Rangsh commented on issue #12058: URL: https://github.com/apache/seatunnel/issues/12058#issuecomment-5681358366
@SEZ9 @DanielLeens @nzw921rx following up on the requested next experiment: **noisy primary + jmh-worker park-sum alignment**, plus diagnostic **`OperationsPerInvocation(1)`** one-to-one correlation. Production code unchanged; `#12081` stays Related-only. ### What we ran Same baseline as the wait-site pack (`38a37104e`, Corretto 11.0.26, Apple M1, `file:///`, method `checkpointOverviewIncrementalUpdate$`, 0 ns `ThreadPark` JFR). - **N.** Five OPI=100 primary attempts (3 sequential `-f 1` per attempt with explicit `fork-N.jfr` — Corretto 11 here does not expand `%p` inside `StartFlightRecording` filenames). Merged to the usual 15-sample shape. - **O.** Diagnostic OPI=1 (benchmark-fixture only; TearDown durable reload/validate skipped via a diagnostic system property because single-op MapStore reload races; timed `updateOverview` still runs). Restored afterward. ### Noisy hunt (OPI=100) Best CV day: **attempt-04 = 15.59%** (165–ish us/op class; min–max **138–255**). Historical ~32% / ~393 us day was **not** reproduced (other attempts ~9–11% CV). Park↔sample alignment on attempt-04 (jmh-worker × `AbstractInvocationFuture` × `MapProxySupport.invokeOperation`, proportional warmup/measurement windows; stacks truncate above `invokeOperation`, so absolute park-sum/op is inflated — use rank/window correlation): - Overall **Pearson 0.65 / Spearman 0.84** (n=15) - fork 0 almost monotone (Spearman 1.0) - fork 2 spike **254.99 us/op** co-locates with window **max park ≈ 950 µs** (fork-2 max-park↔sample Pearson ≈ 0.91) ### Diagnostic OPI=1 n=60, mean ~778 us/op, **CV 88.7%**, min–max 290–**5385** us (diagnostic Score — not comparable to OPI=100 primary). - Pearson ≈ **0.00**, Spearman ≈ **0.33** - Largest spike **5385 us** has park window only **~478 us** (residual ~4.9 ms) → extreme tails are **not** InvocationFuture-park explained ### Verdict (still no production change) - **H_mixed:** on the noisiest OPI=100 day we captured, sample movement **tracks** jmh-worker park windows (including the 255 us spike vs a long park tail). - OPI=1 shows **additional non-park spike sources**; park is not a universal spike cause. - Historical high-CV day still open (15.6% ≠ ~32%). - **Keep production code unchanged.** `#12081` cannot close this issue. Happy to push a gist of `NOISY-ALIGN-REPORT.md` + raw JSON / park-alignment summaries if useful; raw `.jfr` remains local (binary). -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
