[ https://issues.apache.org/jira/browse/SPARK-36872?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Shekhar Gupta updated SPARK-36872: ---------------------------------- Description: During the graceful decommissioning phase, executors need to transfer all of their shuffle and cache data to the peer executors. However, they get killed before transferring all the data because of the hardcoded timeout value of 60 secs in the decommissioning script. As a result of executors dying prematurely, the spark tasks on other executors fail which causes application failures, and it is hard to debug those failures. To fix the issue, we ended up writing a custom script with a different timeout and rebuilt the spark image but we would prefer an easier solution that does not require rebuilding the image. (was: During the graceful decommissioning phase, executors need to transfer all of their shuffle and cache data to the peer executors. However, they get killed before could transfer all the data because of the hardcoded timeout value of 60 secs in the decommissioning script. As a result of executors dying prematurely, the spark tasks on other executors fail which causes application failures, and it is hard to debug those failures. To fix the issue, we ended up writing a custom script with a different timeout and rebuilt the spark image but we would prefer an easier solution that does not require rebuilding the image. ) > Decommissioning executors get killed before transferring their data because > of the hardcoded timeout of 60 secs > --------------------------------------------------------------------------------------------------------------- > > Key: SPARK-36872 > URL: https://issues.apache.org/jira/browse/SPARK-36872 > Project: Spark > Issue Type: Improvement > Components: Kubernetes > Affects Versions: 3.1.1, 3.1.2, 3.2.0 > Reporter: Shekhar Gupta > Priority: Trivial > > During the graceful decommissioning phase, executors need to transfer all of > their shuffle and cache data to the peer executors. However, they get killed > before transferring all the data because of the hardcoded timeout value of 60 > secs in the decommissioning script. As a result of executors dying > prematurely, the spark tasks on other executors fail which causes application > failures, and it is hard to debug those failures. To fix the issue, we ended > up writing a custom script with a different timeout and rebuilt the spark > image but we would prefer an easier solution that does not require rebuilding > the image. -- This message was sent by Atlassian Jira (v8.3.4#803005) --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org