SEZ9 commented on issue #12058: URL: https://github.com/apache/seatunnel/issues/12058#issuecomment-5673574709
@Rangsh — thanks for pasting the CPU/GC tables here alongside the `AB-REPORT.md` in the gist (`66008c5bad44fb05825db40198349175`); the record on this issue is now complete. Answering both asks directly: **1. CPU / GC / after-side hsync** — Done, nothing further needed. The numbers say what you concluded: `HdfsWriter.flush` / `hsync` share is effectively flat (~0.48% before vs. ~0.53% after) and allocation/GC is unchanged (286,615 vs. 286,729 B/op, one collection each side). Sync is not CPU-hot on either side of the A/B. **2. Flush-collapse experiment status** — Agreed, concluded for #12058 on the `file:///` / LocalFileSystem contract. The isolated A/B is a null result, so I'm treating the storage sync / flush-collapse path as ruled out as the driver of the variance reported here. I also agree with the limits you drew: this is not a statement about a real HDFS/DFS client, no rerun against a non-local filesystem is required to close out this line, and no production change should be proposed from this evidence. #12081 stays Related-only and cannot claim to address #12058 on this basis. What remains, as you already laid out in your last comment: - Keep #12058 open. - Next experiment on unchanged `dev`, one controlled question about the `InvocationFuture.get` / `LockSupport.park` wait — treat it as a wait site, not a proven cause, until the data says otherwise. - Please include in that post: wake-up/notification and scheduler timing sufficient to distinguish a causal wait condition from a passive symptom, raw per-iteration data, and the exact JDK, machine, JMH and storage settings. Nothing else is outstanding from the flush-path work. Looking forward to the wait-site results. <!-- streview-comment:1047 --> -- 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]
