[ 
https://issues.apache.org/jira/browse/HADOOP-14621?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16265452#comment-16265452
 ] 

Steve Loughran commented on HADOOP-14621:
-----------------------------------------

This is possibly now remapped to ConnectionTimeoutException and so 
retried...we'd need to experiment

ConnectionPoolTimeoutException is a sign of a serious problem, as it can be 
raised if the pool of http threads has been used up by files not being closed 
in user code (SPARK-22526), elsewhere (HIVE-13216). At the same time, 
completely spurious if it is raised early and often.

Proposed
# Make sure translation preserves it as its own exception (interruptedIOE with 
text)
# add commentary in troubleshooting doc
# review timeout value  and pool size, consider recommending values (2 per CPU?)
# could we make queue size a gauge in metrics ?
# have a test which creates a very small pool and timeout and triggers a 
failure (GET on small file and read() one byte should be enough to use up a 
connection.

Look @ other object store code and see if the same policy applies

> S3A client raising ConnectionPoolTimeoutException
> -------------------------------------------------
>
>                 Key: HADOOP-14621
>                 URL: https://issues.apache.org/jira/browse/HADOOP-14621
>             Project: Hadoop Common
>          Issue Type: Sub-task
>          Components: fs/s3, test
>    Affects Versions: 3.0.0-beta1
>         Environment: Home network with 2+ other users on high bandwidth 
> activities
>            Reporter: Steve Loughran
>            Priority: Minor
>
> Parallel test with threads = 12 triggering connection pool timeout. 
> Hypothesis? Congested network triggering pool timeout.
> Fix? For tests, could increase pool size
> For retry logic, this should be considered retriable, even on idempotent 
> calls (as its a failure to acquire a connection



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to