A micro-batching window of 3 ms removed 80% of our round trips. Then it made the p99 worse, and the fix was not a bigger window.
2026-09-13 · 9 min readMostly storage engines, queues and the unglamorous parts of running services. Written for my future self, who will absolutely have forgotten all of this.
Last updated 2026-09-22
A micro-batching window of 3 ms removed 80% of our round trips. Then it made the p99 worse, and the fix was not a bigger window.
2026-09-13 · 9 min readWhat I would keep (visibility timeouts, at-least-once) and what I would delete (priorities, and the clever scheduler).
2026-07-22 · 12 min readCounting hits over requests hides the case you care about. Count them per key population instead, and the graph starts telling the truth.
6 min readEvery retry policy is a capacity decision. Jitter is not optional, and a retry budget beats a retry count.
7 min read