[
https://issues.apache.org/jira/browse/HIVE-22850?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17035928#comment-17035928
]
Gopal Vijayaraghavan commented on HIVE-22850:
---------------------------------------------
[~rajesh.balamohan]: The jump-table looks like it can be reversed into a filter
condition so that you can avoid running 2 queries. And the size of the query
packet is very big, maybe we can remove the HL_PARTITION filter and filter the
result after the jump-table condition, if there are any conflicts returned (i.e
build a hash-set of the partition ids iff there are any collisions).
> Optimise lock acquisition in TxnHandler
> ---------------------------------------
>
> Key: HIVE-22850
> URL: https://issues.apache.org/jira/browse/HIVE-22850
> Project: Hive
> Issue Type: Improvement
> Components: Hive
> Reporter: Rajesh Balamohan
> Priority: Major
> Attachments: HIVE-22850.1.patch, HIVE-22850.2.patch,
> HIVE-22850.3.patch, Screenshot 2020-02-07 at 4.14.51 AM.jpg, jumpTableInfo.png
>
>
> With concurrent queries, time taken for lock acquisition increases
> substantially. As part of lock acquisition in the query,
> {{TxnHandler::checkLock}} gets invoked. This involves getting a mutex and
> compare the locks being requested for, with that of existing locks in
> {{HIVE_LOCKS}} table.
> With concurrent queries, time taken to do this check increase and this
> significantly increases the time taken for getting mutex for other threads
> (due to select for update). In a synthetic workload, it was in the order of
> 10+ seconds. This codepath can be optimized when all lock requests are
> SHARED_READ.
>
>
> !Screenshot 2020-02-07 at 4.14.51 AM.jpg|width=743,height=348!
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
