Remove the PhysicalSortExpr restriction on union get meet

[berkaysynnada]

Which issue does this PR close?

Closes #6272 .

Rationale for this change

Union get meet can only consider sort option and corresponding vector index during comparison. No need to check for PhysicalExpr equality.

What changes are included in this PR?

If the PhysicalSortExpr's can be downcasted to Column, compare the Column's indexes and sort options. Otherwise, the existing comparison is sustained.

Are these changes tested?

Existing tests are sufficient.

Are there any user-facing changes?

[mingmwang]

Nice optimization

[mingmwang]

@berkaysynnada

Can you extend the optimization to non-column cases?

For example

select a + b from table_2 order by a + b
union all
select x + y from table_2  order by x + y

[berkaysynnada]

@berkaysynnada

Can you extend the optimization to non-column cases?

For example

select a + b from table_2 order by a + b
union all
select x + y from table_2  order by x + y

My last commit can succeed in the cases of your example. However, for the cases like a+b+c (having two or more depth), the build order of the binary expression is important. We can get false negative results. Actually, I think this is a general problem with binary expressions in Datafusion.

[alamb]

Is there and SQL / plan level test that could be written that would cover this code? It seems like a neat optimization but I don't fully understand how to map it to how it would affect queries

[berkaysynnada]

Is there and SQL / plan level test that could be written that would cover this code? It seems like a neat optimization but I don't fully understand how to map it to how it would affect queries

c1 and c1a columns are provided as ascending, and this PR optimizes the plan such:

query TT
explain
SELECT c1 FROM(
(   
    SELECT c1 FROM t1
)  
UNION ALL
(   
    SELECT c1a FROM t2
))
ORDER BY c1
----
logical_plan
  Sort: t1.c1 ASC NULLS LAST
    Union
      TableScan: t1 projection=[c1]
      Projection: t2.c1a AS t1.c1
        TableScan: t2 projection=[c1a]
physical_plan
  SortPreservingMergeExec: [c1@0 ASC NULLS LAST]
    UnionExec
      CsvExec: file_groups={1 group: [[WORKSPACE_ROOT/testing/data/csv/aggregate_test_100.csv]]}, projection=[c1], output_ordering=[c1@0 ASC NULLS LAST], has_header=true
        ProjectionExec: expr=[c1a@0 as t1.c1]
      CsvExec: file_groups={1 group: [[WORKSPACE_ROOT/testing/data/csv/aggregate_test_100.csv]]}, projection=[c1a], output_ordering=[c1a@0 ASC NULLS LAST], has_header=true

However, the main version's result is:

logical_plan
  Sort: t1.c1 ASC NULLS LAST
    Union
      TableScan: t1 projection=[c1]
      Projection: t2.c1a AS t1.c1
        TableScan: t2 projection=[c1a]
physical_plan
  SortPreservingMergeExec: [c1@0 ASC NULLS LAST]
    UnionExec
      CsvExec: file_groups={1 group: [[WORKSPACE_ROOT/testing/data/csv/aggregate_test_100.csv]]}, projection=[c1], output_ordering=[c1@0 ASC NULLS LAST], has_header=true
      SortExec: expr=[c1@0 ASC NULLS LAST]
        ProjectionExec: expr=[c1a@0 as t1.c1]
          CsvExec: file_groups={1 group: [[WORKSPACE_ROOT/testing/data/csv/aggregate_test_100.csv]]}, projection=[c1a], output_ordering=[c1a@0 ASC NULLS LAST], has_header=true

[alamb]

This looks like a good improvement to me. Thank you @berkaysynnada

@ozankabak and/ or @mustafasrepo do you have time to review this PR (as I think you authored the code originally and most recently modified it): https://github.com/apache/arrow-datafusion/blame/894beb1d3b2b42ecc58c582f98e17800ac2b9527/datafusion/core/src/physical_plan/common.rs#L333

[ozankabak]

I went over the code and checked in a commit simplifying it a little bit.

It looks good to me at this point, feel free to merge when you like.

[alamb]

Thanks @berkaysynnada and @ozankabak

[alamb]

For some reason github claims this PR has conflicts that must be resolved 🤔 since I can't push changes to your branch to resolve this myself, @ozankabak or @berkaysynnada can you please resolve the problem? I will then merge it

[ozankabak]

Done, thank you

[mingmwang]

LGTM