imback82 opened a new pull request #29655:
URL: https://github.com/apache/spark/pull/29655
<!--
Thanks for sending a pull request! Here are some tips for you:
1. If this is your first time, please read our contributor guidelines:
https://spark.apache.org/contributing.html
2. Ensure you have added or run the appropriate tests for your PR:
https://spark.apache.org/developer-tools.html
3. If the PR is unfinished, add '[WIP]' in your PR title, e.g.,
'[WIP][SPARK-XXXX] Your PR title ...'.
4. Be sure to keep the PR description updated to reflect all changes.
5. Please write your PR title to summarize what this PR proposes.
6. If possible, provide a concise example to reproduce the issue for a
faster review.
7. If you want to add a new configuration, please read the guideline first
for naming configurations in
'core/src/main/scala/org/apache/spark/internal/config/ConfigEntry.scala'.
-->
### What changes were proposed in this pull request?
<!--
Please clarify what changes you are proposing. The purpose of this section
is to outline the changes and how this PR fixes the issue.
If possible, please consider writing useful notes for better and faster
reviews in your PR. See the examples below.
1. If you refactor some codes with changing classes, showing the class
hierarchy will help reviewers.
2. If you fix some SQL features, you can provide some references of other
DBMSes.
3. If there is design documentation, please add the link.
4. If there is a discussion in the mailing list, please add the link.
-->
This PR proposes to optimize SortMergeJoin (SMJ) if each of its children has
hash output partitioning which "partially" satisfies the required distribution.
In this case where the child's output partitioning expressions are a subset of
required distribution expressions (join keys expressions), the shuffle can be
removed because rows will be sorted by join keys before rows are joined (the
required child ordering for SMJ is on join keys).
This PR introduces `OptimizeSortMergeJoinWithPartialHashDistribution ` which
removes shuffle for the sort merge join if the following conditions are met:
- The child of ShuffleExchangeExec has HashPartitioning with the same
number of partitions as the other side of join.
- The child of ShuffleExchangeExec has output partitioning which has the
subset of join keys on the respective join side.
This rule can be turned on by setting
`spark.sql.execution.sortMergeJoin.optimizePartialHashDistribution.enabled` to
`true` (`false` by default).
### Why are the changes needed?
<!--
Please clarify why the changes are needed. For instance,
1. If you propose a new API, clarify the use case for a new API.
2. If you fix a bug, you can clarify why it is a bug.
-->
To remove unnecessary shuffles in certain scenarios.
### Does this PR introduce _any_ user-facing change?
<!--
Note that it means *any* user-facing change including all aspects such as
the documentation fix.
If yes, please clarify the previous behavior and the change this PR proposes
- provide the console output, description and/or an example to show the
behavior difference if possible.
If possible, please also clarify if this is a user-facing change compared to
the released Spark versions or within the unreleased branches such as master.
If no, write 'No'.
-->
Suppose the following case where `t1` is bucketed by `i1`, and `t2` by `i2`:
```scala
val df1 = (0 until 100).map(i => (i % 5, i % 13, i.toString)).toDF("i1",
"j1", "k1")
val df2 = (0 until 100).map(i => (i % 3, i % 17, i.toString)).toDF("i2",
"j2", "k2")
df1.write.format("parquet").bucketBy(8, "i1").saveAsTable("t1")
df2.write.format("parquet").bucketBy(8, "i2").saveAsTable("t2")
val t1 = spark.table("t1")
val t2 = spark.table("t2")
```
Now if you join two tables by `t1("i1") === t2("i2") && t1("j1") ===
t2("j2")`
Before this change:
```scala
scala> spark.conf.set("spark.sql.autoBroadcastJoinThreshold", "0")
scala> t1.join(t2, t1("i1") === t2("i2") && t1("j1") === t2("j2")).explain
== Physical Plan ==
*(5) SortMergeJoin [i1#161, j1#162], [i2#167, j2#168], Inner
:- *(2) Sort [i1#161 ASC NULLS FIRST, j1#162 ASC NULLS FIRST], false, 0
: +- Exchange hashpartitioning(i1#161, j1#162, 200), true, [id=#196]
: +- *(1) Filter (isnotnull(i1#161) AND isnotnull(j1#162))
: +- *(1) ColumnarToRow
: +- FileScan parquet default.t1[i1#161,j1#162,k1#163] Batched:
true, DataFilters: [isnotnull(i1#161), isnotnull(j1#162)], Format: Parquet,
Location: InMemoryFileIndex[], PartitionFilters: [], PushedFilters:
[IsNotNull(i1), IsNotNull(j1)], ReadSchema: struct<i1:int,j1:int,k1:string>,
SelectedBucketsCount: 8 out of 8
+- *(4) Sort [i2#167 ASC NULLS FIRST, j2#168 ASC NULLS FIRST], false, 0
+- Exchange hashpartitioning(i2#167, j2#168, 200), true, [id=#205]
+- *(3) Filter (isnotnull(i2#167) AND isnotnull(j2#168))
+- *(3) ColumnarToRow
+- FileScan parquet default.t2[i2#167,j2#168,k2#169] Batched:
true, DataFilters: [isnotnull(i2#167), isnotnull(j2#168)], Format: Parquet,
Location: InMemoryFileIndex[], PartitionFilters: [], PushedFilters:
[IsNotNull(i2), IsNotNull(j2)], ReadSchema: struct<i2:int,j2:int,k2:string>,
SelectedBucketsCount: 8 out of 8
```
After the PR:
```scala
scala>
spark.conf.set("spark.sql.execution.sortMergeJoin.optimizePartialHashDistribution.enabled",
"true")
scala> t1.join(t2, t1("i1") === t2("i2") && t1("j1") === t2("j2")).explain
== Physical Plan ==
*(3) SortMergeJoin [i1#161, j1#162], [i2#167, j2#168], Inner
:- *(1) Sort [i1#161 ASC NULLS FIRST, j1#162 ASC NULLS FIRST], false, 0
: +- *(1) Filter (isnotnull(i1#161) AND isnotnull(j1#162))
: +- *(1) ColumnarToRow
: +- FileScan parquet default.t1[i1#161,j1#162,k1#163] Batched: true,
DataFilters: [isnotnull(i1#161), isnotnull(j1#162)], Format: Parquet, Location:
InMemoryFileIndex[], PartitionFilters: [], PushedFilters: [IsNotNull(i1),
IsNotNull(j1)], ReadSchema: struct<i1:int,j1:int,k1:string>,
SelectedBucketsCount: 8 out of 8
+- *(2) Sort [i2#167 ASC NULLS FIRST, j2#168 ASC NULLS FIRST], false, 0
+- *(2) Filter (isnotnull(i2#167) AND isnotnull(j2#168))
+- *(2) ColumnarToRow
+- FileScan parquet default.t2[i2#167,j2#168,k2#169] Batched: true,
DataFilters: [isnotnull(i2#167), isnotnull(j2#168)], Format: Parquet, Location:
InMemoryFileIndex[], PartitionFilters: [], PushedFilters: [IsNotNull(i2),
IsNotNull(j2)], ReadSchema: struct<i2:int,j2:int,k2:string>,
SelectedBucketsCount: 8 out of 8
```
### How was this patch tested?
<!--
If tests were added, say they were added here. Please make sure to add some
test cases that check the changes thoroughly including negative and positive
cases if possible.
If it was tested in a way different from regular unit tests, please clarify
how you tested step by step, ideally copy and paste-able, so that other
reviewers can test and check, and descendants can verify in the future.
If tests were not added, please describe why they were not added and/or why
it was difficult to add.
-->
Added 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.
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
