[
https://issues.apache.org/jira/browse/IMPALA-5384?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16480165#comment-16480165
]
ASF subversion and git services commented on IMPALA-5384:
---------------------------------------------------------
Commit c1c122a10177920903009420d2faac673d867c4b in impala's branch
refs/heads/master from [~dhecht]
[ https://git-wip-us.apache.org/repos/asf?p=impala.git;h=c1c122a ]
IMPALA-5384, part 2: Simplify Coordinator locking and clarify state
The is the final change to clarify and break up the Coordinator's lock.
The state machine for the coordinator is made explicit, distinguishing
between executing state and multiple terminal states. Logic to
transition into a terminal state is centralized in one location and
executes exactly once for each coordinator object.
Derived from a patch for IMPALA_5384 by Marcel Kornacker.
Testing:
- exhaustive functional tests
- stress test on minicluster with memory overcommitment. Verified from
the logs that this exercises all these paths:
- successful queries
- client requested cancellation
- error from exec FInstances RPC
- error reported asynchronously via report status RPC
- eos before backend execution completed
- loop query_test & failure for 12 hours with no dchecks or crashes
(This had previously reproduced IMPALA-7030 and IMPALA-7033 with
the previous version of this change).
Change-Id: I6dc08da1295f1df3c9dce6d35d65d887b2c00a1c
Reviewed-on: http://gerrit.cloudera.org:8080/10440
Reviewed-by: Dan Hecht <dhe...@cloudera.com>
Tested-by: Impala Public Jenkins <impala-public-jenk...@cloudera.com>
> Simplify coordinator locking protocol
> -------------------------------------
>
> Key: IMPALA-5384
> URL: https://issues.apache.org/jira/browse/IMPALA-5384
> Project: IMPALA
> Issue Type: Improvement
> Affects Versions: Impala 2.9.0
> Reporter: Marcel Kornacker
> Assignee: Dan Hecht
> Priority: Major
> Fix For: Impala 2.13.0, Impala 3.1.0
>
>
> The coordinator has a central lock (lock_) which is used very liberally to
> synchronize state changes that don't need to be synchronized, creating a
> concurrency bottleneck.
> Also, the coordinator contains a number of data structures related to INSERT
> finalization that don't need to be part of and synchronized with the rest of
> the coordinator state.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-all-unsubscr...@impala.apache.org
For additional commands, e-mail: issues-all-h...@impala.apache.org
