[ https://issues.apache.org/jira/browse/STORM-956?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15013674#comment-15013674 ]
ASF GitHub Bot commented on STORM-956: -------------------------------------- Github user revans2 commented on the pull request: https://github.com/apache/storm/pull/647#issuecomment-158085350 To make a feature like this work for storm, we need to provide people a lot of control over it. I personally would prefer to see something where the Bolt/Spout can indicate life without completely processing a tuple. Probably an API in OutputCollector that would allow it to heartbeat in saying don't shoot me. Any other interaction with an OutputCollection should also prevent the component from being shot. The timeout should be configurable on a per-component basis, including turning them off. Having all of them set to the same cluster wide setting is not good. Also if we detect that we are in a bad state there is no reason to stop heart-beating, and wait for a supervisor to shoot us. If we think we are bad just exit. It will let recovery happen much faster. I agree with @bastiliu too that one of the options should be not to shot ourselves, but to alert that we are seeing problems. I would like to see this done through both the metrics system and through writing an error into zookeeper that would show up on the UI for the component that is stuck. The metrics would allow for automated alerting and the UI would let users manually see what is happening. > When the execute() or nextTuple() hang on external resources, stop the > Worker's heartbeat > ----------------------------------------------------------------------------------------- > > Key: STORM-956 > URL: https://issues.apache.org/jira/browse/STORM-956 > Project: Apache Storm > Issue Type: Improvement > Components: storm-core > Reporter: Chuanlei Ni > Assignee: Chuanlei Ni > Priority: Minor > Original Estimate: 6h > Remaining Estimate: 6h > > Sometimes the work threads produced by mk-threads in executor.clj hang on > external resources or other unknown reasons. This makes the workers stop > processing the tuples. I think it is better to kill this worker to resolve > the "hang". I plan to : > 1. like `setup-ticks`, send a system-tick to receive-queue > 2. the tuple-action-fn deal with this system-tick and remember the time that > processes this tuple in the executor-data > 3. when worker do local heartbeat, check the time the executor writes to > executor-data. If the time is long from current (for example, 3 minutes), the > worker does not do the heartbeat. So the supervisor could deal with this > problem. -- This message was sent by Atlassian JIRA (v6.3.4#6332)