[
https://issues.apache.org/jira/browse/HIVE-16949?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Birger Brunswiek updated HIVE-16949:
------------------------------------
Description:
The commit
[20210de|https://github.com/apache/hive/commit/20210dec94148c9b529132b1545df3dd7be083c3]
which was part of HIVE-15546 [introduced a thread
pool|https://github.com/apache/hive/blob/824b9c80b443dc4e2b9ad35214a23ac756e75234/ql/src/java/org/apache/hadoop/hive/ql/exec/Utilities.java#L3109]
which is not shutdown upon completion of its threads. This leads to a leak of
threads for each query which uses more than 1 partition. They are not removed
by the GC. When queries spanning multiple partitions are made the number of
threads increases and is never reduced. On my machine hiveserver2 starts to get
slower and slower once 10k threads are reached.
Thread pools only shutdown automatically in special circumstances (see
[documentation section
_Finalization_|https://docs.oracle.com/javase/7/docs/api/java/util/concurrent/ThreadPoolExecutor.html]).
I am not sure why this is not the case. I would add a _pool.shutdown()_ just
[after the pool has completed its
work|https://github.com/apache/hive/blob/824b9c80b443dc4e2b9ad35214a23ac756e75234/ql/src/java/org/apache/hadoop/hive/ql/exec/Utilities.java#L3137]
to make sure the threads are really shutdown. This, however, would only fix
normal operation. There are other exit points, namely through exceptions, which
would still lead to the same leak of threads.
My current workaround is to set {{hive.exec.input.listing.max.threads = 1}}.
This prevents the the thread pool from being spawned
[\[1\]|https://github.com/apache/hive/blob/824b9c80b443dc4e2b9ad35214a23ac756e75234/ql/src/java/org/apache/hadoop/hive/ql/exec/Utilities.java#L2118]
[\[2\]|https://github.com/apache/hive/blob/824b9c80b443dc4e2b9ad35214a23ac756e75234/ql/src/java/org/apache/hadoop/hive/ql/exec/Utilities.java#L3107].
The same issue probably also applies to the [Get-Input-Summary thread
pool|https://github.com/apache/hive/blob/824b9c80b443dc4e2b9ad35214a23ac756e75234/ql/src/java/org/apache/hadoop/hive/ql/exec/Utilities.java#L2193].
was:
The commit
[20210de|https://github.com/apache/hive/commit/20210dec94148c9b529132b1545df3dd7be083c3]
which was part of HIVE-15546 [introduced a thread
pool|https://github.com/apache/hive/blob/824b9c80b443dc4e2b9ad35214a23ac756e75234/ql/src/java/org/apache/hadoop/hive/ql/exec/Utilities.java#L3109]
which is not shutdown upon completion of its threads. This leads to a leak of
threads for each query which uses more than 1 partition. They are not removed
by the GC. When queries spanning multiple partitions are made the number of
threads increases and is never reduced. On my machine hiveserver2 starts to get
slower and slower once 10k threads are reached.
Thread pools should be should be [shutdown
automatically|https://docs.oracle.com/javase/7/docs/api/java/util/concurrent/ThreadPoolExecutor.html].
I am not sure why this is not the case. I would add a _pool.shutdown()_ just
[after the pool has completed its
work|https://github.com/apache/hive/blob/824b9c80b443dc4e2b9ad35214a23ac756e75234/ql/src/java/org/apache/hadoop/hive/ql/exec/Utilities.java#L3137]
to make sure the threads are really shutdown. This, however, would only fix
normal operation. There are other exit points, namely through exceptions, which
would still lead to the same leak of threads.
My current workaround is to set {{hive.exec.input.listing.max.threads = 1}}.
This prevents the the thread pool from being spawned
[\[1\]|https://github.com/apache/hive/blob/824b9c80b443dc4e2b9ad35214a23ac756e75234/ql/src/java/org/apache/hadoop/hive/ql/exec/Utilities.java#L2118]
[\[2\]|https://github.com/apache/hive/blob/824b9c80b443dc4e2b9ad35214a23ac756e75234/ql/src/java/org/apache/hadoop/hive/ql/exec/Utilities.java#L3107].
The same issue probably also applies to the [Get-Input-Summary thread
pool|https://github.com/apache/hive/blob/824b9c80b443dc4e2b9ad35214a23ac756e75234/ql/src/java/org/apache/hadoop/hive/ql/exec/Utilities.java#L2193].
> Leak of threads from Get-Input-Paths thread pool when more than 1 used in
> query
> -------------------------------------------------------------------------------
>
> Key: HIVE-16949
> URL: https://issues.apache.org/jira/browse/HIVE-16949
> Project: Hive
> Issue Type: Bug
> Components: HiveServer2
> Reporter: Birger Brunswiek
>
> The commit
> [20210de|https://github.com/apache/hive/commit/20210dec94148c9b529132b1545df3dd7be083c3]
> which was part of HIVE-15546 [introduced a thread
> pool|https://github.com/apache/hive/blob/824b9c80b443dc4e2b9ad35214a23ac756e75234/ql/src/java/org/apache/hadoop/hive/ql/exec/Utilities.java#L3109]
> which is not shutdown upon completion of its threads. This leads to a leak
> of threads for each query which uses more than 1 partition. They are not
> removed by the GC. When queries spanning multiple partitions are made the
> number of threads increases and is never reduced. On my machine hiveserver2
> starts to get slower and slower once 10k threads are reached.
> Thread pools only shutdown automatically in special circumstances (see
> [documentation section
> _Finalization_|https://docs.oracle.com/javase/7/docs/api/java/util/concurrent/ThreadPoolExecutor.html]).
> I am not sure why this is not the case. I would add a _pool.shutdown()_ just
> [after the pool has completed its
> work|https://github.com/apache/hive/blob/824b9c80b443dc4e2b9ad35214a23ac756e75234/ql/src/java/org/apache/hadoop/hive/ql/exec/Utilities.java#L3137]
> to make sure the threads are really shutdown. This, however, would only fix
> normal operation. There are other exit points, namely through exceptions,
> which would still lead to the same leak of threads.
> My current workaround is to set {{hive.exec.input.listing.max.threads = 1}}.
> This prevents the the thread pool from being spawned
> [\[1\]|https://github.com/apache/hive/blob/824b9c80b443dc4e2b9ad35214a23ac756e75234/ql/src/java/org/apache/hadoop/hive/ql/exec/Utilities.java#L2118]
>
> [\[2\]|https://github.com/apache/hive/blob/824b9c80b443dc4e2b9ad35214a23ac756e75234/ql/src/java/org/apache/hadoop/hive/ql/exec/Utilities.java#L3107].
> The same issue probably also applies to the [Get-Input-Summary thread
> pool|https://github.com/apache/hive/blob/824b9c80b443dc4e2b9ad35214a23ac756e75234/ql/src/java/org/apache/hadoop/hive/ql/exec/Utilities.java#L2193].
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
