SEZ9 commented on issue #12058: URL: https://github.com/apache/seatunnel/issues/12058#issuecomment-5611678243
@Rangsh thanks for the follow-up — this answers the diagnostics ask from @nzw921rx as far as I can tell. The gist (`e21302908815c2954edf4392eafc62c8`) has the pieces we asked for: `run-a-results.json` / `run-b-results.json`, the full `-v EXTRA` logs for all forks, the per-fork `-Xlog:gc*=info` logs, and the `profile-gc-summary.md` / `profile-gc-result.jmh.json` output from `tools/benchmarks/profile_benchmarks.sh profile gc`. Losing the Sep 8 raw files is fine; the fresh A/A on the same host/JDK/settings at `origin/dev` @ `8bea8c681` is what we needed, and keeping #12081 as Related-only is the right call. Your A/A takeaway matches how I read the tables: ~32% CV with 393 us spikes one day, ~10–11% CV with a max of ~190–206 us the next, on identical inputs. That means a single local Score ± Error cannot be used to claim or disprove a production mechanism, so no fix is validated yet — agreed. What is still open from the diagnostics point: 1. @nzw921rx asked for a screenshot of the analysis view or the original HTML flame graph from the tool run. The gist has the GC summary and JMH JSON, but I don't see the flame graph/HTML or a screenshot listed in the file table. Please add those (or say if the `profile gc` mode doesn't produce them and which mode does). 2. Please confirm whether the GC logs from run A/B show anything correlating with the 205.64 us iteration (fork 1) and the 115.29 us iteration (run B, fork 0) — e.g. a pause or heap resize aligned with those iterations — or whether the GC timeline is flat across all forks. That is the first concrete mechanism check we can do on the unchanged code. 3. As @nzw921rx said, keep the root-cause search on the unchanged `origin/dev` code and the existing artifacts before introducing any code change; if the GC logs are flat, the next narrow experiment would be on the same commit with JIT/scheduling isolated, not an A/B against a candidate fix. Once the flame graph/screenshot is up and we know whether GC correlates with the outliers, we can decide the next single-variable experiment. <!-- streview-comment:925 --> -- 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]
