob/faedd627513a442ddba780b9c44cd68bac7cd4e1/crsql/crsql_changes.go#L21&#34;&gt;Check out the implementation&lt;/a&gt;, you can simply query the database to get the list of changes needed to synchronize two databases&lt;/p&gt;&#xA;&lt;p&gt;To merge changes, simply &lt;code&gt;INSERT&lt;/code&gt; into &lt;code&gt;crsql_changes&lt;/code&gt; and your rows/columns will seamlessly be updated. This is really cool because it&#39;s just plain old SQL, and it&#39;s dead simple to interact with from Go, using the &lt;code&gt;database/sql&lt;/code&gt; package. You can check out the implementation at the time of writing &lt;a href=&#34;https://github.com/ebuckley/cr-sqlite-go/blob/faedd627513a442ddba780b9c44cd68bac7cd4e1/crsql/crsql_changes.go#L45&#34;&gt;here&lt;/a&gt;.&lt;/p&gt;&#xA;&lt;p&gt;See this all working together in the &amp;quot;dumb version&amp;quot; implemented in &lt;a href=&#34;https://github.com/ebuckley/cr-sqlite-go/blob/main/crsql/sync_test.go&#34;&gt;this test&lt;/a&gt;. The test creates two databases in memory, inserts values into one and merges changes between each other to validate that the two databases end up in the same state. After implementing this I&#39;m feeling confident that the method will work, and I have everything needed for the back-end server to implement merging of database states.&lt;/p&gt;&#xA;&lt;p&gt;Lets move on.&lt;/p&gt;&#xA;&lt;h2&gt;Implementing the server&lt;/h2&gt;&#xA;&lt;p&gt;How shall we merge the &lt;code&gt;[]Change&lt;/code&gt; values in a meaningful way over the network? My test just passes them through memory and that&#39;s not really realistic for our desired application. The requirement I have is to support many web clients and a simple server in the background for caching and orchestrating the web clients. The server is just like a client, except it&#39;s always online and available for peers to use as a source for synchronizing data.&lt;/p&gt;&#xA;&lt;p&gt;We can solve this with a service! Now there are many ways I could have started, but I decided to test out gRPC along the way. This was a mistake. I hoped for the best, but gRPC ended up not being a good choice for the web client. Why? you ask. The gRPC protocol works with all the bells and whistles of http when used server to server, but web clients are not as great. The Javascript client is dependent on http 2.0, and it requires a proxy like Envoy to work with a browser. What&#39;s more, I didn&#39;t love the structure of the generated web client. So through the process of working on this &#39;local first stack&#39; I actually got sucked in to a big rabbit hole in making the rpc system work. I ended up going with &lt;a href=&#34;https://connectrpc.com/&#34;&gt;Connect&lt;/a&gt; which is a tool that can create a service from a protobuf service definition, that also talks a simple http 1.1 protocol. What ultimately sold me on this solution as the best is that it also came with a very nice to use web client generation, and even plugs in to my favorite react http helper &lt;a href=&#34;https://tanstack.com/query/v4/docs/react/reference/useQuery&#34;&gt;useQuery&lt;/a&gt;.&lt;/p&gt;&#xA;&lt;p&gt;So with the backend all done and implemented, I have a simplified service definition &lt;a href=&#34;https://github.com/ebuckley/cr-sqlite-go/blob/main/api/v1/service.proto&#34;&gt;here&lt;/a&gt;, an implementation of the server in go and a client implemented in typescript.&lt;/p&gt;&#xA;&lt;h2&gt;Implementing the frontend&lt;/h2&gt;&#xA;&lt;p&gt;The coolest thing in this whole stack is that we have the same exact database on the frontend as the backend. A build of the whole SQLite project, including the CR-Sqlite plugin. This is cutting edge stuff, let&#39;s go dive in and see how it works.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;The database is this really high performance low level piece of systems programming, how the heck is it running on the web?&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;Great question, dear reader! Despite the database being written in C and Rust, we can use this from javascript using a simple database oriented API, in the exact same way as we interacted with the db in our backend.&lt;/p&gt;&#xA;&lt;p&gt;The technology that unlocks this is Web Assembly (WASM). A very low level language that can run in the browser, and can expose API&#39;s to your plain old javascript. This is something that comes in the browser, and more and more it is becoming widely used on the web. Compiling the C and rust code with LLVM enables the CR-SQLite project to run on the web. I can use all this really advanced and deeply complicated stuff through simple npm imports. Got to love the 2023 way of plugging in thousands of lines of advanced database with a few lines of &lt;code&gt;bun install&lt;/code&gt; :D&lt;/p&gt;&#xA;&lt;h2&gt;Wrapping it all up&lt;/h2&gt;&#xA;&lt;p&gt;Now the frontend is far from perfect so far, there are in-fact critical bugs with the synchronization work, but I have the bones of the project setup -- and over the next weeks I plan to continue refining and improving on it so that I can test my theory about Local First software being a simpler, faster and better way to build software. It&#39;s amazing how much work I can leverage already in the CRDT space, thanks to the way the web as a platform has evolved and the focus and efforts of people to develop great open source tools for RPC, user interface and databses.&lt;/p&gt;&#xA;&lt;p&gt;And one last note, if you&#39;ve made it this far. Thanks for reading! It would be awesome to hear from you, so feel free to reach out on the twitter, or the comments, help me push this project (writing) a bit further. My goal has been to write once a week, and I&#39;m hoping that the process is one that will help me hone my craft.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>crdt</category><category>sqlite</category><category>side projects</category><category>connect rpc</category><category>grpc</category><guid isPermaLink="true">https://www.ersin.nz/articles/creating-the-local-first-stack</guid></item><item><title>Local First Software</title><link>https://www.ersin.nz/articles/local-first-software</link><description>This week I explore the concept of local first software, and how it can help you build better software.</description><pubDate>10 Dec 2023 22:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;&lt;img src=&#34;/public/new-fangled-vs-walled-garden.png&#34; alt=&#34;&#34;&gt;&lt;/p&gt;&#xA;&lt;p&gt;The 2023 way is to lock your data in a walled garden, keeping it safe and secure in someone else&#39;s computer (a server). You put trust in the service and hope for the best.  Hope for no  leaks, hacks, security breaches or rogue employees. When the service ends hopefully there are backups available and a way to extract your own data. This pattern is one most of you are familiar with. It&#39;s the way web 2.0 works, and by 2023 it&#39;s old hat!&lt;/p&gt;&#xA;&lt;p&gt;Today; we talk about a niche suite of algorithms that enables a better way, Conflict Free Replicated Data Types (CRDTs) enable an eventual data sync data between peers. This can be, servers/browsers/phones or anything in between. It&#39;s a cool alternative to locking data behind a server, and it makes some hard features easy&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Users always having a copy of their own data&lt;/li&gt;&#xA;&lt;li&gt;Easy support for offline mode&lt;/li&gt;&#xA;&lt;li&gt;Simplified architecture&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;The local-first software movement aims to put that control back in users&#39; hands. Groups like Ink &amp;amp; Switch Research have been doing some fascinating work in this space. Their insights into eventually consistent, yet resilient systems inspired me to embark on my own learning journey with CRDTs.&lt;/p&gt;&#xA;&lt;p&gt;In this series, I plan to document my progress building a small prototype application using conflict-free replicated data types. CRDTs provide a neat way to replicate application state across devices without worrying about sync conflicts. The goal is to see these patterns in action by implementing database synchronization between server, desktop, mobile, and browser-based peers of the app.&lt;/p&gt;&#xA;&lt;h2&gt;What are CRDTs?&lt;/h2&gt;&#xA;&lt;p&gt;&lt;img src=&#34;/public/CRDT-merge.png&#34; alt=&#34;&#34;&gt;&lt;/p&gt;&#xA;&lt;p&gt;So what exactly are conflict-free replicated data types? They are specially designed data structures that allow for eventual consistency of application state. What I mean by eventual consistency is that nodes can update their local copy of state independently and immediately then propagate changes to other nodes asynchronously, yet all nodes eventually converge to the same global state due to a mathematically-guaranteed merge procedure.&lt;/p&gt;&#xA;&lt;p&gt;This magic lies in the CRDTs&#39; commutative, associative, and idempotent merge function.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;Basically, no matter what order changes or updates arrive across nodes, applying them results in the same output state.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;The most simple example of this is a Set. If multiple nodes perform an &#39;Add&#39; operation, the same eventual state is reached. When you need to &#39;merge&#39; state you simply do a &#39;Union&#39; operation of each state and a final shared state is reached. The fantastic part is it doesn&#39;t matter what order of events happen on each of the peers, it can seamlesly be merged together at some point of time in the future. Pretty nifty! This is the key property that eliminates update conflicts even across disconnected and independent replicas.&lt;/p&gt;&#xA;&lt;p&gt;By removing the need for coordinating events, you get this awesome property of local first editing. Immediate feedback can be given to the user, and whenever a sync is possible it can happen in the background&lt;/p&gt;&#xA;&lt;h2&gt;The prototype&lt;/h2&gt;&#xA;&lt;p&gt;Theory is great, but how can we apply this in practice? Instead of starting from 0, and writing a CRDT, let&#39;s try and leverage an existing project to do the heavy lifting. My choice is &lt;a href=&#34;https://github.com/vlcn-io/cr-sqlite&#34;&gt;crSQLITE&lt;/a&gt;, an extension for SQLite to support CRDT merging of databases. Under the hood, the extension creates tables to track changes and allow inserting into an event log for merging states of separated peers.&lt;/p&gt;&#xA;&lt;p&gt;What&#39;s great about choosing SQLite as a platform, is it is a ubiquitous development environment. I can get this built for every platform, and all my favorite libraries are setup to work with SQL already.&lt;/p&gt;&#xA;&lt;p&gt;Now, I wish I could share with you the skeleton of requirements for a big amazing project, but I&#39;m just not quite there yet. Ideas I&#39;ve been thinking about include:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;An extension to my markdown editor &lt;a href=&#34;https://www.ersin.nz/articles/markdown-editor-with-wails-react-tailwind&#34;&gt;markdown editor in tailwind react and wails&lt;/a&gt;&lt;/li&gt;&#xA;&lt;li&gt;Simple diagram tool&lt;/li&gt;&#xA;&lt;li&gt;Time tracker&lt;/li&gt;&#xA;&lt;li&gt;????&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;No matter what, my goal with this is to test out this CRSqlite thing, and build something desktop based using wails, which can be run on my desktop, laptop(s), phone and anything with a web browser. So without further ado, I&#39;m introducing the newfangled &lt;a href=&#34;https://github.com/ebuckley/cr-sqlite-go&#34;&gt;cr-sqlite-go&lt;/a&gt;, a simple library that wraps the CRSQLite extension and provides a simple API for interacting with the database.&lt;/p&gt;&#xA;&lt;p&gt;My vision is to eventually build an application that seamlessly synchronizes across peers, with offline support and including as many platforms as possible. Hopefully along the way we can learn some things about CRDT, and give my stack choice a good workout.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>crdt</category><category>computer science</category><category>side projects</category><category>technical blog</category><guid isPermaLink="true">https://www.ersin.nz/articles/local-first-software</guid></item><item><title>Should I hack on the weekends?</title><link>https://www.ersin.nz/articles/balancing-growth-with-personal-time</link><description>Balancing professional growth and personal time is hard. This week I answer: Is it important to keep working on side projects after that first job?</description><pubDate>01 Dec 2023 22:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;&lt;img src=&#34;/public/contributions-gh.png&#34; alt=&#34;910 contributions in the last year, not bad ersin!&#34;&gt;&lt;/p&gt;&#xA;&lt;p&gt;Balancing professional growth and personal time is a constant challenge in our industry.  Inspired by the question in episode &lt;a href=&#34;https://softskills.audio/2020/05/18/episode-210-study-time-and-caring-less/&#34;&gt;210&lt;/a&gt; of the soft skills engineering podcast,I am reflecting on my own approach to hiring and interviewing for jobs and how we should think about side projects and the github profile.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;Before I got my job as a web developer, I was studying 2 hours per night. Now that I&#39;ve started working (been here for two years) I want to come home and relax! How much time do you spend reading about new technology, or working on side projects? I plan on getting a new job in the future, and I feel like I need to start studying again. My github is empty since I started this job, should I be worried about that?&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;When you first start in the industry, you will be going from this time where you probably spent every waking hour looking to make yourself ready for work by doing every leetcode/sideproject that will make you look like a good candidate. This changes almost immediately as you sink in to a 40 hour week of actually doing the work. It&#39;s completely natural to not want to code at home. This is healthy and good. Some of the best engineers I have worked with haven&#39;t written a single line of code at home. You learn on the job, and you get to grow with every new problem you solve.&lt;/p&gt;&#xA;&lt;p&gt;An empty Github Profile might feel confronting when it comes to moving from work mode to searching mode, but reality is, it doesn&#39;t matter! A delicious github graph filled with green squares might be a great dopamine hit for you, but it&#39;s not going to be the deciding factor for hiring someone who already has industry experience.&lt;/p&gt;&#xA;&lt;p&gt;Now, for those of you who know me, you know I&#39;m a huge fan of side projects. That is like 50% of the purpose of this blog is to share from my own experiences from coding for fun. But even I haven&#39;t always felt the need to hardcore code all day and all night. When I&#39;ve had really engaging or demanding professional work, I&#39;m more likely to go outside and touch grass, or spend time with my family. I think that is a healthy balance. There are times when it&#39;s the opposite though, if my day job isn&#39;t involving much code, and more in the analysis space I will find time to pique curiosity with side-code.&lt;/p&gt;&#xA;&lt;p&gt;So on to the next part; you&#39;re looking for that second job? Well, the github profile isn&#39;t going to be a factor (sorry). A full profile may be pivotal in your first role, but in my experience it plays a much smaller part in each subsequent role. What&#39;s more important is that you can talk about the work you have done professionally. Be very precise in telling the story of your work; Describe the problem, explain the context and then dive in to how you use your tools to solve it. This is the kind of thing that will get you hired.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;When does it make sense to pick up side projects ahead of a job search?&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;This might make sense if you are looking at picking up a new technology. If you can find a way to branch out within your current role, do it. Experience with real world projects is 1000x more valuable for lateral shifts. Finally, if you can&#39;t shift your current role to the new stack/skill, then totally put your time into a nice real world project simulation.&lt;/p&gt;&#xA;&lt;p&gt;Focus on distilling a project to just the parts you might need in interviewing in the new role, and aggressively cut scope on anything you already have experience in within your current role. This is the kind of thing that will optimize your time, and get the most pay-off for shifting to a new skillset.&lt;/p&gt;&#xA;&lt;p&gt;In my own experience, side projects have infrequently come up as a topic of discussion. What does come up more frequently is the knowledge gained from working on the project. Being able to compare my experience on a project to the role I&#39;m interviewing for has been helpful in giving context about who I am as a technical contributor, and how I might approach the trade-offs from technology decisions.&lt;/p&gt;&#xA;&lt;p&gt;Navigating this balance between professional growth and personal growth is something that will always be a tension. There is no obvious one-size-fits-all approach. Whether you hack on the weekends or not, prioritize a balanced approach. Try not to get caught up on expectations for your side projects to mean anything in the interview. Be real with yourself about what you are working on achieving here. Don&#39;t Fixate on the Github contributions graph. Finding a healthy work-life balance is the most important thing of all, and the best way to do that is to listen to your gut, and don&#39;t take things too seriously.&lt;/p&gt;&#xA;&lt;p&gt;Thanks for reading! I&#39;m eager to hear your own experience with side projects and job seeking :)&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>beginner</category><category>career</category><category>side projects</category><category>technical blog</category><guid isPermaLink="true">https://www.ersin.nz/articles/balancing-growth-with-personal-time</guid></item><item><title>A Decade of Golang</title><link>https://www.ersin.nz/articles/should-i-learn-golang</link><description>Reflections on a decade of Golang, how to pick niche technologies for your software engineering career.</description><pubDate>24 Nov 2023 22:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;&lt;img src=&#34;/public/old-office.webp&#34; alt=&#34;&#34;&gt;&#xA;The year is 2013, you&#39;ve just landed your first real engineering job and you get to churn out angular.js and PHP. What is even better, you are paid a handsome junior salary to do it. In other words you are living the dream.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;If I could ask one thing of my future self at the time it would be. Should I learn the Go language?&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;The Go language was really fresh at the time. It had come up in one or two conversations with my mentor. The hype online was starting to build, interesting tools like docker were being built on the language, and it was popping up all the time on the front page of Hacker News.  What&#39;s more, it was supported by google.&lt;/p&gt;&#xA;&lt;p&gt;New Zealand had very few opportunities to use Go professionally. We tend to be microsoft centric down under. On the face of it, Golang doesn&#39;t look like the best tool to pick up next. If job optionality was most important, one would be better off learning C# or Java.&lt;/p&gt;&#xA;&lt;p&gt;Despite all this, you should learn Go first young Ersin. Start building side projects, don&#39;t fret that there are no job opportunities yet, even if it doesn&#39;t end up in paid work, you will learn along the way. You will learn a new type system, you will be exposed to the kind of problems the language is good at solving (concurrency, distributed services, systems programming). The types of blog posts and tutorials coming out centered around go will stretch your knowledge and make your more well rounded. This is the kind of thing that compounds and makes you better years down the track.&lt;/p&gt;&#xA;&lt;p&gt;The side projects you build might seem inconsequential, but each new project you build hones that muscle. Reading code and trying new things are how you grow. When the day job becomes routine, a side project will keep you invested in your craft. The effort will pay off eventually, I promise.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;If you take one point from this, always take a bet on niche and emerging technologies.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;The prevailing wisdom is that you should stick with learning the most popular tools, to give you the best chance of landing jobs. I think doing just the safe bet is a mistake. We should adopt a barbell strategy in technology. Build a solid base with safe choices to earn your way in to the industry -- then put time into many niche, emerging and wildly interesting technologies. This gives you the security of stable work, while broadening your horizons in a direction that brings you joy. Eventually, an opportunity will come up in one of these niches.&lt;/p&gt;&#xA;&lt;p&gt;It is an amazing thing to be ahead of the curve on an emerging niche technology. Projects are more likely to be green-field. The people who are drawn to organizations that use them are typically more engaged in the work, and you will find a higher degree of alignment in the team. People aren&#39;t just there for the money or because they want to get a job at all costs. Not to mention if that tool becomes wildly successful, you will be an outlier expert of which no one can catch up&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;What about in 2023?&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;So a decade later, If I was asked the same question &amp;quot;Should I learn go?&amp;quot;; I would say Yea! but the reasons are different now. Although it&#39;s still a niche technology in New Zealand, there is always at least a couple jobs on the board. You could bet your career on Go as the &#39;safe&#39; option.&lt;/p&gt;&#xA;&lt;p&gt;After that first job, you might be better off focusing your efforts on a more radical technology. Instead of Go, branch out and pick up another emerging niche technology. You never know where it will take you, so find what piques your interest and dive in head first! It&#39;s hard to picture where you will be in ten years time, so best not get overly focused on the outcome but instead focus on the journey.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>beginner</category><category>learning</category><category>golang</category><category>technical blog</category><guid isPermaLink="true">https://www.ersin.nz/articles/should-i-learn-golang</guid></item><item><title>Crafting a lightweight markdown editor</title><link>https://www.ersin.nz/articles/markdown-editor-with-wails-react-tailwind</link><description>This week we create a tiny markdown editor using Wails, React and Tailwind.</description><pubDate>16 Nov 2023 22:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;&lt;img src=&#34;/public/writer-screenshot.png&#34; alt=&#34;&#34;&gt;&lt;/p&gt;&#xA;&lt;h1&gt;Building a Markdown Editor with Wails, React and Tailwind&lt;/h1&gt;&#xA;&lt;p&gt;Are you tired of dealing with the complexities and massive builds associated with Electron for desktop applications? Do traditional frameworks like QT, GTK, or Win32 make you want to cry?&#xA;Today we are going to build a markdown editor using Wails, React and Tailwind.&lt;/p&gt;&#xA;&lt;h2&gt;Simplifying Desktop App Development&lt;/h2&gt;&#xA;&lt;p&gt;Building a desktop application has been an ugly duck since the dawn of web 2.0. The situation has been even more dire for golang enjoyers.&#xA;You are forced into a few options, mostly wrapping existing UI libraries, and non of which really bring the productivity and joy of writing backend&#xA;services using golang. The most popular option, Electron, is a huge stack of javascript, and while it&#39;s easy to get started, it&#39;s bloated and you are forced&#xA;to write so much javascript that it&#39;s going to make you pull your hair out.&lt;/p&gt;&#xA;&lt;p&gt;Let&#39;s delve into a better way, by creating a markdown editor using Wails, React and Tailwind.&lt;/p&gt;&#xA;&lt;p&gt;Our basic requirements for this app will be:&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code&gt;    preview markdown text&#xA;    edit markdown text&#xA;    create a file&#xA;    save a file&#xA;    export a file (as pdf)&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;h2&gt;Step 1: Bootstrap!&lt;/h2&gt;&#xA;&lt;p&gt;Wails follows a Rails-like philosophy, providing a batteries are included and opinions come thick and fast. The opinion it does not have, is what html framework you use for the view layer. To make things really easy to get started, it makes it straight forward to kick start from templates.&lt;/p&gt;&#xA;&lt;p&gt;Let&#39;s get started from this awesome pre-made vite+react+ts+tailwind template.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;wails init -n write -t https://github.com/hotafrika/wails-vite-react-ts-tailwind-template&#xA;cd write&#xA;wails dev&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;Queue waiting forever to let npm install + build for the first time. This is a one time thing, and it&#39;s worth the wait, I promise!&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;Take a second now to get oriented, the structure is like so, and every wails app will look the same:&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code&gt;- build (contains built binaries, installers, etc)&#xA;- frontend (contains the frontend code)&#xA;- wailsjs (contains the go code)&#xA;- main.go (the entrypoint for the go code)&#xA;- wails.json (the wails configuration file)&#xA;- app.go (the main wails application file)&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Let&#39;s start with the design: here is what I &lt;a href=&#34;https://v0.dev/r/Z0Ln82RKNlP&#34;&gt;used&lt;/a&gt;, thanks to v0.dev for saving me all the time! This is pretty much a copy/paste job of the html into &lt;a href=&#34;https://github.com/ebuckley/write/blob/335bc096973f3b5a368e0257c1ffcd053a6c513c/write/frontend/src/App.tsx&#34;&gt;app.tsx&lt;/a&gt;. I chose not to import the v0 library -- I&#39;m just not convinced it&#39;s worth it for such a simple design.&lt;/p&gt;&#xA;&lt;h2&gt;Step 2: Render and update markdown&lt;/h2&gt;&#xA;&lt;p&gt;Wails provides a server/client model, seamlessly wiring the Golang Backend to the Javascript frontend. In our markdown editor component you will see some simple snippets like this:&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code&gt;// at the start of the file&#xA;import {Save, SendMarkdownToRenderer, Open, Export} from &amp;quot;../wailsjs/go/main/App&amp;quot;;&#xA;&#xA;&#xA;// inside the Component&#xA;&#xA; const updateValue = async (event: React.ChangeEvent&amp;lt;HTMLTextAreaElement&amp;gt;) =&amp;gt; {&#xA;&#xA;   const value = event.target.value&#xA;&#xA;   const html = await SendMarkdownToRenderer(value)&#xA;&#xA;   setMd(html);&#xA;&#xA; }&#xA;&#xA;// many lines skipped ...&#xA;&#xA;&amp;lt;textarea&#xA;&#xA;           onChange={updateValue}&#xA;&#xA;           ref={ref}&#xA;&#xA;           className=&amp;quot;flex min-h-[80px] rounded-md border border-input px-3 py-2 text-sm ring-offset-background placeholder:text-muted-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2 disabled:cursor-not-allowed disabled:opacity-50 w-full h-full bg-white dark:bg-gray-700&amp;quot;&#xA;&#xA;           id=&amp;quot;markdown-input&amp;quot; placeholder=&amp;quot;Type your markdown here...&amp;quot;&amp;gt;&amp;lt;/textarea&amp;gt;&amp;lt;/div&amp;gt;&#xA;&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;As you can see, we have a really simple textarea with an onChange function. When the textarea changes, an event is fired, the function gets called and we execute &lt;code&gt;SendMarkdownToRenderer&lt;/code&gt;. The function is a code generated client that maps to Go code. Let&#39;s check that out.&lt;/p&gt;&#xA;&lt;p&gt;https://github.com/ebuckley/write/blob/335bc096973f3b5a368e0257c1ffcd053a6c513c/write/app.go#L55C1-L60C1&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code&gt;&#xA;func (a *App) SendMarkdownToRenderer(content string) string {&#xA;&#xA;   a.markdown = content&#xA;&#xA;   a.renderedHTML = lib.RenderMD(content)&#xA;&#xA;   return a.renderedHTML&#xA;&#xA;}&#xA;&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;This fantastically simply little function is how we persist state to the &#39;server&#39; and go context, and then pass back a string value to the client for rendering via the &lt;code&gt;setMD&lt;/code&gt; state setter.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Important:&lt;/strong&gt; there is a lot of cool stuff going on here. First of all, we have this really awesome code generation that just automatically happens as you save your files. A go function gets seamlessly and magically turned into something callable from a react application. This kind of magic RPC is really effective for a UI application where the server is close to the client, and it&#39;s just quite delightful for working with.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;A fun digression: If you squinted at this you could look at the recent tech twitter memes about &lt;code&gt;&#39;use server&#39;&lt;/code&gt; and call this &lt;code&gt;&#39;use golang&#39;&lt;/code&gt;.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;h2&gt;Step 3: New file, Save File, Open file&lt;/h2&gt;&#xA;&lt;p&gt;To implement this trio of features, we are going to use the same exact pattern using &lt;code&gt;onClick&lt;/code&gt; handlers for buttons. Again, we leverage the codegen that wails provides to automagically wire up. Let&#39;s dive into the Save function in app.go:&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-golang&#34;&gt;&#xA;func (a *App) Save() (string, error) {&#xA;&#xA;   hd, err := homedir.Dir()&#xA;&#xA;   if err != nil {&#xA;&#xA;      return &amp;quot;&amp;quot;, err&#xA;&#xA;   }&#xA;&#xA;&#xA;   chosenPath, err := runtime.SaveFileDialog(a.ctx, runtime.SaveDialogOptions{&#xA;&#xA;      DefaultDirectory:          path.Join(hd, &amp;quot;Documents&amp;quot;),&#xA;&#xA;      DefaultFilename:           a.filename,&#xA;&#xA;      Title:                     &amp;quot;Save file&amp;quot;,&#xA;&#xA;      ShowHiddenFiles:           false,&#xA;&#xA;      CanCreateDirectories:      true,&#xA;&#xA;      TreatPackagesAsDirectories: true,&#xA;&#xA;   })&#xA;&#xA;   if err != nil {&#xA;&#xA;      return &amp;quot;&amp;quot;, err&#xA;&#xA;   }&#xA;&#xA;   err = os.WriteFile(chosenPath, []byte(a.markdown), 0644)&#xA;&#xA;   if err != nil {&#xA;&#xA;      return &amp;quot;&amp;quot;, err&#xA;&#xA;   }&#xA;&#xA;   return chosenPath, nil&#xA;&#xA;}&#xA;&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;The framework provides a runtime package which allows calling a few functions which provide common cross-platform utility. Here, we call &lt;code&gt;SaveFileDialog&lt;/code&gt; to present the user with a native and familiar way of choosing a filename and location to put the markdown. The function returns a path &lt;code&gt;chosenPath&lt;/code&gt; and we use that in a straight up &lt;code&gt;WriteFile&lt;/code&gt; call.&lt;/p&gt;&#xA;&lt;p&gt;Our state &lt;code&gt;a.markdown&lt;/code&gt; is kept in sync with the frontend via the previous &lt;code&gt;onChange&lt;/code&gt; handler from step 1.&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;Travel warning: The SaveFileDialog takes a context, it&#39;s really important that this is a runtime context, the runtime context is stored on the app struct, and you need to always remember to pass this so that all the magic works! This planned to change in wails v3, but that is still in development at the time of writing this.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;&lt;strong&gt;challenge&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;Take a look and see if you can find out how the New File and Open File work. Start out in app.tsx and trace your way back to app.go!&lt;/p&gt;&#xA;&lt;h2&gt;Step 4: Yak shaving markdown export&lt;/h2&gt;&#xA;&lt;p&gt;Ok this turned out to be a &lt;strong&gt;&lt;em&gt;H U G E&lt;/em&gt;&lt;/strong&gt; timewaste, but it was a &lt;em&gt;lot&lt;/em&gt; of fun. As you might be aware, there are a lot of techniques for exporting PDF files, I decided to go for the native and non browser based pure pdf library fpdf. Thankfully, our markdown -&amp;gt; html compiler also supports extending via a renderer function and walking the parsed AST. I had a great time adapting a scuffed and minimalist little pdf renderer. You can check it out &lt;a href=&#34;https://github.com/ebuckley/write/blob/main/write/lib/pdf.go&#34;&gt;here&lt;/a&gt;.&lt;/p&gt;&#xA;&lt;p&gt;For the wails part, it&#39;s again very simply an &lt;code&gt;onclick&lt;/code&gt; handler for a button alongside an exported method on &lt;code&gt;App.go&lt;/code&gt; We can simply take that markdown content and put it through our renderer, passing a path to save as a result from the &lt;code&gt;SavefileDialog&lt;/code&gt;&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;Warning: don&#39;t use this pdf renderer code, it crashes on codeblocks! It&#39;s just an example approach which I think could be extended quite nicely.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;h2&gt;Step 5: CI/CD&lt;/h2&gt;&#xA;&lt;p&gt;Eat your vegetables, make your code hygienic and never leave a bad build on &lt;code&gt;origin main&lt;/code&gt;.&lt;/p&gt;&#xA;&lt;p&gt;I went to the effort of setting up CI/CD for this project to show how you can cross compile for windows and linux. To do this I setup a github action which tests/builds on push and releases when you tag the repo with something matching &lt;code&gt;v*&lt;/code&gt;.&lt;/p&gt;&#xA;&lt;p&gt;Yes. You are seeing this correctly, it took me 16 builds to get a release. Anyway... So that you don&#39;t need to waste as much time, check out my deploy action &lt;a href=&#34;https://github.com/ebuckley/write/blob/main/.github/workflows/main.yml&#34;&gt;here&lt;/a&gt;.&lt;/p&gt;&#xA;&lt;p&gt;The interesting parts are:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;You need to install a few dependencies &lt;code&gt;sudo apt install libgtk-3-dev libwebkit2gtk-4.0-dev nsis&lt;/code&gt;&lt;/li&gt;&#xA;&lt;li&gt;Don&#39;t forget to get the paths right&lt;/li&gt;&#xA;&lt;li&gt;Don&#39;t forget to use the permissions block in your action to allow it to cut new releases&lt;/li&gt;&#xA;&lt;li&gt;Copy the binary out of the build directory between builds, because &lt;code&gt;build/bin&lt;/code&gt; will be overwritten on builds&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;As a result, I think we have a solid base for a project which can be extended, bug smashed and it all gets automagically released through github actions. A nice place to stop for this little exercise.&lt;/p&gt;&#xA;&lt;h2&gt;Wrapping it all up with a disruptive and transformative conclusion&lt;/h2&gt;&#xA;&lt;p&gt;Wails is a gamechanger for GUI applications written in Golang -- simple, productive and downright magical.&#xA;You can think of it as a &#39;web style&#39; MVC for your frontend, or &lt;code&gt;&#39;use golang&#39;&lt;/code&gt; for desktop applications. What sets it apart? Bring any frontend view layer! HTML to Svelte, it all works. Binaries are tiny 10Mb instead of electrons 100+Mb. As far as Golang desktop frameworks go, I think Wails is undisputed as our best option.&lt;/p&gt;&#xA;&lt;p&gt;You can check out the code I wrote for this project &lt;a href=&#34;https://github.com/ebuckley/write&#34;&gt;here&lt;/a&gt;, and get started with wails &lt;a href=&#34;https://wails.io/&#34;&gt;here&lt;/a&gt;.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>tutorial</category><category>wails</category><category>golang</category><category>react</category><category>electron</category><category>technical blog</category><guid isPermaLink="true">https://www.ersin.nz/articles/markdown-editor-with-wails-react-tailwind</guid></item><item><title>How I solved: code cracker</title><link>https://www.ersin.nz/articles/how-i-solved-kata-code-cracker</link><description>This week i solve the code cracker problem in rust</description><pubDate>10 Nov 2023 22:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;Are you looking for just the challenge with no spoilers? Stop now, go back to &lt;a href=&#34;/articles/quick-kata-code-cracker&#34;&gt;last weeks post&lt;/a&gt;!&lt;/p&gt;&#xA;&lt;p&gt;Still here? Great! Let&#39;s dive in to the solution.&#xA;Why did I do this in rust? Well.. it&#39;s not because I had the need for extreme memory safety. I didn&#39;t need the zero cost abstractions, but i guess they are nice.&#xA;The thick and well supplied third party ecosystem wasn&#39;t needed. Didn&#39;t need the safe concurrency.&lt;/p&gt;&#xA;&lt;p&gt;I chose rust because I&#39;m interested in the language, and I wanted to pick it up again.&lt;/p&gt;&#xA;&lt;p&gt;Recently, I went along to the local rust meetup in Christchurch, and I enjoyed the vibe. People are interested in talking allocations, safe handling of errors, and&#xA;working with systems that require some high performance.&lt;/p&gt;&#xA;&lt;p&gt;Writing systems that need to go fast is really fun. You get a sweet dopamine hit from making memory go down and cpu flame graphs chill out.&lt;/p&gt;&#xA;&lt;p&gt;My hope is that rust is a tool that helps me stretch that urge, and at some point in the future I would very much like to apply it to a professional problem.&lt;/p&gt;&#xA;&lt;p&gt;So let&#39;s dive in to the code. I&#39;m going to skip over the boring bits, and focus on the interesting parts.&lt;/p&gt;&#xA;&lt;h2&gt;Part one, finding possible words:&lt;/h2&gt;&#xA;&lt;p&gt;Let&#39;s start by building our dictionary of possible words. I&#39;m depending on the built in dictionary file in linux, on my machine I chose &lt;code&gt;/usr/share/dict/american-english&lt;/code&gt;.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-rust&#34;&gt;fn get_dictionary(dictionary_path: &amp;amp;str) -&amp;gt; Vec&amp;lt;String&amp;gt; {&#xA;    let mut words = Vec::new();&#xA;    for line in read_to_string(dictionary_path).unwrap().lines() {&#xA;        // if it contains &#39; then strip it out and save it...&#xA;        let mut thisline = line.to_string();&#xA;        if line.contains(&amp;quot;&#39;&amp;quot;) {&#xA;            thisline = thisline.replace(&amp;quot;&#39;&amp;quot;, &amp;quot;&amp;quot;).to_string();&#xA;        }&#xA;        words.push(thisline.to_uppercase());&#xA;    }&#xA;    words&#xA;}&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;As you can see, I&#39;m reading the file in to a string, and then iterating over the lines. I&#39;m also stripping out the &lt;code&gt;&#39;&lt;/code&gt; character, because it is not a valid character in our code cracker alphabet.&lt;/p&gt;&#xA;&lt;p&gt;Next we need to represent our hint and substitute alphabet as a type. As what you might find after years and years of programming, most problems can be solved with simple vectors and hashmaps. So let&#39;s represent our substitution alphabet with an array, and our hint with a vec.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-rust&#34;&gt;// a hint is a Vec&amp;lt;i32&amp;gt; where each item in the hint represents an index into our substituion alphabet&#xA;&#xA;struct Problem {&#xA;    code: [Option&amp;lt;char&amp;gt;; 26],&#xA;    hints: Vec&amp;lt;Vec&amp;lt;i32&amp;gt;&amp;gt;&#xA;}&#xA;&#xA;&#xA;fn find_options(hint: &amp;amp;Vec&amp;lt;i32&amp;gt;, code:[Option&amp;lt;char&amp;gt;; 26], words: &amp;amp;Vec&amp;lt;String&amp;gt;) -&amp;gt; Vec&amp;lt;String&amp;gt; {&#xA;    let word_length = hint.len();&#xA;    let mut options = Vec::new();&#xA;    &#39;outer: for word in words.iter() {&#xA;        if word.len() != word_length {&#xA;            continue&#xA;        }&#xA;        // this might be it!&#xA;        options.push(word.clone())&#xA;    }&#xA;    options&#xA;}&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;In this super simple first cut, we just iterate through all the possible words, and check if the length of the word matchces the hint, if it does, push it on to the options vector.&lt;/p&gt;&#xA;&lt;h2&gt;Part two, optimizing options:&lt;/h2&gt;&#xA;&lt;p&gt;We are going to extend find_options so that we can select just the words that make sense. To do this we need a function to fill in the known parts of our hint word with the actual &lt;code&gt;char&lt;/code&gt; from the substitution alphabet.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-rust&#34;&gt;fn to_letter_map(hint: &amp;amp;Vec&amp;lt;i32&amp;gt;, code: [Option&amp;lt;char&amp;gt;; 26]) -&amp;gt; HashMap&amp;lt;usize, char&amp;gt; {&#xA;    let mut out = HashMap::new();&#xA;    for n in 0..hint.len() {&#xA;        let elem = hint[n];&#xA;        let codeindex = (elem - 1) as usize;&#xA;        let opt = code[codeindex];&#xA;        match opt {&#xA;            Some(v) =&amp;gt; {&#xA;                out.insert(n, v);&#xA;            }&#xA;            None =&amp;gt; {&#xA;                continue&#xA;            }&#xA;        }&#xA;    }&#xA;    out&#xA;}&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;This one builds up a HashMap that maps the substitution alphabet value to it&#39;s actual character, so we can easily lookup what letters to search for.&lt;/p&gt;&#xA;&lt;p&gt;Now we have what we need to breakdown words with letters that are already known, lets attempt to filter the word list even more by checking that they match the &#39;pattern&#39; laid out by the hint.&#xA;So for our hint &lt;code&gt;1 7 7 R 7 21&lt;/code&gt; we would expect the word to have the same letter for the 2nd, 3rd and 5th letters. We implement a &lt;code&gt;check_word&lt;/code&gt; function to do so.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-rust&#34;&gt;fn check_word(hint: &amp;amp;Vec&amp;lt;i32&amp;gt;, code:[Option&amp;lt;char&amp;gt;; 26], word: &amp;amp;String) -&amp;gt; bool {&#xA;    assert_eq!(hint.len() , word.len(), &amp;quot;should be the same length&amp;quot;);&#xA;&#xA;    let mut code = code.clone();&#xA;&#xA;    for (i, c) in hint.iter().enumerate() {&#xA;        let current_word_char = word.chars().nth(i).unwrap();&#xA;        let id = (c-1)as usize;&#xA;        let matches_code = code[id];&#xA;        match matches_code {&#xA;            Some(code_value) =&amp;gt; {&#xA;                // if the current value exists as a code but doesn&#39;t match the current word&#xA;                if current_word_char != code_value {&#xA;                    return false&#xA;                }&#xA;&#xA;            },&#xA;            None =&amp;gt; {&#xA;                code[id] = Some(current_word_char);&#xA;            }&#xA;        }&#xA;    }&#xA;    return true&#xA;}&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Of note in this function, we use a &#39;clone()` on our substitution alphabet (code). This is important because we want to not mutate the original one, and this function will be called many times as we check each candidate word.&lt;/p&gt;&#xA;&lt;p&gt;Now we can extend the find_options function, putting together our word_map and check_word functions into the loop.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-rust&#34;&gt;fn find_options(hint: &amp;amp;Vec&amp;lt;i32&amp;gt;, code:[Option&amp;lt;char&amp;gt;; 26], words: &amp;amp;Vec&amp;lt;String&amp;gt;) -&amp;gt; Vec&amp;lt;String&amp;gt; {&#xA;    let word_length = hint.len();&#xA;    // word_vec is a map of index of hint -&amp;gt; char, built from referencing the code (substitution alphabet)&#xA;    let word_vec = to_letter_map(hint, code);&#xA;    let mut options = Vec::new();&#xA;    &#39;outer: for word in words.iter() {&#xA;        if word.len() != word_length {&#xA;            continue&#xA;        }&#xA;        // NEW: in this loop we check that the candidate word matches the letters already known from the hint.&#xA;        for (index, word_char) in word_vec.iter() {&#xA;            if word.chars().nth(*index).is_some_and(|c| c != *word_char){&#xA;                continue &#39;outer // if it&#39;s not then bail to the outer loop `outer&#xA;            }&#xA;        }&#xA;&#xA;        let matches_pattern = check_word(hint, code, word);&#xA;        if !matches_pattern {&#xA;            continue &#39;outer&#xA;        }&#xA;        // this might be it!&#xA;        options.push(word.clone())&#xA;    }&#xA;    options&#xA;}&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;This is how far I got with the challenge! you can check out the code on github.&lt;/p&gt;&#xA;&lt;p&gt;I&#39;m really new to rust, and this was a fun one to flex my Rustacean skills. Looking forward to putting new skills to use in real projects some time in the near future :)&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>codecracker</category><category>kata</category><category>puzzle</category><category>technical blog</category><guid isPermaLink="true">https://www.ersin.nz/articles/how-i-solved-kata-code-cracker</guid></item><item><title>Quick kata: code cracker</title><link>https://www.ersin.nz/articles/quick-kata-code-cracker</link><description>A quick and fun code kata is to implement this algorithm to solve a code cracker puzzle.</description><pubDate>02 Nov 2023 22:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;This code cracker challenge is a great exercise to flex your programming muscles, especially in something new (like I did with rust). Let&#39;s break it down:&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Challenge 1: Finding Possible Words&lt;/strong&gt;&#xA;Create a function that takes input text and outputs the substitution alphabet and each clue. For example:&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;example input&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code&gt;14 R&#xA;1 7 7 14 7 21&#xA;14 7 3 7 14 1 7&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;&lt;strong&gt;example output&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code&gt;Alphabet:&#xA;1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26&#xA;_ _ _ _ _ _ _ _ _  _  _  _  _  R  _  _  _  _  _  _  _  _  _  _  _  _&#xA;&#xA;Clue 1:&#xA;1 7 7 R 7 21&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Now, search through a dictionary (e.g., &lt;code&gt;/usr/dict/words&lt;/code&gt;) to find words matching the clue&#39;s length.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Challenge 2: Optimizing Options&lt;/strong&gt;&#xA;Print out all possible words that match the pattern. For &lt;code&gt;1 7 7 R 7 21&lt;/code&gt;, the output would be:&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code&gt;1 7 7 R 7 21&#xA;&#xA;0: DEERES&#xA;1: JEERED&#xA;2: LEERED&#xA;3: PEERED&#xA;4: VEERED&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;&lt;strong&gt;Challenge 3: Play the Game&lt;/strong&gt;&#xA;Allow the player to fill in new letters in the substitution alphabet to play the next word. This could be a handy hint buddy for cracker enthusiasts.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Challenge 4: Solving a Full Game&lt;/strong&gt;&#xA;Create the ability to input an entire game, letting the program figure out a solution.&lt;/p&gt;&#xA;&lt;p&gt;This challenge is not only a great exercise but also a fantastic way to learn a new programming language. Happy coding!&lt;/p&gt;&#xA;&lt;p&gt;Looking for some more samples, check out the original authors work! &lt;a href=&#34;https://codecracker.co.nz/&#34;&gt;codecracker.co.nz&lt;/a&gt;&lt;/p&gt;&#xA;&lt;p&gt;As for my own solution? I&#39;ll post it soon, promise!&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>codecracker</category><category>kata</category><category>puzzle</category><category>technical blog</category><guid isPermaLink="true">https://www.ersin.nz/articles/quick-kata-code-cracker</guid></item><item><title>Developer spends 10 years building blog platform</title><link>https://www.ersin.nz/articles/10-year-blog-platform</link><description>How I&#39;m relaunching my blog after 10 years of working on it.</description><pubDate>29 Oct 2023 15:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;Ok I&#39;m kidding. I may have been writing code for cash for ten years now, but this &lt;em&gt;tiny&lt;/em&gt; platform has been the product of two weeks casually plugging away in my editor, in between looking after my 5 month old daughter and working full time on &lt;a href=&#34;starboard.nz&#34;&gt;starboard.nz&lt;/a&gt;.&lt;/p&gt;&#xA;&lt;p&gt;This blog exists for three reasons.  I wanted to learn something new, liberate my data from the centralized blogging platforms, and experiment with SEO.&lt;/p&gt;&#xA;&lt;p&gt;My goal is to build a platform that can last for 100 years.&lt;/p&gt;&#xA;&lt;p&gt;So Buckle up, we&#39;re going deep.&lt;/p&gt;&#xA;&lt;p&gt;Yea I know it&#39;s not going to be here for even a half century.  I&#39;m setting my sites long term, and I&#39;ll be happy if it lasts a quarter of that. Shoot for the moon, hit the outer atmosphere. Big goals are important, no?&lt;/p&gt;&#xA;&lt;p&gt;My rationale is that making my blog tech so simple, so boring will let me &lt;em&gt;finish&lt;/em&gt; and move on to the next thing. The biggest impediment to building a 100 year blog, is gaining the inertia to keep writing, and build an audience that can pull you towards continuing in your writing.&lt;/p&gt;&#xA;&lt;p&gt;You are probably asking; Why not post on a major platform Ersin? Major platforms suck, they don&#39;t help me with distribution or gaining an audience, and they don&#39;t let me mess around with some code and learn something new! At least through this process I am getting to &lt;em&gt;build&lt;/em&gt; and I&#39;m &lt;em&gt;learning&lt;/em&gt; something new at the same time.&lt;/p&gt;&#xA;&lt;p&gt;Don&#39;t get me wrong, if you&#39;re already famous you probably should use a platform for posting your content. But I&#39;m not there yet. I&#39;m still building.&lt;/p&gt;&#xA;&lt;p&gt;If you&#39;re not interested in running a $5 vm, you definitely shouldn&#39;t be doing this. If your life isn&#39;t spent in git, then you should run now. Stop reading! Go away!&lt;/p&gt;&#xA;&lt;p&gt;Great, now the noobs have been filtered out, let&#39;s continue.&lt;/p&gt;&#xA;&lt;p&gt;This blog has been back-filled with some things I have written over on &lt;a href=&#34;https://dev.to/ebuckley&#34;&gt;dev.to&lt;/a&gt;. Not only is it important to bring my old stuff in to the future, it&#39;s important to keep my content rolling. When you are working on software, having &lt;em&gt;real&lt;/em&gt; data is also amazing and gives you the best possible experience for design. Let&#39;s not spend our days looking at lorem ipsum boilerplate eh?&lt;/p&gt;&#xA;&lt;p&gt;My old blogging platform helpfully provided a tool to export your data (thanks gdpr), and this comes in the form of a simple JSON structure, this let me build a little tool which converts old blogs to new blogs.&lt;/p&gt;&#xA;&lt;p&gt;So what does a new blog look like? Glad you asked!&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code&gt;Blog posts are in hashicorp config language&#xA;The content is markdown&#xA;They can sit in a file under a folder and will automagically be loaded and embedded in to the binary.&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;pre&gt;&lt;code&gt;page &amp;quot;render-massive-lists-on-the-web-2h3c-temp-slug-7179880&amp;quot; {&#xA;    title = &amp;quot;Render massive lists on the web&amp;quot;&#xA;    description = &amp;quot;Rendering a huge list is almost never the right solution, but this post is here to help you out if it...&amp;quot;&#xA;    keywords = [&amp;quot;javascript&amp;quot;]&#xA;    content = &amp;lt;&amp;lt;EOT&#xA;Rendering a huge list is almost never the right solution, but this post is here to help you out if it is, and you want to do it as fast as possible.&#xA;EOT&#xA;  posted = &amp;quot;2023/04/05&amp;quot;&#xA;}&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;I am aiming to avoid needing a database for this application. All the content will be in the git repo, and the fact that it uses HCL is more of a quirk in that I wanted to pique my interest of the language. In 2023 I have picked up terraform for the first time, and it&#39;s just so much better than writing YAML. I want to spend more time writing code than HCL, and hopefully the least amount of time writing YAML.&lt;/p&gt;&#xA;&lt;p&gt;Let&#39;s be real, no one really likes the look of tailwind when they first pick it up. But after 4 years of skepticism, I&#39;m all in on utility first CSS. Now I&#39;m not the best at slicing and dicing boxes on the web, but I&#39;ve been centering divs since &lt;code&gt;float:&lt;/code&gt; was the best technology we had. Utility first CSS is something that has been sitting in the background for longer than tailwind has. Tailwind is having it&#39;s moment because the old way was too rigid. We can copy a design system and easily extend it to new patterns when using tailwind, that&#39;s much much harder with the old way.&lt;/p&gt;&#xA;&lt;p&gt;HTMX is my other pick here. Right now I&#39;m only using it in two places. We are &lt;a href=&#34;https://htmx.org/attributes/hx-boost/&#34;&gt;boosting&lt;/a&gt;, which should result in a marginally faster click through for you. And we are building the subscribe/contact form using our dear htmx. A tool that is gaining huge popularity now, based on a technique that has been happening for about 33 years? I&#39;m all about it &lt;a href=&#34;https://www.ersin.nz/articles/htmx-is-old-but-gold-4i9i&#34;&gt;htmx is old but gold&lt;/a&gt;.&lt;/p&gt;&#xA;&lt;p&gt;If you&#39;ve made it this far. Thanks for hanging! What can you expect in the future? More technical content, my goal with this blog is to help make you a better programmer, and give you some ingisht in to the mind of a decade experienced software engineer. If I&#39;m smashing goals out of the park, I will post weekly, but at a minimum I will post something once a month. For now you can expect blog posts, source code releases and the infrequent real world meetup talk.&lt;/p&gt;&#xA;&lt;p&gt;What&#39;s next? Well, after a really energizing session at the local rust meetup, I&#39;m hoping to pickup my learning journey there. On another note, I have countless dead projects that might be worth a deep dive. Or we could dive in to the real world of software development, and maybe I could answer a reader question.&lt;/p&gt;&#xA;&lt;p&gt;Thanks for hanging!&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>golang</category><category>htmx</category><category>tailwind</category><category>technical blog</category><guid isPermaLink="true">https://www.ersin.nz/articles/10-year-blog-platform</guid></item><item><title>A Gophers hack for marketing: The Tampermonkey and AI Strategy You Need</title><link>https://www.ersin.nz/articles/a-gophers-hack-for-marketing-the-tampermonkey-and-ai-strategy-you-need-3719</link><description>I had a very specific need. Finding the current state of an Airbnb market in a specific location....</description><pubDate>15 Apr 2023 05:41 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;I had a very specific need. Finding the current state of an Airbnb market in a specific location. Yea, tools for this exist, but that cost money. Here is a more interesting and indirect route. The result is a weird mash up of AI with a simple tampermonkey script. Yea, it did take twice as long as needed, but I had fun and learned a bit more about userscripts for your browser.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Part one: Get that Data&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;&lt;em&gt;Full disclosure: I asked the AI to write this section, based on the tamper monkey code I wrote here. Prompt &amp;quot;describe this code from the perspective of a gonzo journalist gopher&amp;quot;. Comments in the code are my own.&lt;/em&gt;&lt;/p&gt;&#xA;&lt;p&gt;Hold on to your burrows, fellow mammals, because this gopher has got some JavaScript code to analyze! This script is a browser extension that digs deep into an Airbnb listing to download data automatically.&lt;/p&gt;&#xA;&lt;p&gt;To start, there&#39;s a function called waitForElementToExist, which is like a gopher waiting to pounce on some tasty roots. This function keeps burrowing through the page until it finds an element that matches the provided CSS selector. When it finds it, it sends a signal back to the main script that the root is ready to be harvested.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-js&#34;&gt;function waitForElementToExist(selector) {&#xA;  return new Promise((resolve, reject) =&amp;gt; {&#xA;    // An element might not be on the page, so we retry every 100 ms to find the element.&#xA;    // Yea, it might block forever but it&#39;s good enough!&#xA;    let intervalId = setInterval(() =&amp;gt; {&#xA;      let element = document.querySelector(selector);&#xA;      if (element) {&#xA;        clearInterval(intervalId);&#xA;        resolve(element);&#xA;      }&#xA;    }, 100);&#xA;  })&#xA;}&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Once all the juicy elements have been harvested, the script goes into full tunnel mode with Promise.all. It waits for all of the promises returned by waitForElementToExist to be fulfilled, then snatches the innerText from each element and stores it like a gopher storing food for the winter.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-js&#34;&gt;  // get at the data we want&#xA;  const [cost, title, descriptionText] = await Promise.all([&#xA;    // Airbnb uses these &amp;quot;lovely&amp;quot;, &amp;quot;semantic&amp;quot; class names (not!).&#xA;    waitForElementToExist(&#39;._tyxjp1&#39;).then(e =&amp;gt; e.innerText),&#xA;    // I think we can thank a css in js tool for this :grin:&#xA;    waitForElementToExist(&#39;.hghzvl1&#39;).then(e =&amp;gt; e.innerText),&#xA;    waitForElementToExist(&#39;.ll4r2nl&#39;).then(e =&amp;gt; e.innerText)]);&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;After that, the script builds a burrow of data by creating a JavaScript object with all the data it has gathered, including the cost, title, description text, URL, and number of rooms. Then it uses its keen eyesight to log that data to the console like a gopher signaling its friends.&lt;/p&gt;&#xA;&lt;p&gt;Here&#39;s where things get really interesting. The script creates a Blob object from the data as a JSON string, and then creates a tunnel-like URL for the Blob object. What is a Blob object, you ask? It&#39;s like a burrow of soil, only it&#39;s a binary large object that can be used to store and transfer data. And that URL? It&#39;s like a tunnel that leads straight to a JSON file where the data resides.&lt;/p&gt;&#xA;&lt;p&gt;Finally, the script surfaces like a gopher to create a link element, sets some attributes like download and href to the juicy data-filled Blob URL, and then simulates a click on the link. Just like that, the JSON file containing all the data is downloaded faster than a gopher can dig a new burrow.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-js&#34;&gt;// Create a Blob object from the data as a JSON string&#xA;var blob = new Blob([JSON.stringify( {cost, title, descriptionText, url, rooms}, null, 2)], {type: &#39;application/json&#39;});&#xA;&#xA;// Create a URL for the Blob object&#xA;var url2 = URL.createObjectURL(blob);&#xA;&#xA;// Create a link element and click it to download the file&#xA;var link = document.createElement(&#39;a&#39;);&#xA;link.download = title.toLowerCase().split(&#39;&#39;).filter(c =&amp;gt; c.match(/[a-z]/i)).join(&amp;quot;&amp;quot;) + &#39;.json&#39;;&#xA;link.href = url2;&#xA;link.click();&#xA;// the end result here is that when you visit a page, it will automagically download a json file with the&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;So, my fellow gophers, that&#39;s how this script works. It&#39;s like the gopher of browser extensions, digging deep to gather data and bringing it back to the surface for all to see.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;step 2: Use open AI to classify and find the most relevant listing to your property&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;For the next part, I load all this data I scraped from airbnb into a database, then analyze each row of the data and get some unique AI insights from each one.&lt;/p&gt;&#xA;&lt;p&gt;Interesting? let me know.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>webdev</category><category>javascript</category><category>beginners</category><category>productivity</category><guid isPermaLink="true">https://www.ersin.nz/articles/a-gophers-hack-for-marketing-the-tampermonkey-and-ai-strategy-you-need-3719</guid></item><item><title>RSMQ for golang</title><link>https://www.ersin.nz/articles/rsmq-for-golang-2ej5</link><description>RSMQ is the most simple Queue Implementation in the known universe. My own implementation here comes...</description><pubDate>22 Mar 2022 22:03 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;RSMQ is the most simple Queue Implementation in the known universe. My own implementation &lt;a href=&#34;https://github.com/ebuckley/rsmq&#34;&gt;here&lt;/a&gt; comes in at under 500 lines of code. Join me for a short tour of the code, and how it works.&lt;/p&gt;&#xA;&lt;p&gt;People might be asking why I went to this effort when there are already so many implementations? Educational reasons. Implementing RSMQ is a fine and tightly scoped project that can be completed in a couple days of after work effort. In the day job we have been using a PHP implementation of the library but I wanted to deepen my knowledge of this simple tool by creating an implementation in my favorite language!&lt;/p&gt;&#xA;&lt;p&gt;If you have Redis already, RSMQ is probably the easiest way to add messaging between different languages, processes and machines!&#xA;There are PHP/Python/Javascript/Java/C#/Rust/Go implementations now available for your usage. Want to write your own? No problem! Read this and become convinced that this is a simple and basic implementation task.&lt;/p&gt;&#xA;&lt;h2&gt;How does RSMQ work?&lt;/h2&gt;&#xA;&lt;p&gt;Create Queues, send Messages, receive them on the diaspora of devices/processes and languages. Queues contain messages. Messages are delivered in a first in-first out basis (think about the queue in line at the McDonloads).&lt;/p&gt;&#xA;&lt;p&gt;A Message is only delivered to one consumer. A consumer, being the thing that takes the message and finishes the work, deleting the message at the end. There is a time limit for how long a message can be processed for. When the time runs out, the message re-enters the queue and can be picked up by another consumer.&lt;/p&gt;&#xA;&lt;p&gt;A typical worker implementation will receive a message, do the work, and then delete the message once it has processed successfully. If the worker crashes during the work, the message will be re-entered on the queue automatically, and another consumer will be able to pick it up.&lt;/p&gt;&#xA;&lt;h1&gt;Implementing our own.&lt;/h1&gt;&#xA;&lt;p&gt;Some Redis knowledge assumed. You should become familiar with the basics. Today we will be using the following data structures.&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Sets (contains the list of queues)&lt;/li&gt;&#xA;&lt;li&gt;Hashes (contains the stats for a specific queue, and messages)&lt;/li&gt;&#xA;&lt;li&gt;Sorted Sets (Zset)&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;h2&gt;Creating a new message Queue&lt;/h2&gt;&#xA;&lt;p&gt;A new queues is simply a unique Hash and a member added to a set.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-go&#34;&gt;key := rsmq.ns + &amp;quot;:&amp;quot; + opts.QName + &amp;quot;:Q&amp;quot;&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;A queue is uniquely identified by the key. This is how we reference the queue, send messages to it, pop from it, update attributes about the queue. You will be using similar identifiers to this throughout your code in a lot of places.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-go&#34;&gt;_, err = rsmq.cl.HMSet(ctx, key, map[string]interface{}{&#xA;    &amp;quot;createdby&amp;quot;: &amp;quot;ersin&amp;quot;,&#xA;    &amp;quot;vt&amp;quot;:        30, // TODO allow this to be set with CreateQueueRequestOptions&#xA;    &amp;quot;delay&amp;quot;:     0,&#xA;    &amp;quot;maxsize&amp;quot;:   65536,&#xA;    &amp;quot;created&amp;quot;:   unixMilliseonds,&#xA;    &amp;quot;modified&amp;quot;:  unixMilliseconds,&#xA;}).Result()&#xA;if err != nil {&#xA;    return fmt.Errorf(&amp;quot;CreateQueue: set queue params: %w&amp;quot;, err)&#xA;}&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;We define a hash for the attributes of the queue, including informative things like the delay in seconds before a message becomes receivable and the maximum size of the message in the queue.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-go&#34;&gt;_, err = rsmq.cl.SAdd(ctx, rsmq.ns+&amp;quot;:QUEUES&amp;quot;, opts.QName).Result()&#xA;if err != nil {&#xA;    return fmt.Errorf(&amp;quot;CreateQueue: add queue to QUEUES set: %w&amp;quot;, err)&#xA;}&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Finally, we add the name of the queue to a set key. This allows our library to list all queues that are created under a given namespace.&lt;/p&gt;&#xA;&lt;h2&gt;Sending a message to the queue&lt;/h2&gt;&#xA;&lt;p&gt;Code first, description second. It looks like this!&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-go&#34;&gt;pipe := rsmq.cl.Pipeline()&#xA;pipe.ZAdd(ctx, key, &amp;amp;Redis.Z{&#xA;    Score:  float64(q.TimeSent.Add(sendTime).UnixMilli()),&#xA;    Member: q.UID,&#xA;})&#xA;pipe.HSet(ctx, key+&amp;quot;:Q&amp;quot;, q.UID, opts.Message)&#xA;pipe.HIncrBy(ctx, key+&amp;quot;:Q&amp;quot;, &amp;quot;totalsent&amp;quot;, 1)&#xA;_, err = pipe.Exec(ctx)&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Sending a message is a three step process that must happen as Redis &#39;pipeline&#39;. A pipeline ensures that the actions do not require a round trip between client and server.&lt;/p&gt;&#xA;&lt;ol&gt;&#xA;&lt;li&gt;Add the message ID to the Sorted Set &lt;code&gt;rsmq.ns + &amp;quot;:&amp;quot; + opts.QName&lt;/code&gt; with a score being the unix milliseconds of when it should send. Later on, we will re-score this message which has the effect of hiding it from other Q consumers.&lt;/li&gt;&#xA;&lt;li&gt;Set the message content as a value on the &lt;code&gt;rsmq.ns + &amp;quot;:&amp;quot; + opts.QName + &amp;quot;:Q&amp;quot;&lt;/code&gt; hash. The key of the content is the message ID&lt;/li&gt;&#xA;&lt;li&gt;Increment the &lt;code&gt;totalSent&lt;/code&gt; field of the Q hash used in step (2).&lt;/li&gt;&#xA;&lt;/ol&gt;&#xA;&lt;p&gt;Note that we execute these in a &#39;pipe&#39;, which ensures that these steps happen atomically on the Redis data store, so we can be confident that the 3 commands always happen in sequence.&lt;/p&gt;&#xA;&lt;h2&gt;ReceiveMessage&lt;/h2&gt;&#xA;&lt;p&gt;Our final operation for this code review is to receive a message from the queue. This step would be called by the workers that do the actual processing.&lt;/p&gt;&#xA;&lt;p&gt;When a message is received it comes with an explicit deadline set for processing the message. This is known as the &#39;Visibility Timeout&#39; and it is the number of second in which this message will be exclusively available to the one worker. If the visibility timeout expires, then the message will end up being sent to another worker.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-golang&#34;&gt;results, err := rsmq.cl.EvalSha(&#xA;    ctx,&#xA;            // this references the embedded LUA procedure&#xA;    *rsmq.receiveMessageSha1,&#xA;            // these parameters are presented as KEY to the LUA script &#xA;    []string{key, timeSentUnix, timeVisibilityExpiresUnix}).Slice()&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Only a single line on the Go side of things. We call a Lua script which runs on Redis to do the heavy lifting. An embedded Lua script runs in an atomic an thread-safe way on the datastore. This is how we get the guarantees of only one worker receiving a message at a time.&lt;/p&gt;&#xA;&lt;p&gt;The following parameters to the Lua:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;KEYS[1]. Is the name of the sorted set containing the Q &lt;code&gt;rsmq.ns + &amp;quot;:&amp;quot; + opts.QName&lt;/code&gt;&lt;/li&gt;&#xA;&lt;li&gt;KEYS[2]. Is the time when a message has been requested from the Q.&lt;/li&gt;&#xA;&lt;li&gt;KEYS[3]. Is the time when a message should become visible to other consumers again.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;Now let&#39;s step through the Lua procedure.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-lua&#34;&gt;local msg = Redis.call(&amp;quot;ZRANGEBYSCORE&amp;quot;, KEYS[1], &amp;quot;-inf&amp;quot;, KEYS[2], &amp;quot;LIMIT&amp;quot;, &amp;quot;0&amp;quot;, &amp;quot;1&amp;quot;)&#xA;&#x9;&#x9;&#x9;if #msg == 0 then&#xA;&#x9;&#x9;&#x9;&#x9;return {}&#xA;&#x9;&#x9;&#x9;end&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Remember that there is a sorted set scored by timestamp sent for each message in the queue. The above code returns exactly one or zero message that is due for return between the &lt;code&gt;-inf&lt;/code&gt; score and the current timestamp provided &lt;code&gt;KEYS[2]&lt;/code&gt;.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-lua&#34;&gt;Redis.call(&amp;quot;ZADD&amp;quot;, KEYS[1], KEYS[3], msg[1])&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Next we re-score the current message found with the new visibility timeout defined in the &lt;code&gt;KEYS[3]&lt;/code&gt; parameter. This step is important, because it keeps our queue in the data store. Other workers will be able to pick up this message again if it is not deleted before the visibility time has expired.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-lua&#34;&gt;Redis.call(&amp;quot;HINCRBY&amp;quot;, KEYS[1] .. &amp;quot;:Q&amp;quot;, &amp;quot;totalrecv&amp;quot;, 1)&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;Now we increment the &lt;code&gt;totalrecv&lt;/code&gt; counter on the Hash which contains the Q.&lt;/p&gt;&#xA;&lt;p&gt;The final block of code concerns returning the message.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-lua&#34;&gt;local mbody = Redis.call(&amp;quot;HGET&amp;quot;, KEYS[1] .. &amp;quot;:Q&amp;quot;, msg[1])&#xA;local rc = Redis.call(&amp;quot;HINCRBY&amp;quot;, KEYS[1] .. &amp;quot;:Q&amp;quot;, msg[1] .. &amp;quot;:rc&amp;quot;, 1)&#xA;local o = {msg[1], mbody, rc}&#xA;if rc==1 then&#xA;&#x9;Redis.call(&amp;quot;HSET&amp;quot;, KEYS[1] .. &amp;quot;:Q&amp;quot;, msg[1] .. &amp;quot;:fr&amp;quot;, KEYS[2])&#xA;&#x9;table.insert(o, KEYS[2])&#xA;else&#xA;&#x9;local fr = Redis.call(&amp;quot;HGET&amp;quot;, KEYS[1] .. &amp;quot;:Q&amp;quot;, msg[1] .. &amp;quot;:fr&amp;quot;)&#xA;&#x9;table.insert(o, fr)&#xA;end&#xA;return o&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;An HINCRBY is called to increment our count of how many times this message has been received.&lt;/p&gt;&#xA;&lt;p&gt;To wrap it all up, the message content is returned along with a few additional infomrative attributes.&lt;/p&gt;&#xA;&lt;h2&gt;Want to implement your own?&lt;/h2&gt;&#xA;&lt;p&gt;My dreams would come true if I inspired you to write something like this for your own learning purposes. It would make my week complete if you left a comment.&lt;/p&gt;&#xA;&lt;p&gt;All Credit for the original implementation of RSMQ goes to https://github.com/smrchy/rsmq :)&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>go</category><category>redis</category><category>showdev</category><guid isPermaLink="true">https://www.ersin.nz/articles/rsmq-for-golang-2ej5</guid></item><item><title>Learn functional, learn systems and learn object oriented</title><link>https://www.ersin.nz/articles/learn-functional-learn-systems-and-learn-object-oriented-10ie</link><description>If you are a junior or intermediate, you should consider picking up projects or languages that help...</description><pubDate>11 Oct 2021 19:00 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;If you are a junior or intermediate, you should consider picking up projects or languages that help you round out the functional, object oriented and systems trio of languages. After 10 years of putting text files into compilers and interpreters, here is my take.&lt;/p&gt;&#xA;&lt;h2&gt;why we should do it&lt;/h2&gt;&#xA;&lt;p&gt;We should do it as a form of practice and professional development, even if we do not need it right now at our day job. Learning new languages increases enjoyment of coding, and brings fresh ideas to your craft.&lt;/p&gt;&#xA;&lt;h2&gt;Why functional?&lt;/h2&gt;&#xA;&lt;p&gt;Understanding pure functions is the reason you should pick up a functional programming language. This is one of the key ideas to writing test-able code. If you are new to functional programing, the experience can be a very steep hill climb. The reward in my opinion is the most impactful on the quality of code regardless of what other language you are using.&lt;/p&gt;&#xA;&lt;p&gt;My recommended functional language is Clojure. The reason I picked up a lisp is that the language has a very strong foundation and an incredibly simple syntax. I believe it to be a &#39;stripped down&#39; feel which helps you really zero in on the core concepts in the paradigm.&lt;/p&gt;&#xA;&lt;h2&gt;Why OOP?&lt;/h2&gt;&#xA;&lt;p&gt;An object oriented language is important to learn because it introduces a very familiar vernacular for modelling process, business and the world. Simply put, it helps you communicate about software projects with other people. On a technical level. Object oriented languages fill a pretty broad source of options, so it could be a difficult choice. For my own purpose I have chosen to become an expert in golang. Although a purist may rightly claim that Go is not OOP, I believe it fills that same niche. You can use the &lt;code&gt;interface&lt;/code&gt; and &lt;code&gt;struct&lt;/code&gt; features to achieve polymorphism.&lt;/p&gt;&#xA;&lt;p&gt;The different object oriented languages I have used in my day job to this stage is vast. As you pick up the next one, it becomes faster and easier to get productive in the language.but The concepts boil down to just a few differences, and they all share many strong similarities in how they approach state, assignment and memory management.  PHP, Python, Java, Golang, c#, Javascript being the ones I have personally used.&lt;/p&gt;&#xA;&lt;h1&gt;Why Systems?&lt;/h1&gt;&#xA;&lt;p&gt;A systems language completes the set of types of languages you should learn. You will learn to appreciate the high level of abstraction you can achieve with different languages. When I started working on projects with systems languages I also learned a greater depth about my operating system, infrastructure and memory management.&lt;/p&gt;&#xA;&lt;p&gt;As programmers, our job is to create magic inside the box. When you understand the lower level of the abstraction you get to unveil the magic for what it is. Having a solid basis in a systems language will help you unveil problems when garbage collection or operating systems features are causing issues with performance&lt;/p&gt;&#xA;&lt;p&gt;Right now I think the go language is my favorite pick for a systems language. It allows you to access the OS API with relative ease, and you get the compiled language which opens up really interesting project opportunities in the ops, sysadmin and SRE space.&lt;/p&gt;&#xA;&lt;p&gt;In addition to this, it is worthwhile understanding the power of manual memory management. You can get this with languages like, c, rust or d. I wouldn&#39;t go so far as to say it is the most important concept to learn, but it can give you magical super powers when you really need code to perform in a reliable and fast way.&lt;/p&gt;&#xA;&lt;h1&gt;What it means to have a solid base&lt;/h1&gt;&#xA;&lt;p&gt;In an average coders career, you will learn a lot of different languages. Practicing learning languages opens up opportunities for picking up the most interesting projects. It broadens the array of problems you can solve. Hones your craft no matter what tool you are using to write that next best project.&lt;/p&gt;&#xA;&lt;p&gt;It matters not so much which language you pick, but try to target the variety of niches which let you bring in the good ideas from the other platform.&lt;/p&gt;&#xA;&lt;p&gt;What are your three picks under the FP, OOP and systems category?&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>systems</category><category>javascript</category><category>clojure</category><category>go</category><guid isPermaLink="true">https://www.ersin.nz/articles/learn-functional-learn-systems-and-learn-object-oriented-10ie</guid></item><item><title>htmx is Old but Gold</title><link>https://www.ersin.nz/articles/htmx-is-old-but-gold-4i9i</link><description>What is htmx?  htmx is this tool I have been flirting with for creating dynamic web applications. It...</description><pubDate>11 May 2021 09:16 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;&lt;strong&gt;What is htmx?&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;htmx is this tool I have been flirting with for creating dynamic web applications. It is a heretical tool of sorts. Cutting my teeth, writing Javascript every day with the latest framework is something I have been doing day to day for the last 10 years. This tool (htmx) is a pragmatic choice for developing modern web applications using a technique that would be common in a web app from 2004. Our industry runs in a cycle, and this tool is the return of a productive and useful trend in my opinion.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Old ways and New ways&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;Back in 2004, javascript was entering it&#39;s heyday and becoming a useful tool for building dynamic web applications. But we still had sites like craigslist, ebay and online banking. Fast forward, today we can do all of this and much more. Javascript is a much better language in 2021 than 2004. So what is the problem?&lt;/p&gt;&#xA;&lt;p&gt;Modern web frameworks and architecture is optimized for big teams. It is a tool born out of the need to scale huge amounts of developers to all work on one interface.  We do not need the overhead of the current frameworks. They bring more weight than they need.&lt;/p&gt;&#xA;&lt;p&gt;We ship so much javascript to the client side, it just slows down the experience. Your browser is becoming a huge drain on your battery power, and the rest of us are being excluded from the web.&lt;/p&gt;&#xA;&lt;p&gt;The modern web application stack can take a few different shapes and forms, so bear with me explaining the straw man. You start with a component framework like React. Sprinkle in something for state, let&#39;s say RXJS. Put on an &lt;code&gt;npm i react-router&lt;/code&gt; to manage routing to pages. Drop in the data layer, your API is graphql so you put on an apollo client.&lt;/p&gt;&#xA;&lt;p&gt;At this stage you&#39;ve loaded 100&#39;s of kb&#39;s of javascript and you&#39;re not even rendering the page yet.&lt;/p&gt;&#xA;&lt;p&gt;The architecture is optimized for teams with a backend and frontend separation, but that does not represent all the teams and produces and ways that we rwork.&lt;/p&gt;&#xA;&lt;p&gt;Testing becomes a nightmare as you have more layers for units, and debugging means getting to the bottom of more connection and boiler plate code.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;What is old but gold?&lt;/strong&gt; plain old forms and a sprinkle of javascript to make it pop. You can still get the benefit of using ajax to make dynamic sites. htmx is the tool of choice for wiring it together, without needing to write a heavy javascript app.&lt;/p&gt;&#xA;&lt;p&gt;You load htmx in the &lt;code&gt;&amp;lt;head&amp;gt;&lt;/code&gt; and it brings about ~10kb of minified javascript to the page. From here, you can sprinkle the &lt;code&gt;hx-&lt;/code&gt; directives in to your markup to instruct the framework on how the browser should make AJAX requests to the backend. Your backend responds with plain old html templates that will be substituted in to the page . A casual smile builds on your face as you realize you&#39;ve delivered the boring part of your app without the hassle of wiring components, state and back-end API&#39;s together.&lt;/p&gt;&#xA;&lt;p&gt;When you write your application with a tool like htmx, you get to spend more time focusing on the cool parts. Javascript is still the right choice for making components that extend the dom to do new and excellent things.&lt;/p&gt;&#xA;&lt;p&gt;Writing javascript is the right choice if you are building maps, games, interactive visualizations or advanced editing experiences. You can mix htmx with code, and write nice isolated components for the purpose.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Controlling the experience&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;The future internet will run on a several generations old android, an overused battery with a healthy amount of vendor bloatware. This is a reality I have been living the last few months as my solid modern android phone was lost and replaced by an Acatel 1E. Sure it cost $40, but it comes with a web browser, and those of us with a $40 android should be able to use the web too.  Most web application suck on a phone like this, loading more than two mb of javascript probably means you are going to have a less than ideal experience.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;I&#39;m convinced, how do I get started!&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;The original reason why I got in to this was because my good friend Matt asked me to give a talk at Christchurch Javascript meetup. I explained to a room of 20 people why they shouldn&#39;t write so much code, and it actually was really well recieved. It turns out that a lot of people who have been writing code for the front end share the same pain points, and are interested in a new (but actually old) way.&lt;/p&gt;&#xA;&lt;p&gt;My example code is posted on github.&lt;/p&gt;&#xA;&lt;p&gt;I love the process of explaining this stuff to people and I would love to share some infrequent writing with you too! If this interests you, please put your internet points towards this, and I will attempt to bring new and interesting stuff to your feed. I&#39;m not internet famous, so feel free to reach out&lt;/p&gt;&#xA;&lt;p&gt;Get started with the documentation at https://htmx.org/.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>javascript</category><category>html</category><category>python</category><category>go</category><guid isPermaLink="true">https://www.ersin.nz/articles/htmx-is-old-but-gold-4i9i</guid></item><item><title>Fighting boilerplate with code generation</title><link>https://www.ersin.nz/articles/fighting-boilerplate-with-code-generation-b41</link><description>Greetings fellow gophers.  Go is optimized for readability, but it can be a bit repetitive for the wr...</description><pubDate>10 Apr 2021 22:45 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;Greetings fellow gophers.&lt;/p&gt;&#xA;&lt;p&gt;Go is optimized for readability, but it can be a bit repetitive for the writing case.  Adding features can require a lot of grunt work and boiler plate code.&lt;/p&gt;&#xA;&lt;p&gt;Is boiler plate a problem?&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Marshaling Responses&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;Marshaling responses from web services. If we are optimizing for readability, we should have strongly typed structs to marshal in to. This leads to  grunt work in defining the type struct that maps to the web service.  The code generation solution is using a tool like &lt;a href=&#34;https://mholt.github.io/json-to-go/&#34;&gt;https://mholt.github.io/json-to-go/&lt;/a&gt; .&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;Interacting with a SQL persistence layer&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;Getting your data in and out of a SQL database requires boilerplate.&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Generating SQL DDL. Creating tables, indexes, relationships.&lt;/li&gt;&#xA;&lt;li&gt;Row to struct scanning.&lt;/li&gt;&#xA;&lt;li&gt;CRUD methods.&lt;/li&gt;&#xA;&lt;li&gt;Complex Queries with type safe interactions.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;There are a tone of tools in this place depending on the type of API you want to use with your crud methods, a couple worth evaluating are:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;a href=&#34;https://github.com/xo/xo&#34;&gt;https://github.com/xo/xo&lt;/a&gt;&lt;/li&gt;&#xA;&lt;li&gt;&lt;a href=&#34;https://sqlc.dev/&#34;&gt;https://sqlc.dev/&lt;/a&gt;&lt;/li&gt;&#xA;&lt;li&gt;&lt;a href=&#34;https://github.com/drone/sqlgen&#34;&gt;https://github.com/drone/sqlgen&lt;/a&gt;&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;&lt;strong&gt;Building out an HTTP API&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;Taking your current data structures and operations and exposing them over an HTTP API can be rife with boilerplate.&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Validating input&lt;/li&gt;&#xA;&lt;li&gt;Generating consistent route structure&lt;/li&gt;&#xA;&lt;li&gt;Marshaling and Unmarshaling data structures&lt;/li&gt;&#xA;&lt;li&gt;Generating consistent query parameter interface (ordering, filtering, pagination)&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;Projects that help  solve this:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;a href=&#34;https://goswagger.io/&#34;&gt;https://goswagger.io/&lt;/a&gt; (makes swagger compatible API&#39;s from definitions or code)&lt;/li&gt;&#xA;&lt;li&gt;&lt;a href=&#34;https://goa.design/&#34;&gt;https://goa.design/&lt;/a&gt; (makes a &#39;hexagonal&#39; architecture micro-service)&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;&lt;strong&gt;Building clients for your service&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;Once you have exposed your web service you almost certainly want to consume it. Possibly you want to consume it over a number of different languages and environments.  Thankfully, we can stand on the shoulders of giants here and use the open API IDL (&lt;a href=&#34;https://en.wikipedia.org/wiki/Interface_description_language&#34;&gt;Interface description langauge&lt;/a&gt;).  The IDL will let you generate client libraries for numerous programming languages, and tools.&lt;/p&gt;&#xA;&lt;p&gt;Check out everything you can codegen from the swagger IDL here &lt;a href=&#34;https://swagger.io/tools/swagger-codegen/&#34;&gt;https://swagger.io/tools/swagger-codegen/&lt;/a&gt;&lt;/p&gt;&#xA;&lt;p&gt;​&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;T L D R&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;Go code is a pleasure to read, but can be a bit of a grind to write all that boiler plate. You can remove some of the grunt work through code generation.  The design of Go has made it an easy tool for code generation. A mature ecosystem of tools makes it easy to get started.&lt;/p&gt;&#xA;&lt;p&gt;Are you using code generation in your projects / day jobs and hacks?&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>go</category><category>languages</category><category>beginners</category><guid isPermaLink="true">https://www.ersin.nz/articles/fighting-boilerplate-with-code-generation-b41</guid></item><item><title>How to ask for help</title><link>https://www.ersin.nz/articles/how-to-ask-for-help-3noc</link><description>Asking questions well is one of the most important skills to develop as a new professional. Really gr...</description><pubDate>04 Nov 2020 21:07 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;Asking questions well is one of the most important skills to develop as a new professional. Really great communication will help you get mentors, and increase your productivity.&lt;/p&gt;&#xA;&lt;p&gt;Here&#39;s the 3 things you should be doing when asking a question&lt;/p&gt;&#xA;&lt;p&gt;:bow: Start with a soft opener, and don&#39;t be a robot troll.&lt;/p&gt;&#xA;&lt;p&gt;:exclamation: Explain the situation and the complication.&lt;/p&gt;&#xA;&lt;p&gt;:book: List what you have tried to resolve the&lt;/p&gt;&#xA;&lt;p&gt;Working on a team, or individually there will always be occasions when you need to seek help from someone.  You could be asking about why a decision is made, or how to get a component to work, or maybe it seems completely broken!&lt;/p&gt;&#xA;&lt;p&gt;If you&#39;ve been around software for any number of days you will have had experience with some really bad answers. Perhaps it seems like certain people are guarding information, or maybe they just seem like a bit of a troll.&lt;/p&gt;&#xA;&lt;p&gt;This raises the question. How do you get a better answers to your questions?&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;:bow: Start with a soft opener&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;Establish that your question is being asked by a real person! Introduce yourself if you are a stranger. Throw in a complement about the tool/technology in question. Ask them about their pet/hobby/weekend.&lt;/p&gt;&#xA;&lt;p&gt;Most importantly, never :poop: on any code, you should never assume that code someone else is wrong, even when you&#39;re at a 80% confidence that it is. Situations where you need to explain that code is wrong, you should still phrase it as if it is right, but explain why it looks wrong. You will be surprised how often a 80% certainty that something is wrong will turn out to be right.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;:exclamation: Explain the situation and the complication&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;Explain your high level goal and then summarize what your approach has been. Before someone can answer your question, they need the context to understand what you are doing. When I am asking a question, I like to use it as an opportunity to get some more related knowledge. As a junior engineer, I found that in many situations I would gain a lot of valuable knowledge beyond the scope of my question when providing the right background context to the question.&lt;/p&gt;&#xA;&lt;p&gt;At the end of this, you need to get to your question, make it as short and bold as possible. You are able to do this now due to the explanation and friendliness in your communication.&lt;/p&gt;&#xA;&lt;p&gt;&lt;strong&gt;:book: List what you have tried&lt;/strong&gt;&lt;/p&gt;&#xA;&lt;p&gt;You should always list out what you have tried to solve your problem. Again, this show&#39;s respect to the answerer by letting them avoid many iterations to determine how you should debug the situation.&lt;/p&gt;&#xA;&lt;p&gt;There is another meta benefit to this.  The process of listing out what you have tried will help you brain storm different ways to self resolve. In the process of forming your question you may realize that the question does not need to be answered because you come up with a way to solve your own problem.&lt;/p&gt;&#xA;&lt;p&gt;Finally, the person answering can help you get better by pointing out other avenues of investigation for solving this type of problem. This is kind of gold quality you can get from a great answer that really rises your skill as a developer.&lt;/p&gt;&#xA;&lt;h2&gt;Examples&lt;/h2&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;&lt;strong&gt;Bad&lt;/strong&gt; uat.foo.acme.org api is down.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;&lt;strong&gt;Good&lt;/strong&gt; Hi Jane :wave:&#xA;I&#39;m working on some new features for facebook99 and for some reason all my requests are being thrown from foo.acme.org.&#xA;I have tried connecting through CURL, and I&#39;ve tried our UAT site and the integration seems to be down, and my development instance isn&#39;t responding either!&#xA;Can you help me out check the integration is working?&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;&lt;strong&gt;Bad&lt;/strong&gt; How do I get this shitty code generation to work? IT ALWAYS BREAKS :angry:&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;&lt;strong&gt;Good&lt;/strong&gt; Hi Team,  I&#39;ve been working on an urgent feature for $XCompany; and this has been tested and validated on my developer instance by the customer. Now I need to merge this feature and I&#39;m getting this really strange error 79 code from the CI system. $XCompany expects the release to be today so I&#39;m urgently needing some help here :grimacing:&#xA;&lt;strong&gt;Can someone please help me resolve this and give me a little bit of history on the codegen module?&lt;/strong&gt;&#xA;I&#39;ve tried running codegen locally and it works.&#xA;I&#39;m using the standard developer machine for testing&#xA;I&#39;ve updated the modules on the project so they &lt;em&gt;should&lt;/em&gt; be up to date in CI.&#xA;The feature has been revieiwed {here is the merge request}.&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;The age old developer aphorism of &#39;garbage in&#39; &#39;garbage out&#39; applies here. If you are getting bad answers, you may be asking bad questions.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>beginners</category><category>productivity</category><guid isPermaLink="true">https://www.ersin.nz/articles/how-to-ask-for-help-3noc</guid></item><item><title>Build infinite scroll in Svelte</title><link>https://www.ersin.nz/articles/build-infinite-scroll-in-svelte-4clk</link><description>Svelte is an interesting framework because it is not really a framework. To be accurate, it is a comp...</description><pubDate>19 Oct 2020 22:55 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;Svelte is an interesting framework because it is &lt;strong&gt;not&lt;/strong&gt; really a framework. To be accurate, it is a compiler which turns your components into incredibly lean javascript that has no external dependencies on framework code. The component structure is appealing. You ship your CSS/HTML/Scripts in a single &#39;svelte&#39; component file, where everything is bound and scoped locally.&lt;/p&gt;&#xA;&lt;h2&gt;The Problem&lt;/h2&gt;&#xA;&lt;p&gt;In my latest project, I want to display an infinite list of documents, each document could be up to a couple kb of markdown long, and there is an arbitrary number of documents. I want to focus on learning the framework, not a library or pre-made component, so I decided to implement the ability to quickly scroll this list on my own.&lt;/p&gt;&#xA;&lt;p&gt;The underlying data-source is simple. Two API endpoints. But you could easily switch this out with your own API which involves a list too long to render all upfront.&lt;/p&gt;&#xA;&lt;p&gt;When building a javascript heavy application, it&#39;s important to take advantage of what the platform affords us, and lazy loading an inifinite list is a very common technique used all across the web.&lt;/p&gt;&#xA;&lt;h2&gt;The Design&lt;/h2&gt;&#xA;&lt;p&gt;&lt;em&gt;Endpoints&lt;/em&gt;:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;&lt;code&gt;/list&lt;/code&gt; list all the available keys.&lt;/li&gt;&#xA;&lt;li&gt;&lt;code&gt;/r/:key&lt;/code&gt; get back the document for the underlying key.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;&lt;em&gt;The Psuedo code&lt;/em&gt;&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code&gt;Get the list of keys&#xA;Take the first 10 items into an array&#xA;render the first 10 items&#xA;hook up an event that triggers when the user has scrolled to the end of the page&#xA;render the next 10 items&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;h2&gt;The Implementation&lt;/h2&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-html&#34;&gt;&amp;lt;script&amp;gt;&#xA;// psuedo code -- but you get the idea right?!&#xA;const PAGE_LENGTH = 10&#xA;let docs = []&#xA;let loadingNextPage = false&#xA;let currentOffset = 0&#xA;&#xA;start(0)&#xA;&#xA;function start(offset) {&#xA;   getList() // BYO implementation of returning the list of data&#xA;     .then(keys =&amp;gt; docs = keys.slice(0, offset + PAGE_LENGTH))&#xA;}&#xA;function scrollHandler() {&#xA;  // only allow the loading of new data if it is not already loading new data&#xA;  if (!loadingNextPage &amp;amp;&amp;amp; (window.innerHeight + window.scrollY) &amp;gt;= document.body.offsetHeight) {&#xA;    const previousScrollY = window.scrollY;&#xA;    loadingNextPage = true&#xA;    setTimeout(() =&amp;gt; {&#xA;      window.scroll(0, previousScrollY)&#xA;      currentOffset += PAGE_LENGTH&#xA;      start(currentOffset)&#xA;      loadingNextPage = false&#xA;    })&#xA;  }&#xA;}&#xA;&amp;lt;/script&amp;gt;&#xA;// hooks up the data to the loadingNextPage &amp;amp;&amp;amp; (window.innerHeight + window.scrollY) &amp;gt;= document.body.offsetHeightnew &#xA;&amp;lt;svelte:window on:scroll={scrollHandler} /&amp;gt;&#xA;&amp;lt;div&amp;gt;&#xA;  {#each docs as doc}&#xA;    &amp;lt;Document name={doc}/&amp;gt;&#xA;  {/each}&#xA;&amp;lt;/div&amp;gt;&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;The key things we are depending on to make this all work are the &lt;code&gt;&amp;lt;svelte:window on:scroll={scrollHandler}&amp;gt;&lt;/code&gt; and the framework keeping track of the local binding of the &lt;code&gt;docs&lt;/code&gt; array.&lt;/p&gt;&#xA;&lt;p&gt;The on scroll event is exactly the same as the &lt;a href=&#34;https://developer.mozilla.org/en-US/docs/Web/API/Document/scroll_event&#34;&gt;scroll&lt;/a&gt; dom event.&lt;/p&gt;&#xA;&lt;p&gt;The main gotcha with this is that you will generate many many scroll events and you only want to start loading new data when you reach near the end of the window. This is solved by the condition inside our scroll event handler &lt;code&gt;loadingNextPage &amp;amp;&amp;amp; (window.innerHeight + window.scrollY) &amp;gt;= document.body.offsetHeight&lt;/code&gt;&lt;/p&gt;&#xA;&lt;h1&gt;The Final note&lt;/h1&gt;&#xA;&lt;p&gt;Thanks for making it to the end of this tiny little tutorial. I am very open to feedback, on the format or content here.&lt;/p&gt;&#xA;&lt;p&gt;Has this helped you understand Svelte, or the web?&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>javascript</category><category>svelte</category><category>frontend</category><category>beginners</category><guid isPermaLink="true">https://www.ersin.nz/articles/build-infinite-scroll-in-svelte-4clk</guid></item><item><title>You don&#39;t always need to love the code</title><link>https://www.ersin.nz/articles/you-don-t-always-need-to-love-the-code-4nf3</link><description>In a community programming forum I heard a new developer say:   I quit programming, here to chat. It...</description><pubDate>16 Sep 2020 15:10 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;In a community programming forum I heard a new developer say:&lt;/p&gt;&#xA;&lt;blockquote&gt;&#xA;&lt;p&gt;I quit programming, here to chat. It was fun when I did development as a kid, but since then it&#39;s no longer enjoyable&lt;/p&gt;&#xA;&lt;/blockquote&gt;&#xA;&lt;p&gt;I&#39;m here to say; That&#39;s OK. You are still welcome to hang if you don&#39;t program. I&#39;ve &#39;quit&#39; coding a few times myself. Came back and found even more joy in the industry, the people and sometimes the code. You don&#39;t need to code for fun, or have fun coding in order to get through the gates to our community.&lt;/p&gt;&#xA;&lt;p&gt;When I haven&#39;t loved writing code, I&#39;ve found a lot of joy in other places. To list a few:&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Solving a problem for someone. For people outside the technical space, what we do is magic, working hard on solving their exact problem with the most boring code is my top joy in development.&lt;/li&gt;&#xA;&lt;li&gt;Learn something new. Get creative and pick up a new skill. I find the initial learning curve to be one of the most thrilling things that keeps me in love with my career.&lt;/li&gt;&#xA;&lt;li&gt;Understanding the magic. Getting to the bottom of a legacy code base, reading the source behind your favorite library or framework.&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;What we do can be done as a profession or a Hobby. But for some it&#39;s not healthy to do both. You can still be a part of the community.&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>career</category><category>advice</category><guid isPermaLink="true">https://www.ersin.nz/articles/you-don-t-always-need-to-love-the-code-4nf3</guid></item><item><title>Get better at writing with 3 lines of bash</title><link>https://www.ersin.nz/articles/get-better-at-writing-with-3-lines-of-bash-2cgj</link><description>Everyone knows that great communication skills are the most important trait we want in our team membe...</description><pubDate>28 Mar 2020 00:36 GMT</pubDate><content>&lt;![CDATA[&lt;div class=&#34;markdown-content&#34;&gt;&lt;p&gt;Everyone knows that great communication skills are the most important trait we want in our team members. I have been deliberately practicing writing over the last three years and today I want to share with you my most impactful tools. Here are three lines of bash that are responsible.&lt;/p&gt;&#xA;&lt;pre&gt;&lt;code class=&#34;language-bash&#34;&gt;#!/bin/bash&#xA;&#xA;NOTESPATH=&amp;quot;~/Notes/&amp;quot;&#xA;&#xA;$EDITOR &amp;quot;$NOTESPATH$(date &amp;quot;+%Y-%-m-%-d&amp;quot;)&amp;quot;&#xA;&lt;/code&gt;&lt;/pre&gt;&#xA;&lt;p&gt;How do I install it?&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;put it on your path, I suggest &lt;code&gt;/usr/bin&lt;/code&gt;&lt;/li&gt;&#xA;&lt;li&gt;create the Notes directory in your user home.&lt;/li&gt;&#xA;&lt;li&gt;set your $EDITOR variable to your favorite tool ( if it isn&#39;t already ).&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;What does it do?&lt;/p&gt;&#xA;&lt;ul&gt;&#xA;&lt;li&gt;Creates a file with the current date if it doesn&#39;t exist&lt;/li&gt;&#xA;&lt;li&gt;Opens your today&#39;s file if it already exists&lt;/li&gt;&#xA;&lt;/ul&gt;&#xA;&lt;p&gt;After many iterations on writing tools, this is still my favorite way to get my thoughts on &#39;paper&#39;. Day to day, I spend my time using the terminal, and for me this script is the easiest way to context switch in to writing.&lt;/p&gt;&#xA;&lt;p&gt;The hardest part about deliberate practice is starting, and this tool has allowed me to write something much more frequently.&lt;/p&gt;&#xA;&lt;p&gt;Hope you find this useful! &amp;lt;3&lt;/p&gt;&#xA;&lt;/div&gt;]]&gt;</content><category>showdev</category><category>productivity</category><category>linux</category><guid isPermaLink="true">https://www.ersin.nz/articles/get-better-at-writing-with-3-lines-of-bash-2cgj</guid></item></channel></rss>