[FLINK-10694] ZooKeeperHaServices Cleanup

• Type: Bug
• Status: Closed
• Priority: Critical
• Resolution: Duplicate
• Affects Version/s: 1.6.1, 1.7.0
• Component/s: Runtime / Coordination
• Assignee: Unassigned
• Reporter: Mikhail Pryakhin
• Created: 26/Oct/18 15:58
• Updated: 01/Mar/24 08:16
• Resolved: 11/Jun/20 17:00

Description

When a streaming job with Zookeeper-HA enabled gets cancelled all the job-related Zookeeper nodes are not removed. Is there a reason behind that?
I noticed that Zookeeper paths are created of type "Container Node" (an Ephemeral node that can have nested nodes) and fall back to Persistent node type in case Zookeeper doesn't support this sort of nodes.
But anyway, it is worth removing the job Zookeeper node when a job is cancelled, isn't it?

zookeeper version 3.4.10
flink version 1.6.1

1. The job is deployed as a YARN cluster with the following properties set

    high-availability: zookeeper
    high-availability.zookeeper.quorum: <a list of zookeeper hosts>
    high-availability.zookeeper.storageDir: hdfs:///<recovery-folder-path>
    high-availability.zookeeper.path.root: <flink-root-path>
    high-availability.zookeeper.path.namespace: <flink-job-name>
   

1. The job is cancelled via flink cancel <job-id> command.

What I've noticed:
when the job is running the following directory structure is created in zookeeper

/<flink-root-path>/<flink-job-name>/leader/resource_manager_lock
/<flink-root-path>/<flink-job-name>/leader/rest_server_lock
/<flink-root-path>/<flink-job-name>/leader/dispatcher_lock
/<flink-root-path>/<flink-job-name>/leader/5c21f00b9162becf5ce25a1cf0e67cde/job_manager_lock
/<flink-root-path>/<flink-job-name>/leaderlatch/resource_manager_lock
/<flink-root-path>/<flink-job-name>/leaderlatch/rest_server_lock
/<flink-root-path>/<flink-job-name>/leaderlatch/dispatcher_lock
/<flink-root-path>/<flink-job-name>/leaderlatch/5c21f00b9162becf5ce25a1cf0e67cde/job_manager_lock
/<flink-root-path>/<flink-job-name>/checkpoints/5c21f00b9162becf5ce25a1cf0e67cde/0000000000000000041
/<flink-root-path>/<flink-job-name>/checkpoint-counter/5c21f00b9162becf5ce25a1cf0e67cde
/<flink-root-path>/<flink-job-name>/running_job_registry/5c21f00b9162becf5ce25a1cf0e67cde

when the job is cancelled some ephemeral nodes disappear, but most of them are still there:

/<flink-root-path>/<flink-job-name>/leader/5c21f00b9162becf5ce25a1cf0e67cde
/<flink-root-path>/<flink-job-name>/leaderlatch/resource_manager_lock
/<flink-root-path>/<flink-job-name>/leaderlatch/rest_server_lock
/<flink-root-path>/<flink-job-name>/leaderlatch/dispatcher_lock
/<flink-root-path>/<flink-job-name>/leaderlatch/5c21f00b9162becf5ce25a1cf0e67cde/job_manager_lock
/<flink-root-path>/<flink-job-name>/checkpoints/
/<flink-root-path>/<flink-job-name>/checkpoint-counter/
/<flink-root-path>/<flink-job-name>/running_job_registry/

Here is the method [1] responsible for cleaning zookeeper folders up [1] which is called when a job manager has stopped [2].
And it seems it only cleans up the running_job_registry folder, other folders stay untouched. I suppose that everything under the /<flink-root-path>/<flink-job-name>/ folder should be cleaned up when the job is cancelled.

[1] https://github.com/apache/flink/blob/8674b69964eae50cad024f2c5caf92a71bf21a09/flink-runtime/src/main/java/org/apache/flink/runtime/highavailability/zookeeper/ZooKeeperRunningJobsRegistry.java#L107
[2] https://github.com/apache/flink/blob/f087f57749004790b6f5b823d66822c36ae09927/flink-runtime/src/main/java/org/apache/flink/runtime/dispatcher/Dispatcher.java#L332

Issue Links

• Blocked: FLINK-34557 When the Flink task ends in application mode, there may be issues with the Znode and HDFS files not being deleted
• duplicates: FLINK-11336 Flink HA didn't remove ZK metadata
• is related to: FLINK-10333 Rethink ZooKeeper based stores (SubmittedJobGraph, MesosWorker, CompletedCheckpoints)