Rangsh commented on issue #12058: URL: https://github.com/apache/seatunnel/issues/12058#issuecomment-5659578915
@SEZ9 confirming the two remaining asks for closing out the flush-path experiment on this issue. ### 1. CPU / GC / after-side hsync (record completeness) Yes — the full numbers are in `AB-REPORT.md` in the A/B gist: https://gist.github.com/Rangsh/66008c5bad44fb05825db40198349175 Pasting them here as well so the issue thread is complete even if the prior comment rendering was truncated: **CPU** (`profile_benchmarks.sh profile cpu`, diagnostic `-f 1`) | Side | samples | `HdfsWriter.flush` / `hsync` share (approx) | | --- | ---: | ---: | | BEFORE | 209 | ~0.48% | | AFTER | 189 | **~0.53%** | **GC** (`profile_benchmarks.sh profile gc`, diagnostic `-f 1`) | Side | Alloc/op | GC count | GC time | | --- | ---: | ---: | ---: | | BEFORE | 286,615 B/op | 1 | 5 ms | | AFTER | 286,729 B/op | 1 | 4 ms | No meaningful GC delta. Sync remains not CPU-hot on either side. ### 2. Flush-collapse experiment status for #12058 I consider the flush-collapse experiment **concluded for #12058** on the agreed `file:///` / LocalFileSystem contract. - The isolated A/B was a **null result** (Score/CV did not improve; park / `InvocationFuture.get` hotspot remained). - That null result is **filesystem-specific** to this harness. I am **not** generalising it to a real HDFS/DFS client (where multi-sync + trailing `hflush` could still matter for cost under a separate contract). - I do **not** intend to rerun this same single-variable flush-collapse A/B against a non-local filesystem as part of closing #12058. - **#12081 stays Related only** and still cannot claim to address #12058 on the basis of this local null result. Please treat the storage sync / flush-collapse path as **ruled out** as the driver of the variance reported on this issue. Happy to keep the investigation focused next on the `InvocationFuture.get` / `LockSupport.park` wait. -- 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]
