[
https://issues.apache.org/jira/browse/HIVE-6579?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13930012#comment-13930012
]
xieyuchen commented on HIVE-6579:
---------------------------------
[~xuefuz] thanks for your reply.
Except the testing code, there are actually 2 callers of this constructor:
o.a.h.h.ql.Driver.acquireReadWriteLocks and o.a.h.h.ql.exec.DDLTask.lockTable
The first situation will be fix in this patch. I think the second one will not
cause oom.
But if we have to keep HiveLockObjectData.queryStr trimmed, we can keep the
trimming call in the constructor.
> HiveLockObjectData constructor makes too many queryStr instance causing oom
> ---------------------------------------------------------------------------
>
> Key: HIVE-6579
> URL: https://issues.apache.org/jira/browse/HIVE-6579
> Project: Hive
> Issue Type: Improvement
> Reporter: xieyuchen
> Attachments: HIVE-6579.1.patch.txt
>
>
> We have a huge sql which full outer joins 10+ partitoned tables, each table
> has at least 1k partitions. The sql has 300kb in length(it constructed
> automatically of cause).
> So when we running this sql, there are over 10k HiveLockObjectData instances.
> Because of the constructor of HiveLockObjectData trim the queryStr, there
> will be 10k individual String instances, each has 300kb in length! Then the
> Hive client will get an oom exception.
> Trying to trim the queryStr in Driver.compile function instead of doing it in
> HiveLockObjectData constructor to reduce memory wasting.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
