Practical articles on AI, DevOps, Cloud, Linux, and infrastructure engineering.
Our best engineer quit citing on-call. We rebuilt the whole thing: saner rotations, runbooks that actually help at 3am, and escalation that doesn't punish asking for help.
Moving our fleet from x86 to Graviton promised 20% savings. We got 31%, but only after fixing native dependencies, a broken base image, and one nasty perf regression.
We used to ship code and turn it on in the same breath, so every deploy was a bet. Feature flags split those two events apart and made rollbacks a config toggle.
A p99 that jumped to 3.4 seconds during traffic ramps turned out to be cold starts. Here's how we measured them properly and cut the tail, with real init timings.
When our single LLM provider had a 40-minute outage, every AI feature went dark. A gateway with routing and fallback fixed that, and cut spend 30% as a bonus.
Reliability arguments used to be shouting matches between SRE and product. An error budget turned them into arithmetic. Here's how we made the number drive the roadmap.
A 180k-token context window is not a license to stuff everything in. Here's how we cut prompt size 60% without hurting answer quality, and what to trim first.
Most SLI dashboards track things nobody notices. Here's how we picked the handful of signals that map to real user pain, and dropped the vanity metrics.
The top model on the MTEB leaderboard made our search worse and our bill bigger. Here's how we actually picked an embedding model for a real RAG system.
Our failover config looked perfect in the console and did nothing during a real outage. Here's the health-check design that actually flipped regions when it mattered.
We had the same 180-line build workflow copy-pasted into 60 repos. Fixing one bug meant 60 PRs. Here's the reusable-workflow setup that made it one.
Users hit stop, but our server kept paying for tokens for another 40 seconds. Here's how we wired real cancellation and backpressure into an SSE streaming endpoint.
