Github user tgravescs commented on the issue: https://github.com/apache/spark/pull/18388 200k+ connections seems to be your problem then. Is this all a single application? You say 6000 nodes with 64 executors on each host, how many cores per executor? Or do you mean basically each host can run max 64 tasks in parallel. (6000*64) = 384000 which would be your 200K. I'd be surprised if every reducer is hitting the all nodes at the same time. We are randomizing the blocks to fetch in hope they don't hit all the same one at once. have you tried using spark.reducer.maxReqsInFlight? Rejecting connection could also slow things down or even worse make them fail. You have a wait between retries and if you hit the max retries and fail tasks that is much worse then flow control. Reconnection does have a cost but either way you are going to wait some between retries, you don't actually want to retry to quickly or you will just have same issue. What problem are you seeing with the close? I agree that I think both are good to have but personally think the reject connections should be the last thing you want to do.
--- If your project is set up for it, you can reply to this email and have your reply appear on GitHub as well. If your project does not have this feature enabled and wishes so, or if the feature is enabled but not working, please contact infrastructure at infrastruct...@apache.org or file a JIRA ticket with INFRA. --- --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org