[jira] [Commented] (FLINK-12547) Deadlock when the task thread downloads jars using BlobClient

["Qi (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/FLINK-12547?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16848313#comment-16848313 ] 

Qi commented on FLINK-12547:
----------------------------

Hi Haibo,

 

We had added SO_TIMEOUT to the blob socket (SO_TIMEOUT = 30min) in our internal Flink release, but this deadlock issue still occurs as described in https://issues.apache.org/jira/browse/FLINK-12426. The TM hangs for over an hour (longer than the SO_TIMEOUT).

 

We've observed that the JM is under heavy CPU load during the deadlock issue, but recovers after a few minutes. This looks pretty strange to me.

 

Thanks,

Qi

> Deadlock when the task thread downloads jars using BlobClient
> -------------------------------------------------------------
>
>                 Key: FLINK-12547
>                 URL: https://issues.apache.org/jira/browse/FLINK-12547
>             Project: Flink
>          Issue Type: Bug
>          Components: Runtime / Coordination
>    Affects Versions: 1.8.0
>            Reporter: Haibo Sun
>            Assignee: Haibo Sun
>            Priority: Major
>              Labels: pull-request-available
>          Time Spent: 10m
>  Remaining Estimate: 0h
>
> The jstack is as follows (this jstack is from an old Flink version, but the master branch has the same problem).
> {code:java}
> "Source: Custom Source (76/400)" #68 prio=5 os_prio=0 tid=0x00007f8139cd3000 nid=0xe2 runnable [0x00007f80da5fd000]
> java.lang.Thread.State: RUNNABLE
> at java.net.SocketInputStream.socketRead0(Native Method)
> at java.net.SocketInputStream.socketRead(SocketInputStream.java:116)
> at java.net.SocketInputStream.read(SocketInputStream.java:170)
> at java.net.SocketInputStream.read(SocketInputStream.java:141)
> at org.apache.flink.runtime.blob.BlobInputStream.read(BlobInputStream.java:152)
> at org.apache.flink.runtime.blob.BlobInputStream.read(BlobInputStream.java:140)
> at org.apache.flink.runtime.blob.BlobClient.downloadFromBlobServer(BlobClient.java:164)
> at org.apache.flink.runtime.blob.AbstractBlobCache.getFileInternal(AbstractBlobCache.java:181)
> at org.apache.flink.runtime.blob.PermanentBlobCache.getFile(PermanentBlobCache.java:206)
> at org.apache.flink.runtime.execution.librarycache.BlobLibraryCacheManager.registerTask(BlobLibraryCacheManager.java:120)
> - locked <0x000000062cf2a188> (a java.lang.Object)
> at org.apache.flink.runtime.taskmanager.Task.createUserCodeClassloader(Task.java:968)
> at org.apache.flink.runtime.taskmanager.Task.run(Task.java:604)
> at java.lang.Thread.run(Thread.java:834)
> Locked ownable synchronizers:
> - None
> {code}
>  
> The reason is that SO_TIMEOUT is not set in the socket connection of the blob client. When the network packet loss seriously due to the high CPU load of the machine, the blob client connection fails to perceive that the server has been disconnected, which results in blocking in the native method. 



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)