Core: Fix ParallelIterable memory leak where queue continues to be populated even after iterator close #9402

Heltman · 2024-01-03T12:14:44Z

This patch is a part for fix #7843 , we will need to add BlockingParallelIterable to fix the whole problem.

Heltman · 2024-01-03T12:16:23Z

see #7844 for whole discuss

Heltman · 2024-01-15T07:57:40Z

@findepi @electrum cc

etra · 2024-05-21T13:29:01Z

Any plans for when this will be released?

Thank you

raunaqmorarka · 2024-06-27T16:55:02Z

@Heltman could you please rebase your PR
cc: @nastra

core/src/test/java/org/apache/iceberg/util/TestParallelIterable.java

core/src/main/java/org/apache/iceberg/util/ParallelIterable.java

nastra · 2024-06-28T07:03:29Z

Overall the change makes sense, just a few small comments to address

…added even if iterator exited

Heltman · 2024-06-28T07:40:40Z

@nastra @raunaqmorarka Thanks for review and attention about this.

nastra

LGTM, I verified that this fixes the underlying issue. I'll leave the PR open for a bit to give others a chance to review this as well

amogh-jahagirdar

Thanks @Heltman for fixing this leak, I also agree with this solution since the population of the queue can't really be "interrupted". I'll go ahead and merge.

amogh-jahagirdar · 2024-07-01T21:33:42Z

and thanks @nastra @Fokko @findepi for the reviews !

…pulated even after iterator close (apache#9402)

* Core: Fix ParallelIterable memory leak where queue continues to be populated even after iterator close (apache#9402) (cherry picked from commit d3cb1b6) * Core: Limit ParallelIterable memory consumption by yielding in tasks (apache#10691) ParallelIterable schedules 2 * WORKER_THREAD_POOL_SIZE tasks for processing input iterables. This defaults to 2 * # CPU cores. When one or some of the input iterables are considerable in size and the ParallelIterable consumer is not quick enough, this could result in unbounded allocation inside `ParallelIterator.queue`. This commit bounds the queue. When queue is full, the tasks yield and get removed from the executor. They are resumed when consumer catches up. (cherry picked from commit 7831a8d) * Drop ParallelIterable's queue low water mark (apache#10978) As part of the change in commit 7831a8d, queue low water mark was introduced. However, it resulted in increased number of manifests being read when planning LIMIT queries in Trino Iceberg connector. To avoid increased I/O, back out the change for now. (cherry picked from commit bcb3281) --------- Co-authored-by: Helt <[email protected]> Co-authored-by: Piotr Findeisen <[email protected]>

…pulated even after iterator close (apache#9402)

github-actions bot added the core label Jan 3, 2024

Heltman mentioned this pull request Jan 3, 2024

Core: Add param to limit manifest parallel reader queue size #7844

Closed

Heltman force-pushed the parallel_memory_leak_fix branch from 9793b4a to 5e40a9e Compare January 4, 2024 01:53

Heltman force-pushed the parallel_memory_leak_fix branch from 5e40a9e to 27f74d1 Compare June 28, 2024 06:38