Latency budgets, not averages
Why p99 is the only number that survives contact with production, and how to keep a request path honest when a dependency slows down.
Short write-ups on networking, storage and the boring parts of running software. Updated when there is something worth writing down.
Why p99 is the only number that survives contact with production, and how to keep a request path honest when a dependency slows down.
Queues that grow without bound only move the failure later. A short note on explicit limits and what to do when they are reached.
A practical checklist for tracing one slow connection end to end, starting from the client and ending at the accept queue.
Splitting a monolith buys independent failure domains only when the boundaries carry their own data ownership.