DanielLeens commented on issue #12058: URL: https://github.com/apache/seatunnel/issues/12058#issuecomment-5663652099
Thanks for making the record complete. The new table supports a narrow and useful conclusion for the existing `file:///` harness: the isolated flush-collapse A/B is a null result, `HdfsWriter.flush` / `hsync` CPU share is effectively unchanged (about 0.48% before vs. 0.53% after), and allocation/GC are likewise unchanged (286,615 vs. 286,729 B/op; one collection on each side). That rules out this particular flush-collapse change as an explanation or optimization for the observed local benchmark variance. It does not establish that a real HDFS/DFS client has the same cost profile, so please do not generalize the local result into a filesystem-wide claim. The remaining `InvocationFuture.get` / `LockSupport.park` hotspot is a wait site, not yet a proven cause. Please keep #12058 open and keep the next experiment on unchanged `dev`, with one controlled question about the wait: capture the wake-up/notification and scheduler timing needed to distinguish a causal wait condition from a passive symptom, together with raw per-iteration data and the exact JDK, machine, JMH, and storage settings. No production change should be proposed from the current evidence. #12081 remains Related-only and cannot close this issue; at this read its Build check is failing, and its correctness work must be evaluated independently of this benchmark conclusion. -- 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]
