turboFei edited a comment on issue #24740: [SPARK-27876][CORE] Split large shuffle partition to multi-segments to enable transfer oversize shuffle partition block URL: https://github.com/apache/spark/pull/24740#issuecomment-497559772 > You shouldn't have this limit anymore (from spark 2.4 onwards) as long as you're also running a recent shuffle service. this uses fetching shuffle blocks to disk, instead of memory, which should be enabled by default for large blocks (https://issues.apache.org/jira/browse/SPARK-24297). > > But if you're seeing a failure with that, can you share some more details? Thanks. I see this failure with spark-2.3.2. I'm sorry for that I did not notice your PR, which set maxRemoteBlockSizeFetchToMem to a value little than 2GB. But when resources are available, it is a good idea that we can set this value large than 2GB to reduce the I/O overhead. So, shall we support the shuffle data transmission no matter the shuffle blocks' size when maxRemoteBlockSizeFetchToMem is large than 2GB?
---------------------------------------------------------------- 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]
