Repository navigation
Conversation
moqperf_test_client reports latency as a run average plus each second's min/avg/max. That is not enough to see p95/p99 shifts. The existing Prometheus histogram doubles its bucket width, so it can't resolve p99, and it covers the whole run including ramp-up. The client now also records each object in a log-linear histogram: 1 ms buckets below 128 ms, then buckets at most 1/32 of their value wide up to 16 s. Each second it prints that second's non-empty buckets: [LATENCY_HIST] [42s] 7:12 8:40 ... 203:3 Each key is the largest latency the bucket holds. A harness can add up the seconds inside its own steady-state window and read any percentile. The final summary adds whole-run p50/p90/p95/p99/p99.9/max. The [AGGREGATE] line and the Prometheus export are unchanged.
|
No actionable comments were generated in the recent review. Walkthrough
Merge Risk:
|
The perf client only reports latency as a run average plus each second's min/avg/max. The moqx nightly perf run needs p95/p99 to see changes in tail latency (openmoq/moqx#886). The existing Prometheus histogram can't give p99: its buckets double in width, and it covers the whole run, including ramp-up.
LatencyQuantiles.h: a log-linear histogram with 1 ms buckets below 128 ms, then buckets at most 1/32 of their value wide up to 16 s, plus an overflow bucket.moqperf_test_clientrecords every object into it. Each second it prints that second's non-empty buckets as[LATENCY_HIST] [42s] 7:12 8:40 ... 203:3. Each key is the largest latency the bucket holds, so the moqx scripts can add up the seconds in their steady-state window without knowing the bucket layout.Object Latency p50/p90/p95/p99/p99.9/max.[AGGREGATE]line and the Prometheus export are unchanged.LatencyQuantilesTests: bucket boundaries, nearest-rank percentiles, overflow, merge, and the per-second reset.The moqx side (parsing, comparison, dashboard) is a separate PR. It works with or without this change.
This change is
Summary by CodeRabbit