Luis Lavena

Last year I released the container image I use to compile @CrystalLanguage apps. You can use it to build and test against different version, 1.2, 1.3 or latest 1.5.1 This started a series of experiments on things I'm working on since then... 1/ 🧵

Uncloud just got its first university adoption! 🎓 Radboud's Faculty of Science is rolling it out to manage Docker clusters for their research groups. They're going for: → 300+ websites across a dozen machines → 100+ people who need to self-serve → each research group gets a VM or cluster they fully control → self-host the entire web stack including DB + automated S3 backups → GitLab environments for deploy-on-green workflows → platform team out of the loop >"It looks very much like docker compose or the (semi-deprecated) docker swarm, and there is good documentation" And a huge thanks to @miekg for digging in, reporting issues and contributing fixes over the past few weeks 🙏 Can't wait to see it running in production!

Just shipped Uncloud v0.18 🎉 → Deploy plan got a glow-up - much nicer to read and understand 👇 → Pre-deploy hooks - run DB migrations before new code goes live → System 'ssh' is now the default - faster commands and full SSH config support → Pin compose.yaml to cluster with 'x-context' 📚 New docs!

🎙️ I went on go podcast() this week to talk about why I’m building Uncloud - Docker Compose for production, bridging the gap between Docker and Kubernetes. We covered: - why most projects don't need k8s and what alternatives we have - a hybrid setup mixing cloud VMs and on-prem - WireGuard mesh networking - unregistry - direct image push to servers without a registry Give it a listen on any podcast platform and let me know what you think! 🎧👇

In the next version of Uncloud: detailed deployment plan with style ✨ The plan shows exactly how your services will be rolled out: - what changes - which containers get replaced - where and in what order - whether there's a downtime risk Review, then deploy. Before → After 👇

@aarondfrancis For me worktrees advantage is that I don't need to cd and git pull the other copy multiple times to get the updates, I just switch and rebase them from the same remote update. Is not about space efficiency but performance and memory (did I remember to pull on this clone?)

Why do people like git worktrees over discrete checkouts? (This isn't bait, it's research)

Just released Uncloud 0.17 🚀 Think multi-machine Docker Compose for production Rolling deployments now monitor crashes + health checks after new containers start. If a container keeps restarting or turns unhealthy, the deployment rolls it back to prevent downtime. Enjoy safer deploys by default.

TIL that K8s rolling deployments are not zero-downtime by default. To achieve zero downtime during rollout (stateless workloads), old containers should typically be terminated the following way: 1. Remove the container endpoint from the ingress/load balancer to stop sending new requests to it but continue processing the active ones. 2. Gracefully stop and remove the container. On stop, the process in the container should stop accepting new requests and finish serving all active ones, then terminate. The updated ingress will guarantee that new requests won't be sent to the stopping container. The strict order is essential. Alternatively, if updating the ingress configuration is not practical, the ingress should be able to automatically detect unhealthy endpoints and retry the failing requests with healthy endpoints (aka passive health checks). The problem is that K8s doesn't coordinate pod termination with the ingress, nor do popular ingress controllers implement retries for failed requests by default. It simply terminates a pod which triggers the ingress update and process shutdown simultaneously. If the process happens to start terminating before the ingress removes the endpoint, new requests will get 5xx - downtime. The workaround is to add a wonky sleep(10s) in the app’s shutdown handler or as a preStop hook in the pod. This ensures the changes are propagated to the ingress before actually stopping the process in the container. It's a shame that the behaviour the majority would expect is not the default. Sensible defaults matter. UX first please

Using @aarondfrancis counselors to build it (thank you!), then the judge evaluates the results (baseline vs improved version) and comes back with a verdict on success/failure, but still playing with it.

Has anyone worked on a testing/validation framework for AI skills? Been setting one up to validate improvements on my own skills but that might be biased, so would love a second opinion😅

@simonw Nice! I've been tasking agents to use agent-browser to validate features and then write down playwright scripts to avoid regressions: github.com/luislavena/ski…

I built two new tools to help coding agents demonstrate their work beyond just running automated tests: Showboat and Rodney simonwillison.net/2026/Feb/10/sh…

@nateberkopec One-shot style plus+ ralphing: github.com/luislavena/jus… Learn that there are better approaches for API design and results than brute force and thousands tokens.

You’re about to see a lot more “port of X to Y” software. It hits the sweet spot of what agents are good at: translation, and looping against a huge, known spec (the previous implementation)

@nateberkopec @pangramlabs There are some settings like temp and others that could impact the results (also the length of it). Maybe a ping to @N8Programs? 😊

@luislavena @pangramlabs Not even close to good enough. You can read it for yourself, still tons of obvious AI signals and pangram still hits it as 100% ai detected. gist.github.com/nateberkopec/0…

Interesting observation: none of the "humanizer" skills I've tried (and even the ones I've coded myself) have been able to "defeat" @pangramlabs in humanizing a 100% ai-written original text. However, they do successfully "humanize" text which was originally dictated by voice.

Ruby @rubylangorg (RubyInstaller) made to Microsoft Store! That is a lot of work! Congratulations to the team! 🎉🥳👏 rubyinstaller.org/2026/01/27/rub…

... hours later, after tackling bunch of other issues ... Claude: Multiline plain scalars are complex. Let me look at a simpler validation error first. 🙈🤦

Me: Here are some errors, fix them. Claude: This test case involves multiline plain scalars, which is a complex issue. Let me look at other failing tests to find easier wins first. 🤖🐓

How did you verify your code worked before AI? You still need to do that.

Finally sent my first newsletter: A year of building Uncloud (a simpler Kubernetes alternative for small teams and solo developers) What happened, where we are now, and what’s coming in 2026 👇

@nateberkopec So it's the agent connect to slack via MCP and has access to the conversation? Deploying one agent per client? Or having an orchestrator that will dispatch them when needed?

gist.github.com/nateberkopec/d… We're actively using this to serve Japanese clients now (yes, I live in Japan, no, my Japanese is not nearly good enough for full on ビジネス敬語

My Ruby on Rails performance retainer service is now available in any (human) language, anywhere.

Shipped this race condition fix in Uncloud v0.16 🚀 Also: - new 'uc wg show' command for troubleshooting the WireGuard connectivity issues and mesh topology - cap_add, cap_drop, sysctls support in Compose files