[
https://issues.apache.org/jira/browse/IMPALA-9291?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17024572#comment-17024572
]
ASF subversion and git services commented on IMPALA-9291:
---------------------------------------------------------
Commit 6ac51522d8505416e8ef7ae2d911741fd1c4a3e5 in impala's branch
refs/heads/master from Zoltan Borok-Nagy
[ https://gitbox.apache.org/repos/asf?p=impala.git;h=6ac5152 ]
IMPALA-9291: MemTracker's TryConsume() should behave consistent with
Consume()/Release()
Before this fix MemTracker::Consume() and Release() could take
negative numbers as arguments. In that case Consume(-val) would invoke
Release(val) and vice versa.
On the other hand, TryConsume() was a no-op for negative values and
returned success. This could lead to errors in the following case:
TryConsume(-42); // no-op
...
Release(-42); // -> Consume(42);
And in the end the mem tracker would think there is 42 bytes of
unallocated memory.
This commit changes mem-tracker to forbid negative values.
Another inconsistency was found between Consume()/Release() and
TryConsume(). When Consume()/Release() was invoked with zero as
argument they didn't update 'consumption_' when 'consumption_metric_'
was not null. On the other hand, TryConsume() being invoked with
zero did update 'consumption_' from the metric. I fixed this
inconsistency as well, i.e. now 'consumption_' is not updated
when the argument is zero.
Testing:
* ran exhaustive tests
Change-Id: I22e613c49e4a6bd218e7f7f7c0d6bec95b75cff5
Reviewed-on: http://gerrit.cloudera.org:8080/15050
Reviewed-by: Impala Public Jenkins <[email protected]>
Tested-by: Impala Public Jenkins <[email protected]>
> MemTracker's TryConsume() should behave consistent with Consume()/Release()
> ---------------------------------------------------------------------------
>
> Key: IMPALA-9291
> URL: https://issues.apache.org/jira/browse/IMPALA-9291
> Project: IMPALA
> Issue Type: Bug
> Reporter: Zoltán Borók-Nagy
> Assignee: Zoltán Borók-Nagy
> Priority: Major
>
> Impala's MemTracker tracks allocated memory. MemTracker::Consume() is meant
> to be used in tandem with memory allocations and MemTracker::Release() is for
> de-allocations.
> However, both supports negative values, in this case
> MemTracker::Consume(negative_val) will invoke
> MemTracker::Release(*-*negative_val), and vica versa.
> On the other hand, TryConsume() is a no-op for values less than zero. This
> causes problems in the following scenario:
>
> {noformat}
> TryConsume(-42); // no-op, returns true which means SUCCESS
> Release(-42); // => Consume(42)
> {noformat}
> Later when the mem tracker gets closed it will think there is 42 bytes of
> unallocated memory.
> Either we shouldn't allow negative values in
> Consume()/TryConsume()/Release(), or TryConsume() should invoke Release() in
> this case instead of doing nothing.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
