Add fetch to SortPreservingMergeExec and SortPreservingMergeStream
#6811
Conversation
Dandandan
changed the title
SortPreservingMergeExec
Dandandan
changed the title
SortPreservingMergeExecSortPreservingMergeExec and SortPreservingMergeStream
…serving_merge_exec
| &self.expr, | ||
| metrics, | ||
| self.batch_size, | ||
| self.fetch, |
There was a problem hiding this comment.
This was removed a few months ago to use streaming_merge in SortExec causing SortExec to spill even with a (low) fetch.
alamb
left a comment
alamb
left a comment
There was a problem hiding this comment.
I reviewed this carefully -- thank you @Dandandan .
I think it is worth considering adding fetch to the fuzz test cases in https://github.com/apache/arrow-datafusion/blob/main/datafusion/core/tests/fuzz_cases/merge_fuzz.rs to add some additional test coverage
@tustvold could you give this quick review for any issues?
| }) | ||
| .collect::<Result<_>>()?; | ||
|
|
||
| // TODO: Pushdown fetch to streaming merge (#6000) |
| physical_plan | ||
| GlobalLimitExec: skip=0, fetch=10 | ||
| --SortPreservingMergeExec: [supplier_cnt@3 DESC,p_brand@0 ASC NULLS LAST,p_type@1 ASC NULLS LAST,p_size@2 ASC NULLS LAST] | ||
| --SortPreservingMergeExec: [supplier_cnt@3 DESC,p_brand@0 ASC NULLS LAST,p_type@1 ASC NULLS LAST,p_size@2 ASC NULLS LAST], fetch=10 |
datafusion/core/src/physical_plan/sorts/sort_preserving_merge.rs
Outdated
Show resolved
Hide resolved
| self.loser_tree_adjusted = false; | ||
| self.in_progress.push_row(stream_idx); | ||
| if self.in_progress.len() < self.batch_size { | ||
|
|
There was a problem hiding this comment.
This is the hot path for SortPreservingMerge -- I suspect adding an extra check for fetch won't impact performance in any measurable way, but it might be worth while checking the benchmarks
There was a problem hiding this comment.
Makes sense, otherwise we could maybe need two separate implementations (with/without fetch)
There was a problem hiding this comment.
As benchmarks showed no significant difference, I kept this check in the hot path.
Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org>
jychen7
left a comment
jychen7
left a comment
There was a problem hiding this comment.
a small comment, otherwise LGTM.
Dandandan
changed the title
SortPreservingMergeExec and SortPreservingMergeStreamSortPreservingMergeExec and SortPreservingMergeStream
…serving_merge_exec
There was a problem hiding this comment.
I have left some inline comments, other than these, this PR is LGTM!. Thanks @Dandandan for this PR.
Co-authored-by: Mustafa Akur <106137913+mustafasrepo@users.noreply.github.com>
Co-authored-by: Mustafa Akur <106137913+mustafasrepo@users.noreply.github.com>
|
Thanks everyone -- this looks like it was a great team effort 👍 |
apache#6811) * Add fetch to sortpreservingmergeexec * Add fetch to sortpreservingmergeexec * fmt * Deserialize * Fmt * Fix test * Fix test * Fix test * Fix plan output * Doc * Update datafusion/core/src/physical_plan/sorts/sort_preserving_merge.rs Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org> * Extract into method * Remove from sort enforcement * Update datafusion/core/src/physical_plan/sorts/merge.rs Co-authored-by: Mustafa Akur <106137913+mustafasrepo@users.noreply.github.com> * Update datafusion/proto/src/physical_plan/mod.rs Co-authored-by: Mustafa Akur <106137913+mustafasrepo@users.noreply.github.com> --------- Co-authored-by: Daniël Heres <daniel.heres@coralogix.com> Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org> Co-authored-by: Mustafa Akur <106137913+mustafasrepo@users.noreply.github.com>
5 participants
Which issue does this PR close?
Closes #6000
Rationale for this change
Without fetch in
SortPreserveMergeExecand instreaming_merge(which is used inSortExec) we effectively don't use the fetch anymore in sort besides limiting the amount of rows in batches.This used to work before, as the
fetchwas also used to exit early whenever the limit has been reached.No regressions in the
sortbench:What changes are included in this PR?
Are these changes tested?
Covered by existing tests
Are there any user-facing changes?