[
https://issues.apache.org/jira/browse/FLINK-34557?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17823473#comment-17823473
]
Matthias Pohl commented on FLINK-34557:
---------------------------------------
hi [~tltheshy], thanks for sharing your observations. can you share some
JobManager logs as well to get a better understanding of what's going on?
{quote}
After the Flink task is submitted to the cluster, if there is a conflict or
missing jar package, the task will be restarted multiple times by Yarn and
ultimately fail to end. At this point, it will be found that the Znode
persists, and there are files with corresponding appids in the '/.flink'
directory and '/flink/recovery' directory in HDFS;
[...]
When the Flink task is disconnected from zk (we will not analyze the specific
reason for the disconnection), if zk is disconnected from the jm container, the
task will hang and be pulled back by yarn. When the last disconnection occurs,
the task will eventually end and the same result as above will appear;
{quote}
This sounds like the job failed which should result in the job reaching the
globally terminal state with the data being cleaned up. Here, JM logs would be
helpful to understand what happens.
{quote}
When using the yarn kill command to kill a task, the task ends directly and the
final state is killed, with the final result being the same as the first one;
{quote}
Isn't this a normal failover where you don't want to delete the tasks to be
able to recover from it? Or am I misunderstanding you here?
> When the Flink task ends in application mode, there may be issues with the
> Znode and HDFS files not being deleted
> -----------------------------------------------------------------------------------------------------------------
>
> Key: FLINK-34557
> URL: https://issues.apache.org/jira/browse/FLINK-34557
> Project: Flink
> Issue Type: Improvement
> Components: Deployment / YARN, Runtime / Task
> Affects Versions: 1.17.0, 1.16.2
> Reporter: tanliang
> Priority: Major
> Attachments: image-2024-03-01-15-38-48-396.png,
> image-2024-03-01-15-39-13-953.png, image-2024-03-01-15-39-39-524.png
>
>
> In Flink 1.16.2, we all use application mode to submit tasks to Yarn.
> However, there are several situations during use that result in Znode not
> being deleted and some files on HDFS not being deleted. These should be
> deleted after the task is stopped, otherwise it may cause some resource
> occupancy problems. Below are the several situations I have encountered:
> # After the Flink task is submitted to the cluster, if there is a conflict
> or missing jar package, the task will be restarted multiple times by Yarn and
> ultimately fail to end. At this point, it will be found that the Znode
> persists, and there are files with corresponding appids in the '/.flink'
> directory and '/flink/recovery' directory in HDFS;
> # When using the yarn kill command to kill a task, the task ends directly
> and the final state is killed, with the final result being the same as the
> first one;
> # When the Flink task is disconnected from zk (we will not analyze the
> specific reason for the disconnection), if zk is disconnected from the jm
> container, the task will hang and be pulled back by yarn. When the last
> disconnection occurs, the task will eventually end and the same result as
> above will appear;
> !image-2024-03-01-15-38-48-396.png|width=877,height=171!
> !image-2024-03-01-15-39-13-953.png|width=882,height=174!
>
> !image-2024-03-01-15-39-39-524.png|width=1001,height=67!
>
> *Add：*
> Through consulting with the community and other colleagues, we found that the
> community had previously raised the issue of Znode not being deleted. Later,
> by adding the closeAndCleanupAllData# method, it was uniformly deleted at the
> end of a highly available cluster. However, in the aforementioned situations,
> there are still issues with file and data residue. Among them, when using the
> yarn kill command, after successfully submitting a task to the cluster, Flink
> would indicate through console logs that there would be HDFS file residue
> after successfully submitting tasks to the cluster, however, I don't
> understand why the community did not improve this and instead retained the
> existence of this situation. At the same time, we believe that Znode residue
> should not exist, regardless of the task status, it must be cleaned up after
> stopping the task
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
