[ https://issues.apache.org/jira/browse/SPARK-24016?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16445252#comment-16445252 ]
Saisai Shao commented on SPARK-24016: ------------------------------------- I think this can be useful if we enabled "spark.blacklist.killBlacklistedExecutors". NM could avoid relaunching executors on the bad nodes. > Yarn does not update node blacklist in static allocation > -------------------------------------------------------- > > Key: SPARK-24016 > URL: https://issues.apache.org/jira/browse/SPARK-24016 > Project: Spark > Issue Type: Improvement > Components: Scheduler, YARN > Affects Versions: 2.3.0 > Reporter: Imran Rashid > Priority: Major > > Task-based blacklisting keeps track of bad nodes, and updates YARN with that > set of nodes so that Spark will not receive more containers on that node. > However, that only happens with dynamic allocation. Though its far more > important with dynamic allocation, even with static allocation this matters; > if executors die, or if the cluster was too busy at the original resource > request to give all the containers, the spark application will add new > containers in the middle. And we want an updated node blacklist for that. -- This message was sent by Atlassian JIRA (v7.6.3#76005) --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org