[
https://issues.apache.org/jira/browse/TEZ-711?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Rohini Palaniswamy updated TEZ-711:
-----------------------------------
Attachment: OOM-threaddump-711-5-patch.txt
Even with this patch OOM is still happening, but the job succeeds. I am
assuming retries. Haven't dug into AM logs to confirm if there was another
attempt for the failed task.
One difference is that without patch, threads were stuck at socket read in
HttpURLConnection$HttpInputStream.read, but with the patch they are stuck at
EofSensorInputStream.read
> Fix memory leak when not reading from inputs due to caching
> -----------------------------------------------------------
>
> Key: TEZ-711
> URL: https://issues.apache.org/jira/browse/TEZ-711
> Project: Apache Tez
> Issue Type: Bug
> Affects Versions: 0.2.0
> Reporter: Rohini Palaniswamy
> Assignee: Siddharth Seth
> Priority: Critical
> Attachments: OOM-threaddump-711-5-patch.txt,
> OOM-threaddump-till-TEZ-752.txt, TEZ-711.5.txt, TEZ-711.wip.1.txt,
> TEZ-711.wip.2.txt, TEZ-711.wip.3.txt, TEZ-711.wip.4.txt
>
>
> When you are reading from inputs and caching objects with vertex scope, you
> don't have to read the input again when container is reused. But it allocates
> memory and that leaks causing OOM. KeyValueReader does not have a API to
> close the reader to clear allotted memory without reading from it. Also if
> there was a option to pre-close inputs in Processor and not fetch input at
> all over the wire and do shuffle/sort it would be a good optimization.
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
