[
https://issues.apache.org/jira/browse/FLINK-3291?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15144755#comment-15144755
]
Gabor Gevay commented on FLINK-3291:
------------------------------------
{quote}
The initial idea of a contract for MutableObjectiterator.next(reuse) was the
following:
1. The caller may not hold onto reuse any more
2. The iterator implementor may not hold onto the returned object any more.
Given that this was long ago (5 years probably, since I created that
interface), I am pretty sure that contract is not obeyed everywhere.
{quote}
OK, this clears things up; thanks for chiming in, [~StephanEwen]!
Then the things we should do are
1. Add this contract to the javadoc of `MutableObjectiterator.next(reuse)`.
2. Go with [~greghogan]'s solution \[1\] to fix the problem brought up by this
Jira.
3. Check all calls to `MutableObjectiterator.next(reuse)`. (A few suspicious
ones are in `CrossDriver`, `UnionWithTempOperator`,
`MutableHashTable.ProbeIterator.next`, and
`ReusingBuildFirstHashJoinIterator.callWithNextKey`.)
{quote}
Gabor Gevay, at this point are the changes to MergeIterator fixing a bug? Do
you want to fix up and clarify the documentation for MutableObjectIterator and
verify the implementing classes?
{quote}
My changes to `MergeIterator` become unnecessary if the contracts of
`MutableObjectIterator.next(reuse)` are what Stephan said.
I'll open a Jira for fixing the javadoc of `MutableObjectIterator.next(reuse)`
to include the contract (and possibly also add this info to a new wiki page
that will be linked from the new version of the object reuse documentation
brewing at \[2\]), and check all call sites whether they obey the contract.
\[1\] https://github.com/apache/flink/pull/1626
\[2\]
https://docs.google.com/document/d/1cgkuttvmj4jUonG7E2RdFVjKlfQDm_hE6gvFcgAfzXg/edit
> Object reuse bug in MergeIterator.HeadStream.nextHead
> -----------------------------------------------------
>
> Key: FLINK-3291
> URL: https://issues.apache.org/jira/browse/FLINK-3291
> Project: Flink
> Issue Type: Bug
> Components: Distributed Runtime
> Affects Versions: 1.0.0
> Reporter: Gabor Gevay
> Assignee: Gabor Gevay
> Priority: Critical
>
> MergeIterator.HeadStream.nextHead saves a reference into `this.head` of the
> `reuse` object that it got as an argument. This object might be modified
> later by the caller.
> This actually happens when ReduceDriver.run calls input.next (which will
> actually be MergeIterator.next(E reuse)) in the inner while loop of the
> objectReuseEnabled branch, and that calls top.nextHead with the reference
> that it got from ReduceDriver, which erroneously saves the reference, and
> then ReduceDriver later uses that same object for doing the reduce.
> Another way in which this fails is when MergeIterator.next(E reuse) gives
> `reuse` to different `top`s in different calls, and then the heads end up
> being the same object.
> You can observe the latter situation in action by running ReducePerformance
> here:
> https://github.com/ggevay/flink/tree/merge-iterator-object-reuse-bug
> Set memory to -Xmx200m (so that the MergeIterator actually has merging to
> do), put a breakpoint at the beginning of MergeIterator.next(reuse), and then
> watch `reuse`, and the heads of the first two elements of `this.heap` in the
> debugger. They will get to be the same object after hitting continue about 6
> times.
> You can also look at the count that is printed at the end, which shouldn't be
> larger than the key range. Also, if you look into the output file
> /tmp/xxxobjectreusebug, for example the key 999977 appears twice.
> The good news is that I think I can see an easy fix that doesn't affect
> performance: MergeIterator.HeadStream could have a reuse object of its own as
> a member, and give that to iterator.next in nextHead(E reuse). And then we
> wouldn't need the overload of nextHead that has the reuse parameter, and
> MergeIterator.next(E reuse) could just call its other overload.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
