jerqi commented on PR #1514: URL: https://github.com/apache/incubator-uniffle/pull/1514#issuecomment-1942991425
> @jerqi you are right, if both tasks report their shuffle results (block ids), we would not know how many blocks to read for our partition id and map index. For that to work we need to guarantee that only one task attempt succeeds in registering shuffle results (blockids). We then take that taskAttemptId and those blockIds for the partition id and map index. > > This would require a two-phase commit to guard / coordinate the report shuffle result step, which would be a much bigger change: https://github.com/G-Research/incubator-uniffle/pull/4/files > > There are only two situations, where this can happen, though: speculative task execution (which is not recommended for production) and (without speculative task execution) a task failure after shuffle results have been reported, which triggers another task attempt. Spark has already had a commit mechanism. If we set up a new two-phrase commit. Maybe there will be some consistent. -- 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] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
