Zhiyuan Yang created TEZ-3822: --------------------------------- Summary: Default threshold for blacklisting a node is too high Key: TEZ-3822 URL: https://issues.apache.org/jira/browse/TEZ-3822 Project: Apache Tez Issue Type: Bug Reporter: Zhiyuan Yang
By default, a task will be failed if 4 task attempts fail, which consequently fail the vertex and dag. By default, a node will be blacklisted if 10 task attempts fail on it. This number is higher than 4, so a single faulty node server can fail 4 task attempts by shuffle error, which finally fail the job, before this node is blacklisted. Even we can reschedule a task after it's blamed for input read error, we cannot avoid multiple tasks go to the same bad node and continue to cause shuffle error. -- This message was sent by Atlassian JIRA (v6.4.14#64029)