andygrove opened a new pull request, #3632: URL: https://github.com/apache/datafusion-comet/pull/3632
## Which issue does this PR close? <!-- We generally require a GitHub issue to be filed for all bug fixes and enhancements and this helps us generate change logs for our releases. You can link an issue to this PR using the GitHub syntax. For example `Closes #123` indicates that this PR will close issue #123. --> Research towards https://github.com/apache/datafusion-comet/issues/2545 ## Rationale for this change <!-- Why are you proposing this change? If this is already explained clearly in the issue then this section is not needed. Explaining clearly why changes are proposed helps reviewers understand your changes and offer better suggestions for fixes. --> Comet's current experimental hash join operator is not suitable for use in production because it has no spilling support and will OOM if the build side is too large. For example, it was not possible to run the TPC-DS benchmarks with hash joins enabled prior to this PR. This PR replaces it with an experimental Grace hash join operator which does have spilling. ## Benchmark Results | Benchmark | SMJ | Hash Join | Grace Hash Join | |-|-|-| | TPC-H | 278s | 208s | 250s | | TPC-DS | 704s | OOM | 664s | ## What changes are included in this PR? <!-- There is no need to duplicate the description in the issue here but it is sometimes worth providing a summary of the individual changes in this PR. --> - Replace SortMergeJoinExec with ShuffledHashJoinExec via RewriteJoin rule (removes input sorts), executed natively as GraceHashJoinExec - Hash-partition both build and probe sides into N buckets (default 16) using prefix-sum algorithm for cache-friendly O(rows) partitioning - Fast path: when build side fits in memory and no spilling occurred, skip partitioning overhead — single HashJoinExec with streaming probe - Slow path: merge adjacent partitions to ~32 MB groups, join sequentially with per-partition HashJoinExec - Spill to disk via Arrow IPC with 1 MB buffered writes; streaming reads via SpillReaderExec with batch coalescing (~8192 rows) - Recursive repartitioning (up to depth 3 / 4096 effective partitions) when individual partitions exceed hash table memory - Cooperative memory management: single spillable MemoryReservation during partitioning; aggressive "spill all" strategy on probe-side memory pressure to avoid thrashing between concurrent instances - Supports all join types: Inner, Left, Right, Full, LeftSemi, LeftAnti, LeftMark, RightSemi, RightAnti, RightMark - Configurable SMJ replacement guard (maxBuildSize) to keep sort-merge join when both sides are large - Fast path threshold divided by executor cores to bound per-task memory ## How are these changes tested? <!-- We typically require tests for all PRs in order to: 1. Prevent the code from being accidentally broken by subsequent changes 2. Serve as another way to document the expected behavior of the code If tests are not included in your PR, please explain why (for example, are they covered by existing tests)? --> -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
