[ https://issues.apache.org/jira/browse/FLINK-35288?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17844106#comment-17844106 ]
Biao Geng commented on FLINK-35288: ----------------------------------- https://cwiki.apache.org/confluence/display/FLINK/FLIP-364%3A+Improve+the+exponential-delay+restart-strategy#FLIP364:Improvetheexponentialdelayrestartstrategy-1.2Differentsemanticsofrestartattemptscauseregionfailovernotasexpected In the above FLIP, there is some relevant discussion of the 'restart-strategy.fixed-delay.attempts' problem. When 'region-failover' (the default value of *jobmanager.execution.failover-strategy*) is enabled, the org.apache.flink.runtime.executiongraph.failover.ExecutionFailureHandler#handleFailure method is called once a subtask in a region fails, which consumes the job-level 'restart-strategy.fixed-delay.attempts'. As a result, the restart strategy may not work as the documentation described. We have also met such case in the production environment. > Flink Restart Strategy does not work as documented > -------------------------------------------------- > > Key: FLINK-35288 > URL: https://issues.apache.org/jira/browse/FLINK-35288 > Project: Flink > Issue Type: Bug > Reporter: Keshav Kansal > Priority: Minor > > As per the documentation when using the Fixed Delay Restart Strategy, the > *restart-strategy.fixed-delay.attempts* defines the "The number of times that > Flink retries the execution before the job is declared as failed if has been > set to fixed-delay". > However in reality it is the *maximum-total-task-failures*, i.e. it is > possbile that the job does not even attempt to restart. > This is as per documented in > https://cwiki.apache.org/confluence/display/FLINK/FLIP-1%3A+Fine+Grained+Recovery+from+Task+Failures > If there is an outage at a Sink level, for example Elasticsearch outage, all > the independent tasks might fail and the job will immediately fail without > restart (if restart-strategy.fixed-delay.attempts is set lower or equal to > the parallelism of the sink) -- This message was sent by Atlassian Jira (v8.20.10#820010)