[
https://issues.apache.org/jira/browse/FLINK-27236?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17522631#comment-17522631
]
Lijie Wang commented on FLINK-27236:
------------------------------------
[~freeke]
-> At least, the previous execution shouldn't be still allocated to the tm .
Honestly, I have no idea about how to implement it. I think it is inevitable
that the slot states in RM/TM and JM will be inconsistent for a short period of
time, because they are different akka actors and can only synchronize
information via RPC.
In a streaming job (only one pipeline region), all task deployments are
performed in the same RPC Call (but this RPC time may be very long due to slow
deployement, for example 30s). That is, during this period, if RM/TM wants to
notify JM of slot states via RPC, JM can only handle it after 30s.
Unfortunately, at that point the task deployments have completed(and
encountered the mentioned problem).
Do you have any idea about this?
> No task slot allocated for job in larege-scale job
> --------------------------------------------------
>
> Key: FLINK-27236
> URL: https://issues.apache.org/jira/browse/FLINK-27236
> Project: Flink
> Issue Type: Bug
> Affects Versions: 1.13.3
> Reporter: yanpengshi
> Priority: Major
> Attachments: jobmanager.log.26, taskmanager.log, topology.png
>
> Original Estimate: 444h
> Remaining Estimate: 444h
>
> Hey，
>
> We run a large-scale flink job containing six vertices with 3k parallelism.
> The Topology is shown below.
> !topology.png!
> We meets the following exception in jobmanager.log:[^jobmanager.log.26]
> {code:java}
> 2022-03-02 08:01:16,601 INFO [1998]
> [org.apache.flink.runtime.executiongraph.Execution.transitionState(Execution.java:1446)]
> - Source: tdbank_exposure_wx -> Flat Map (772/3000)
> (6cd18d4ead1887a4e19fd3f337a6f4f8) switched from DEPLOYING to FAILED on
> container_e03_1639558254334_10048_01_004716 @ 11.104.77.40
> (dataPort=39313).java.util.concurrent.CompletionException:
> org.apache.flink.runtime.taskexecutor.exceptions.TaskSubmissionException: No
> task slot allocated for job ID 000000000000ed780000000000000087 and
> allocation ID beb058d837c09e8d5a4a6aaf2426ca99. {code}
>
> In the taskmanager.log [^taskmanager.log], the slot is freed due to timeout
> and the taskmanager receives the new allocated request. By increasing the
> value of key: taskmanager.slot.timeout， we can avoid this exception
> temporarily.
> Here are some our guesses：
> # When the job is scheduled, the slot and execution have been bound, and
> then the task is deployed to the corresponding taskmanager.
> # The slot is released after the idle interval times out and notify the
> ResouceManager the slot free. Thus, the resourceManager will assign other
> request to the slot.
> # The task is deployed to taskmanager according the previous correspondence
>
> The key problems are :
> # When the slot is free, the execution is not unassigned from the slot;
> # The slot state is not consistent in JobMaster and ResourceManager
>
> Has anyone else encountered this problem? When the slot is freed, how can we
> unassign the previous bounded execution? Or we need to update the resource
> address of the execution. @[~zhuzh] @[~wanglijie95]
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
