gczsjdy commented on issue #24462: [SPARK-26268][CORE] Do not resubmit tasks when executors are lost URL: https://github.com/apache/spark/pull/24462#issuecomment-519802295 @squito Yeah it saves us much, from a TPC-DS 1T benchmark, 30% queries get 1.1x+ performance boost, 13% get 1.2x + performance boost. There's still remote read, but only once(if index files are not swapped out because of insufficient cache space), and this feature can take advantage of internal network bandwith inside computing cluster, releasing the compute-storage network, which may be the bottleneck of the workload. By 'reasonable balance', did you mean not considering complex conditions? I think probably it's beneficial to make it clear through discussion. Making this work in a long term is also fine by me. I tried to make a point that the current solution to not resubmitting map tasks by modifying `MapStatus` is not enough, due to it only cares about what Executors tell the Driver about the map outputs'(tasks') location. However, we should also grant Driver the right(for example, by add a config like this PR did) to not resubmit the map tasks even if it knows (not empty MapStatus location) which one to resubmit.
---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
