[ https://issues.apache.org/jira/browse/SPARK-20869?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
lyc updated SPARK-20869: ------------------------ Description: In `Master.removeWorker`, master clears executor and driver state, but does not clear app state. App state is cleared when received `UnregisterApplication` and when `onDisconnect`, the first is when driver shutdown gracefully, the second is called when `netty`'s `channelInActive` is called (which is called when channel is closed), both of which can not handle the case when there is a network partition between master and worker. Follow the steps in [SPARK-19900|https://issues.apache.org/jira/browse/SPARK-19900], and see the [screenshots|https://cloud.githubusercontent.com/assets/2576762/26398697/d50735a4-40ac-11e7-80d8-6e9e1cf0b62f.png] when worker1 partitions with master, the app `app-xxx-000` is still running instead of finished because of worker1 is down. cc @CodingCat was: In `Master.removeWorker`, master clears executor and driver state, but does not clear app state. App state is cleared when received `UnregisterApplication` and when `onDisconnect`, the first is when driver shutdown gracefully, the second is called when `netty`'s `channelInActive` is called, both of which can not handle the case when there is a network partition between master and worker. Follow the steps in [SPARK-19900|https://issues.apache.org/jira/browse/SPARK-19900], and see the [screenshots|https://cloud.githubusercontent.com/assets/2576762/26398697/d50735a4-40ac-11e7-80d8-6e9e1cf0b62f.png] when worker1 partitions with master, the app `app-xxx-000` is still running instead of finished because of worker1 is down. cc @CodingCat > Master should clear failed apps when worker down > ------------------------------------------------ > > Key: SPARK-20869 > URL: https://issues.apache.org/jira/browse/SPARK-20869 > Project: Spark > Issue Type: Improvement > Components: Spark Core > Affects Versions: 2.3.0 > Reporter: lyc > Priority: Minor > Original Estimate: 24h > Remaining Estimate: 24h > > In `Master.removeWorker`, master clears executor and driver state, but does > not clear app state. App state is cleared when received > `UnregisterApplication` and when `onDisconnect`, the first is when driver > shutdown gracefully, the second is called when `netty`'s `channelInActive` is > called (which is called when channel is closed), both of which can not handle > the case when there is a network partition between master and worker. > Follow the steps in > [SPARK-19900|https://issues.apache.org/jira/browse/SPARK-19900], and see the > [screenshots|https://cloud.githubusercontent.com/assets/2576762/26398697/d50735a4-40ac-11e7-80d8-6e9e1cf0b62f.png] > when worker1 partitions with master, the app `app-xxx-000` is still running > instead of finished because of worker1 is down. > cc @CodingCat -- This message was sent by Atlassian JIRA (v6.3.15#6346) --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org