[ https://issues.apache.org/jira/browse/IMPALA-5485?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16460323#comment-16460323 ]
Michael Ho commented on IMPALA-5485: ------------------------------------ Yes but that may cause smaller row batches to be returned from exchange node. We can revisit the tradeoff when this JIRA becomes higher priority. > Exchg / data-stream recv-side buffers should operate under a memory > reservation > ------------------------------------------------------------------------------- > > Key: IMPALA-5485 > URL: https://issues.apache.org/jira/browse/IMPALA-5485 > Project: IMPALA > Issue Type: Sub-task > Components: Distributed Exec > Reporter: Henry Robinson > Priority: Major > > Exchange nodes assign a fixed-size buffer to their datastream receivers > that's used to smooth out differences in send / consume rates between the > sender and the receiver. > These buffers should be tracked by the query memtracker, and with the new > min-reservation support we should allow them to be larger than the configured > minimum. Increasing the buffer size decreases the amount of time that a > sender can be blocked on a receiver, and so increases query-parallelism. > Queries that shuffle a lot of data can see significant speedups from larger > buffers. > The buffers need to be sized based on the #senders and the #rows * #avg row > size. They can dynamically expand trivially - contraction is possible, but a > bit harder. -- This message was sent by Atlassian JIRA (v7.6.3#76005) --------------------------------------------------------------------- To unsubscribe, e-mail: issues-all-unsubscr...@impala.apache.org For additional commands, e-mail: issues-all-h...@impala.apache.org