Support non-tuple expression for in-subquery to join by ygf11 · Pull Request #4826 · apache/datafusion

ygf11 · 2023-01-05T07:33:23Z

Which issue does this PR close?

Closes #4724.

Rationale for this change

This sql works in datafusion:

> select * from t1 where t1_id in (select t2_id from t2);

but following sql does not work:

❯ select * from t1 where t1_id + 11 in (select t2_id from t2);
NotImplemented("Physical plan does not support logical expression CAST(t1.t1_id AS Int64) + Int64(11) IN (<subquery>)")

❯ select * from t1 where t1_id  in (select t2_id + 11 from t2);
NotImplemented("Physical plan does not support logical expression t1.t1_id IN (<subquery>)")

❯ select * from t1 where t1_id + 12  in (select t2_id + 1 from t2);
NotImplemented("Physical plan does not support logical expression CAST(t1.t1_id AS Int64) + Int64(12) IN (<subquery>)")

What changes are included in this PR?

Rewrite optimize_where_in in decorrelate_where_in.rs.
Fix tests.

Are these changes tested?

Yes.

Are there any user-facing changes?

ygf11 · 2023-01-07T10:07:26Z

Please take a look, @alamb @jackwener.

ygf11 · 2023-01-07T12:47:40Z

datafusion/optimizer/src/decorrelate_where_in.rs

+                subquery_filters.push(expr.clone());
+            } else {
+                join_filters.push(expr.clone())
+            }


If the expression depends on the outer table, we should move it to the join predicate.

Yeah -- this is the decorrelation logic

ygf11 · 2023-01-07T12:56:39Z

datafusion/optimizer/src/decorrelate_where_in.rs

-        .project(projection)?
-        .alias(&subqry_alias)?
+    let subquery_expr_name = format!("{:?}", unnormalize_col(subquery_expr.clone()));
+    let first_expr = subquery_expr.clone().alias(subquery_expr_name.clone());


For the right side of the in predicate, it looks strange if we show the original qualify name in the plan, so the qualify name is stripped.

alamb · 2023-01-07T15:14:08Z

I plan to review this tomorrow. THank you @ygf11

jackwener

I prepare to review this PR more carefully tonight or tomorrow.

jackwener · 2023-01-09T04:26:20Z

datafusion/core/tests/sql/joins.rs

+    let expected = vec![
+        "+-------+---------+--------+",
+        "| t1_id | t1_name | t1_int |",
+        "+-------+---------+--------+",
+        "| 11    | a       | 1      |",
+        "| 33    | c       | 3      |",
+        "| 44    | d       | 4      |",
+        "+-------+---------+--------+",
+    ];
+
+    let results = execute_to_batches(&ctx, sql).await;
+    assert_batches_sorted_eq!(expected, &results);


Now we already use sqllogicaltest to support Data Driven Tests #4460.
I suggest add a file join.slt | subquery.slt to cover some case of join and subquery.

In UT and integration-test, we just focus on correctness of Plan, data-test is derived by sqllogicaltest

File a ticket #4870.

jackwener · 2023-01-09T04:34:51Z

datafusion/optimizer/src/decorrelate_where_in.rs

+        let expected = "Projection: customer.c_custkey [c_custkey:Int64]\
+        \n  LeftSemi Join:  Filter: customer.c_custkey + Int32(1) = __correlated_sq_1.o_custkey AND customer.c_custkey = __correlated_sq_1.o_custkey [c_custkey:Int64, c_name:Utf8]\
+        \n    TableScan: customer [c_custkey:Int64, c_name:Utf8]\
+        \n    SubqueryAlias: __correlated_sq_1 [o_custkey:Int64]\
+        \n      Projection: orders.o_custkey AS o_custkey [o_custkey:Int64]\
+        \n        TableScan: orders [o_orderkey:Int64, o_custkey:Int64, o_orderstatus:Utf8, o_totalprice:Float64;N]";
+
+        assert_optimized_plan_eq_display_indent(
            Arc::new(DecorrelateWhereIn::new()),
            &plan,
-            "column comparison required",
+            expected,


alamb · 2023-01-10T22:56:32Z

Sorry for the delay in my review. I am traveling this week at a conference so I don't have as much time to devote to reviews and merging as normal

alamb

This is looking good @ygf11 -- thank you. 👌

The only question I had was about queries that had two subqueries and it appears the predicate filter has been pulled up too far. Otherwise this one is looking

alamb · 2023-01-11T21:10:23Z

datafusion/core/tests/sql/joins.rs

+
+    let expected = vec![
+        "Explain [plan_type:Utf8, plan:Utf8]",
+        "  Projection: t1.t1_id, t1.t1_name, t1.t1_int [t1_id:UInt32;N, t1_name:Utf8;N, t1_int:UInt32;N]",


👍 the plan looks good

alamb · 2023-01-11T21:13:00Z

datafusion/core/tests/sql/subqueries.rs

-    SubqueryAlias: __correlated_sq_1
-      Projection: lineitem.l_linestatus AS l_linestatus, lineitem.l_orderkey AS l_orderkey
-        TableScan: lineitem projection=[l_orderkey, l_linestatus]"#;
+


This change is just whitespace, right?

Not only whitespace, the alias except the first expression are removed in the projection.

before:

SubqueryAlias: __correlated_sq_1 Projection: lineitem.l_linestatus AS l_linestatus, lineitem.l_orderkey AS l_orderkey

after:

\n SubqueryAlias: __correlated_sq_1\ \n Projection: lineitem.l_linestatus AS l_linestatus, lineitem.l_orderkey\

The first expression of the projection is the right side of in-equijoin, it may be a non-column expression, so always give it an alias, but others do not(they are alway column).

alamb · 2023-01-11T21:13:52Z

datafusion/core/tests/sql/joins.rs

+}
+
+#[tokio::test]
+async fn three_projection_exprs_subquery_to_join() -> Result<()> {


As a minor point, I am not sure what the extra coverage the multiple predicates in the where clause are adding

I wonder if we need all these tests?

Added more tests -- multiple predicates in outer and subquery where clause.

alamb · 2023-01-11T21:14:34Z

datafusion/optimizer/src/decorrelate_where_in.rs

    }
 }

+/// Optimize the where in subquery to left-anti/left-semi join.


datafusion/optimizer/src/decorrelate_where_in.rs

alamb · 2023-01-11T21:24:45Z

datafusion/optimizer/src/decorrelate_where_in.rs

-        // Query will fail, but we can still transform the plan
        let expected = "Projection: customer.c_custkey [c_custkey:Int64]\
-        \n  LeftSemi Join: customer.c_custkey = __correlated_sq_1.o_custkey [c_custkey:Int64, c_name:Utf8]\
+        \n  LeftSemi Join:  Filter: customer.c_custkey = __correlated_sq_1.o_custkey AND customer.c_custkey = customer.c_custkey [c_custkey:Int64, c_name:Utf8]\


This looks like somewhat of a regression to me (the filter has been pulled out of the subquery and into the join). Is that intended?

No, I do not handle this special case, just treat customer.c_custkey = customer.c_custkey as join filter.

I think it is better to add this to the filter of outer table, then the plan will be like:

"Projection: customer.c_custkey [c_custkey:Int64]\ LeftSemi Join: customer.c_custkey = __correlated_sq_1.o_custkey [c_custkey:Int64, c_name:Utf8] Filter: customer.c_custkey = customer.c_custkey [..]\ TableScan: customer [c_custkey:Int64, c_name:Utf8]\ SubqueryAlias: __correlated_sq_1 [o_custkey:Int64]\ Projection: orders.o_custkey AS o_custkey [o_custkey:Int64]\ TableScan: orders [o_orderkey:Int64, o_custkey:Int64, o_orderstatus:Utf8, o_totalprice:Float64;N]"

Is this the right direction?

After more thought it probably makes sense to do this decorrelation in the general way (no special case0 as you have done and then we will rely on subsequent predicate pushdown logic to push the filters back down if possible

Yes, it is better to do by other rule.
But I find our predicate pushdown logic does not support these predicates(semi-join-on) now.
https://github.com/apache/arrow-datafusion/blob/dee0dd8be6745f6cb798ba56dca6c1b936d90fd6/datafusion/optimizer/src/push_down_filter.rs#L103-L115

Maybe we should support predicate push down for semi-join-on, or we need move these predicates to the outside of join, then predicate push down can work(Like following).

// for sql: select * from t1 where t1.t1_int in(select t2.t2_int from t2 where t1.t1_id > 10); // the output of this rule: Projection: t1.t1_id, t1.t1_name, t1.t1_int Filter: CAST(t1.t1_id AS Int64) > Int64(10) LeftSemi Join: Filter: t1.t1_int = __correlated_sq_2.t2_int TableScan: t1 SubqueryAlias: __correlated_sq_2 Projection: t2.t2_int AS t2_int TableScan: t2

datafusion/optimizer/src/decorrelate_where_in.rs

alamb · 2023-01-11T21:32:16Z

datafusion/optimizer/src/decorrelate_where_in.rs

+                let using_cols: Vec<Column> = expr
+                    .to_columns()?
+                    .into_iter()
+                    .filter(|col| input_schema.field_from_column(col).is_ok())


doesn't this ignore columns that are not found in the subquery? I wonder if the error should be percolated up rather than ignored? Perhaps test coverage of a query like select ... from t where x in (select non_existent_col from foo) would be good?

Although perhaps if it was an error it signals something is wrong in the plan (as a column reference wouldn't be present in the input schema 🤔 )

If this shouldn't happen, maybe we can collect the errors and return them?

doesn't this ignore columns that are not found in the subquery

Yes, it will ignore. But it is used to collect columns which will be add to subquery projection.

Suppose there is a query:
select * from t where x in (select c from foo where foo.a > t.a and foo.b > t.a)

foo.a > t.a and foo.b > t.a will move to join, to make the join work, we need add foo.a and foo.b to the projection of subquery, the above code does the filter work.

If the where clause references unknown columns, we need do more work here, but I think it should be done in the planner.

alamb · 2023-01-11T21:33:36Z

datafusion/optimizer/src/decorrelate_where_in.rs

+                subquery_filters.push(expr.clone());
+            } else {
+                join_filters.push(expr.clone())
+            }


Yeah -- this is the decorrelation logic

datafusion/optimizer/src/decorrelate_where_in.rs

Co-authored-by: Andrew Lamb <[email protected]>

alamb · 2023-01-14T11:31:28Z

datafusion/optimizer/src/decorrelate_where_in.rs

-        // Query will fail, but we can still transform the plan
        let expected = "Projection: customer.c_custkey [c_custkey:Int64]\
-        \n  LeftSemi Join: customer.c_custkey = __correlated_sq_1.o_custkey [c_custkey:Int64, c_name:Utf8]\
+        \n  LeftSemi Join:  Filter: customer.c_custkey = __correlated_sq_1.o_custkey AND customer.c_custkey = customer.c_custkey [c_custkey:Int64, c_name:Utf8]\


After more thought it probably makes sense to do this decorrelation in the general way (no special case0 as you have done and then we will rely on subsequent predicate pushdown logic to push the filters back down if possible

ygf11 · 2023-01-15T10:30:57Z

datafusion/optimizer/src/decorrelate_where_in.rs

+        let expected = "Projection: test.b [b:UInt32]\
+        \n  Filter: test.c > UInt32(1) [a:UInt32, b:UInt32, c:UInt32]\
+        \n    LeftSemi Join:  Filter: test.c * UInt32(2) = __correlated_sq_2.c * UInt32(2) AND test.a > __correlated_sq_2.a [a:UInt32, b:UInt32, c:UInt32]\
+        \n      Filter: test.c > UInt32(1) [a:UInt32, b:UInt32, c:UInt32]\


Filter: test.c > UInt32(1) happen twice, it is better to only add once #4914.
I will fix it in the following pr.

ygf11 · 2023-01-15T10:35:17Z

datafusion/core/tests/sql/joins.rs

+        "      SubqueryAlias: __correlated_sq_1 [CAST(t2_id AS Int64) + Int64(1):Int64;N]",
+        "        Projection: CAST(t2.t2_id AS Int64) + Int64(1) AS CAST(t2_id AS Int64) + Int64(1) [CAST(t2_id AS Int64) + Int64(1):Int64;N]",
+        "          TableScan: t2 projection=[t2_id] [t2_id:UInt32;N]",
+    ];


This test case show that the special case t1.t1_int > 0 is not pushed down after optimizing.

Is that something you would like to fix in this PR or is it something you would like to fix in a follow on?

I mean next pr 🤣.

there is a issue: #4413

alamb · 2023-01-15T12:23:52Z

Thanks for sticking with this @ygf11

ursabot · 2023-01-15T12:32:30Z

Benchmark runs are scheduled for baseline = d49c805 and contender = e2daee9. e2daee9 is a master commit associated with this PR. Results will be available as each benchmark for each run completes.
Conbench compare runs links:
[Skipped ⚠️ Benchmarking of arrow-datafusion-commits is not supported on ec2-t3-xlarge-us-east-2] ec2-t3-xlarge-us-east-2
[Skipped ⚠️ Benchmarking of arrow-datafusion-commits is not supported on test-mac-arm] test-mac-arm
[Skipped ⚠️ Benchmarking of arrow-datafusion-commits is not supported on ursa-i9-9960x] ursa-i9-9960x
[Skipped ⚠️ Benchmarking of arrow-datafusion-commits is not supported on ursa-thinkcentre-m75q] ursa-thinkcentre-m75q
Buildkite builds:
Supported benchmarks:
ec2-t3-xlarge-us-east-2: Supported benchmark langs: Python, R. Runs only benchmarks with cloud = True
test-mac-arm: Supported benchmark langs: C++, Python, R
ursa-i9-9960x: Supported benchmark langs: Python, R, JavaScript
ursa-thinkcentre-m75q: Supported benchmark langs: C++, Java

Support non-tuple expression for in-subquery to join

10bafa9

github-actions bot added core Core DataFusion crate logical-expr Logical plan and expressions optimizer Optimizer rules labels Jan 5, 2023

ygf11 added 3 commits January 7, 2023 03:03

add tests

6b929df

add comment and fix cargo fmt

7767bb1

fix comment

de0fdf3

ygf11 marked this pull request as ready for review January 7, 2023 10:06

ygf11 commented Jan 7, 2023

View reviewed changes

ygf11 added 2 commits January 7, 2023 22:18

clean unused comment

d641b51

Merge remote-tracking branch 'upstream/master' into in-subquery

c7b7a4c

jackwener reviewed Jan 9, 2023

View reviewed changes

jackwener self-assigned this Jan 10, 2023

ygf11 mentioned this pull request Jan 11, 2023

Port other sql_integration test cases to sqllogictest #4870

Closed

alamb reviewed Jan 11, 2023

View reviewed changes

ygf11 and others added 4 commits January 13, 2023 16:05

Update datafusion/optimizer/src/decorrelate_where_in.rs

5d25827

Co-authored-by: Andrew Lamb <[email protected]>

Update datafusion/optimizer/src/decorrelate_where_in.rs

e0055c8

Co-authored-by: Andrew Lamb <[email protected]>

Update datafusion/optimizer/src/decorrelate_where_in.rs

08a66bb

Co-authored-by: Andrew Lamb <[email protected]>

fix comment

d4187f8

alamb approved these changes Jan 14, 2023

View reviewed changes

ygf11 added 4 commits January 15, 2023 02:13

resolve merge conflict

f5ff04a

fix cargo fmt

747957a

add tests

96d26bd

fix cargo fmt

f035229

ygf11 commented Jan 15, 2023

View reviewed changes

alamb merged commit e2daee9 into apache:master Jan 15, 2023

ygf11 deleted the in-subquery branch January 16, 2023 06:42

mingmwang mentioned this pull request Jan 17, 2023

Support non-tuple expression for Exists Subquery to Join #4934

Closed

ygf11 mentioned this pull request Jan 28, 2023

Minor: remove unused methods in datafusion/optimizer/src/utils.rs #5098

Merged

Conversation

ygf11 commented Jan 5, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Which issue does this PR close?

Rationale for this change

What changes are included in this PR?

Are these changes tested?

Are there any user-facing changes?

Uh oh!

ygf11 commented Jan 7, 2023

Uh oh!

ygf11 Jan 7, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ygf11 Jan 7, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

alamb commented Jan 7, 2023

Uh oh!

jackwener left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

alamb commented Jan 10, 2023

Uh oh!

alamb left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ygf11 Jan 12, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ygf11 Jan 14, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ygf11 Jan 13, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ygf11 Jan 15, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ygf11 commented Jan 5, 2023 •

edited

Loading

ygf11 Jan 7, 2023 •

edited

Loading

ygf11 Jan 7, 2023 •

edited

Loading

ygf11 Jan 12, 2023 •

edited

Loading

ygf11 Jan 14, 2023 •

edited

Loading

ygf11 Jan 13, 2023 •

edited

Loading

ygf11 Jan 15, 2023 •

edited

Loading