You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@flink.apache.org by "Yaroslav Tkachenko (Jira)" <ji...@apache.org> on 2021/10/10 19:11:00 UTC
[jira] [Commented] (FLINK-21986) taskmanager native memory not
release timely after restart
[ https://issues.apache.org/jira/browse/FLINK-21986?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17426853#comment-17426853 ]
Yaroslav Tkachenko commented on FLINK-21986:
--------------------------------------------
We had a very similar problem in 1.13. Found this issue [https://github.com/facebook/rocksdb/issues/4112] and decided to try disabling RocksDB block cache:
{quote}state.backend.rocksdb.options-factory: xxx.NoBlockCacheRocksDbOptionsFactory
state.backend.rocksdb.memory.managed: false{quote}
Where NoBlockCacheRocksDbOptionsFactory essentially does:
{quote}val blockBasedTableConfig = new BlockBasedTableConfig()
blockBasedTableConfig.setNoBlockCache(true)
// Needed in order to disable block cache
blockBasedTableConfig.setCacheIndexAndFilterBlocks(false)
blockBasedTableConfig.setCacheIndexAndFilterBlocksWithHighPriority(false)
blockBasedTableConfig.setPinL0FilterAndIndexBlocksInCache(false){quote}
We did NOT see big performance degradation when running on SSDs.
> taskmanager native memory not release timely after restart
> ----------------------------------------------------------
>
> Key: FLINK-21986
> URL: https://issues.apache.org/jira/browse/FLINK-21986
> Project: Flink
> Issue Type: Bug
> Components: Runtime / State Backends
> Affects Versions: 1.11.3, 1.12.1, 1.13.0
> Environment: flink version:1.12.1
> run :yarn session
> job type:mock source -> regular join
>
> checkpoint interval: 3m
> Taskmanager memory : 16G
>
> Reporter: Feifan Wang
> Assignee: Feifan Wang
> Priority: Critical
> Labels: pull-request-available
> Fix For: 1.11.4, 1.13.0, 1.12.3
>
> Attachments: 82544.svg, image-2021-03-25-15-53-44-214.png, image-2021-03-25-16-07-29-083.png, image-2021-03-26-11-46-06-828.png, image-2021-03-26-11-47-21-388.png
>
>
> I run a regular join job with flink_1.12.1 , and find taskmanager native memory not release timely after restart cause by exceeded checkpoint tolerable failure threshold.
> *problem job information:*
> # job first restart cause by exceeded checkpoint tolerable failure threshold.
> # then taskmanager be killed by yarn many times
> # in this case,tm heap is set to 7.68G,bug all tm heap size is under 4.2G
> !image-2021-03-25-15-53-44-214.png|width=496,height=103!
> # nonheap size increase after restart,but still under 160M.
> !https://km.sankuai.com/api/file/cdn/706284607/716474606?contentType=1&isNewContent=false&isNewContent=false|width=493,height=102!
> # taskmanager process memory increase 3-4G after restart(this figure show one of taskmanager)
> !image-2021-03-25-16-07-29-083.png|width=493,height=107!
>
> *my guess:*
> [RocksDB wiki|https://github.com/facebook/rocksdb/wiki/RocksJava-Basics#memory-management] mentioned :Many of the Java Objects used in the RocksJava API will be backed by C++ objects for which the Java Objects have ownership. As C++ has no notion of automatic garbage collection for its heap in the way that Java does, we must explicitly free the memory used by the C++ objects when we are finished with them.
> So, is it possible that RocksDBStateBackend not call AbstractNativeReference#close() to release memory use by RocksDB C++ Object ?
> *I make a change:*
> Actively call System.gc() and System.runFinalization() every minute.
> *And run this test again:*
> # taskmanager process memory no obvious increase
> !image-2021-03-26-11-46-06-828.png|width=495,height=93!
> # job run for several days,and restart many times,but no taskmanager killed by yarn like before
>
> *Summary:*
> # first,there is some native memory can not release timely after restart in this situation
> # I guess it maybe RocksDB C++ object,but I hive not check it from source code of RocksDBStateBackend
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)