gripleaf commented on PR #323: URL: https://github.com/apache/paimon-cpp/pull/323#issuecomment-5630709710
> > end-to-end query latency or a corresponding query speedup. > > The flame graph suggests that Avro decoding time drops significantly after the optimization. I’m also curious whether the end-to-end query latency shows a similarly noticeable improvement, since it’s not entirely clear what fraction of the overall query latency is spent in Avro decoding. Thanks! Thanks for raising this. In our workload, the impact on query latency also comes from contention in the executor used for manifest scanning. Paimon C++ submits a read task for each manifest being scanned and waits for these tasks to complete before proceeding. CPU-intensive Avro decoding keeps the executor’s worker threads occupied, causing additional scan tasks to queue up under concurrent requests. The large number of manifest files amplifies this problem: our previous count was over 7 million manifest files in total. Under sustained load, requests accumulate and end-to-end latency progressively increases. Reducing manifest decoding CPU should therefore help by freeing executor capacity and reducing queueing delays, in addition to making individual scans cheaper. The flame graphs demonstrate the CPU reduction, but we have not yet quantified the corresponding end-to-end latency improvement under the same load. -- 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]
