Trigger the test #33
Closed
Trigger the test #33
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Test build #728735244 has started for PR 33 at commit |
HyukjinKwon
force-pushed
the
master
branch
2 times, most recently
from
f92d52c to
036dc21
Compare
|
Test build #728856288 for PR 33 at commit |
|
Test build #728856288 for PR 33 at commit |
|
Test build #728888674 for PR 33 at commit |
HyukjinKwon
added a commit
to apache/spark
that referenced
this pull request
Apr 14, 2021
…sitories to share the resources
### What changes were proposed in this pull request?
This PR proposes to leverage the GitHub Actions resources from the forked repositories instead of using the resources in ASF organisation at GitHub.
This is how it works:
1. "Build and test" (`build_and_test.yml`) triggers a build on any commit on any branch (except `branch-*.*`), which roughly means:
- The original repository will trigger the build on any commits in `master` branch
- The forked repository will trigger the build on any commit in any branch.
2. The build triggered in the forked repository will checkout the original repository's `master` branch locally, and merge the branch from the forked repository into the original repository's `master` branch locally.
Therefore, the tests in the forked repository will run after being sync'ed with the original repository's `master` branch.
3. In the original repository, it triggers a workflow that detects the workflow triggered in the forked repository, and add a comment, to the PR, pointing out the workflow in forked repository.
In short, please see this example HyukjinKwon#34
1. You create a PR and your repository triggers the workflow. Your PR uses the resources allocated to you for testing.
2. Apache Spark repository finds your workflow, and links it in a comment in your PR
**NOTE** that we will still run the tests in the original repository for each commit pushed to `master` branch. This distributes the workflows only in PRs.
### Why are the changes needed?
ASF shares the resources across all the ASF projects, which makes the development slow down.
Please see also:
- Discussion in the buildsa.o mailing list: https://lists.apache.org/x/thread.html/r48d079eeff292254db22705c8ef8618f87ff7adc68d56c4e5d0b4105%3Cbuilds.apache.org%3E
- Infra ticket: https://issues.apache.org/jira/browse/INFRA-21646
By distributing the workflows to use author's resources, we can get around this issue.
### Does this PR introduce _any_ user-facing change?
No, this is a dev-only change.
### How was this patch tested?
Manually tested at HyukjinKwon#34 and HyukjinKwon#33.
Closes #32092 from HyukjinKwon/poc-fork-resources.
Lead-authored-by: HyukjinKwon <gurwls223@apache.org>
Co-authored-by: Hyukjin Kwon <gurwls223@apache.org>
Signed-off-by: HyukjinKwon <gurwls223@apache.org>
HyukjinKwon
pushed a commit
that referenced
this pull request
May 13, 2021
### What changes were proposed in this pull request?
As title. This PR is to add code-gen support for LEFT SEMI sort merge join. The main change is to add `semiJoin` code path in `SortMergeJoinExec.doProduce()` and introduce `onlyBufferFirstMatchedRow` in `SortMergeJoinExec.genScanner()`. The latter is for left semi sort merge join without condition. For this kind of query, we don't need to buffer all matched rows, but only the first one (this is same as non-code-gen code path).
Example query:
```
val df1 = spark.range(10).select($"id".as("k1"))
val df2 = spark.range(4).select($"id".as("k2"))
val oneJoinDF = df1.join(df2.hint("SHUFFLE_MERGE"), $"k1" === $"k2", "left_semi")
```
Example of generated code for the query:
```
== Subtree 5 / 5 (maxMethodCodeSize:302; maxConstantPoolSize:156(0.24% used); numInnerClasses:0) ==
*(5) Project [id#0L AS k1#2L]
+- *(5) SortMergeJoin [id#0L], [k2#6L], LeftSemi
:- *(2) Sort [id#0L ASC NULLS FIRST], false, 0
: +- Exchange hashpartitioning(id#0L, 5), ENSURE_REQUIREMENTS, [id=#27]
: +- *(1) Range (0, 10, step=1, splits=2)
+- *(4) Sort [k2#6L ASC NULLS FIRST], false, 0
+- Exchange hashpartitioning(k2#6L, 5), ENSURE_REQUIREMENTS, [id=#33]
+- *(3) Project [id#4L AS k2#6L]
+- *(3) Range (0, 4, step=1, splits=2)
Generated code:
/* 001 */ public Object generate(Object[] references) {
/* 002 */ return new GeneratedIteratorForCodegenStage5(references);
/* 003 */ }
/* 004 */
/* 005 */ // codegenStageId=5
/* 006 */ final class GeneratedIteratorForCodegenStage5 extends org.apache.spark.sql.execution.BufferedRowIterator {
/* 007 */ private Object[] references;
/* 008 */ private scala.collection.Iterator[] inputs;
/* 009 */ private scala.collection.Iterator smj_streamedInput_0;
/* 010 */ private scala.collection.Iterator smj_bufferedInput_0;
/* 011 */ private InternalRow smj_streamedRow_0;
/* 012 */ private InternalRow smj_bufferedRow_0;
/* 013 */ private long smj_value_2;
/* 014 */ private org.apache.spark.sql.execution.ExternalAppendOnlyUnsafeRowArray smj_matches_0;
/* 015 */ private long smj_value_3;
/* 016 */ private org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter[] smj_mutableStateArray_0 = new org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter[2];
/* 017 */
/* 018 */ public GeneratedIteratorForCodegenStage5(Object[] references) {
/* 019 */ this.references = references;
/* 020 */ }
/* 021 */
/* 022 */ public void init(int index, scala.collection.Iterator[] inputs) {
/* 023 */ partitionIndex = index;
/* 024 */ this.inputs = inputs;
/* 025 */ smj_streamedInput_0 = inputs[0];
/* 026 */ smj_bufferedInput_0 = inputs[1];
/* 027 */
/* 028 */ smj_matches_0 = new org.apache.spark.sql.execution.ExternalAppendOnlyUnsafeRowArray(1, 2147483647);
/* 029 */ smj_mutableStateArray_0[0] = new org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter(1, 0);
/* 030 */ smj_mutableStateArray_0[1] = new org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter(1, 0);
/* 031 */
/* 032 */ }
/* 033 */
/* 034 */ private boolean findNextJoinRows(
/* 035 */ scala.collection.Iterator streamedIter,
/* 036 */ scala.collection.Iterator bufferedIter) {
/* 037 */ smj_streamedRow_0 = null;
/* 038 */ int comp = 0;
/* 039 */ while (smj_streamedRow_0 == null) {
/* 040 */ if (!streamedIter.hasNext()) return false;
/* 041 */ smj_streamedRow_0 = (InternalRow) streamedIter.next();
/* 042 */ long smj_value_0 = smj_streamedRow_0.getLong(0);
/* 043 */ if (false) {
/* 044 */ smj_streamedRow_0 = null;
/* 045 */ continue;
/* 046 */
/* 047 */ }
/* 048 */ if (!smj_matches_0.isEmpty()) {
/* 049 */ comp = 0;
/* 050 */ if (comp == 0) {
/* 051 */ comp = (smj_value_0 > smj_value_3 ? 1 : smj_value_0 < smj_value_3 ? -1 : 0);
/* 052 */ }
/* 053 */
/* 054 */ if (comp == 0) {
/* 055 */ return true;
/* 056 */ }
/* 057 */ smj_matches_0.clear();
/* 058 */ }
/* 059 */
/* 060 */ do {
/* 061 */ if (smj_bufferedRow_0 == null) {
/* 062 */ if (!bufferedIter.hasNext()) {
/* 063 */ smj_value_3 = smj_value_0;
/* 064 */ return !smj_matches_0.isEmpty();
/* 065 */ }
/* 066 */ smj_bufferedRow_0 = (InternalRow) bufferedIter.next();
/* 067 */ long smj_value_1 = smj_bufferedRow_0.getLong(0);
/* 068 */ if (false) {
/* 069 */ smj_bufferedRow_0 = null;
/* 070 */ continue;
/* 071 */ }
/* 072 */ smj_value_2 = smj_value_1;
/* 073 */ }
/* 074 */
/* 075 */ comp = 0;
/* 076 */ if (comp == 0) {
/* 077 */ comp = (smj_value_0 > smj_value_2 ? 1 : smj_value_0 < smj_value_2 ? -1 : 0);
/* 078 */ }
/* 079 */
/* 080 */ if (comp > 0) {
/* 081 */ smj_bufferedRow_0 = null;
/* 082 */ } else if (comp < 0) {
/* 083 */ if (!smj_matches_0.isEmpty()) {
/* 084 */ smj_value_3 = smj_value_0;
/* 085 */ return true;
/* 086 */ } else {
/* 087 */ smj_streamedRow_0 = null;
/* 088 */ }
/* 089 */ } else {
/* 090 */ if (smj_matches_0.isEmpty()) {
/* 091 */ smj_matches_0.add((UnsafeRow) smj_bufferedRow_0);
/* 092 */ }
/* 093 */
/* 094 */ smj_bufferedRow_0 = null;
/* 095 */ }
/* 096 */ } while (smj_streamedRow_0 != null);
/* 097 */ }
/* 098 */ return false; // unreachable
/* 099 */ }
/* 100 */
/* 101 */ protected void processNext() throws java.io.IOException {
/* 102 */ while (findNextJoinRows(smj_streamedInput_0, smj_bufferedInput_0)) {
/* 103 */ long smj_value_4 = -1L;
/* 104 */ smj_value_4 = smj_streamedRow_0.getLong(0);
/* 105 */ scala.collection.Iterator<UnsafeRow> smj_iterator_0 = smj_matches_0.generateIterator();
/* 106 */ boolean smj_hasOutputRow_0 = false;
/* 107 */
/* 108 */ while (!smj_hasOutputRow_0 && smj_iterator_0.hasNext()) {
/* 109 */ InternalRow smj_bufferedRow_1 = (InternalRow) smj_iterator_0.next();
/* 110 */
/* 111 */ smj_hasOutputRow_0 = true;
/* 112 */ ((org.apache.spark.sql.execution.metric.SQLMetric) references[0] /* numOutputRows */).add(1);
/* 113 */
/* 114 */ // common sub-expressions
/* 115 */
/* 116 */ smj_mutableStateArray_0[1].reset();
/* 117 */
/* 118 */ smj_mutableStateArray_0[1].write(0, smj_value_4);
/* 119 */ append((smj_mutableStateArray_0[1].getRow()).copy());
/* 120 */
/* 121 */ }
/* 122 */ if (shouldStop()) return;
/* 123 */ }
/* 124 */ ((org.apache.spark.sql.execution.joins.SortMergeJoinExec) references[1] /* plan */).cleanupResources();
/* 125 */ }
/* 126 */
/* 127 */ }
```
### Why are the changes needed?
Improve query CPU performance. Test with one query:
```
def sortMergeJoin(): Unit = {
val N = 2 << 20
codegenBenchmark("left semi sort merge join", N) {
val df1 = spark.range(N).selectExpr(s"id * 2 as k1")
val df2 = spark.range(N).selectExpr(s"id * 3 as k2")
val df = df1.join(df2, col("k1") === col("k2"), "left_semi")
assert(df.queryExecution.sparkPlan.find(_.isInstanceOf[SortMergeJoinExec]).isDefined)
df.noop()
}
}
```
Seeing 30% of run-time improvement:
```
Running benchmark: left semi sort merge join
Running case: left semi sort merge join code-gen off
Stopped after 2 iterations, 1369 ms
Running case: left semi sort merge join code-gen on
Stopped after 5 iterations, 2743 ms
Java HotSpot(TM) 64-Bit Server VM 1.8.0_181-b13 on Mac OS X 10.16
Intel(R) Core(TM) i9-9980HK CPU 2.40GHz
left semi sort merge join: Best Time(ms) Avg Time(ms) Stdev(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------------------------------
left semi sort merge join code-gen off 676 685 13 3.1 322.2 1.0X
left semi sort merge join code-gen on 524 549 32 4.0 249.7 1.3X
```
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Added unit test in `WholeStageCodegenSuite.scala` and `ExistenceJoinSuite.scala`.
Closes apache#32528 from c21/smj-left-semi.
Authored-by: Cheng Su <chengsu@fb.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
HyukjinKwon
pushed a commit
that referenced
this pull request
May 18, 2021
### What changes were proposed in this pull request?
As title. This PR is to add code-gen support for LEFT ANTI sort merge join. The main change is to extract `loadStreamed` in `SortMergeJoinExec.doProduce()`. That is to set all columns values for streamed row, when the streamed row has no output row.
Example query:
```
val df1 = spark.range(10).select($"id".as("k1"))
val df2 = spark.range(4).select($"id".as("k2"))
df1.join(df2.hint("SHUFFLE_MERGE"), $"k1" === $"k2", "left_anti")
```
Example generated code:
```
== Subtree 5 / 5 (maxMethodCodeSize:296; maxConstantPoolSize:156(0.24% used); numInnerClasses:0) ==
*(5) Project [id#0L AS k1#2L]
+- *(5) SortMergeJoin [id#0L], [k2#6L], LeftAnti
:- *(2) Sort [id#0L ASC NULLS FIRST], false, 0
: +- Exchange hashpartitioning(id#0L, 5), ENSURE_REQUIREMENTS, [id=#27]
: +- *(1) Range (0, 10, step=1, splits=2)
+- *(4) Sort [k2#6L ASC NULLS FIRST], false, 0
+- Exchange hashpartitioning(k2#6L, 5), ENSURE_REQUIREMENTS, [id=#33]
+- *(3) Project [id#4L AS k2#6L]
+- *(3) Range (0, 4, step=1, splits=2)
Generated code:
/* 001 */ public Object generate(Object[] references) {
/* 002 */ return new GeneratedIteratorForCodegenStage5(references);
/* 003 */ }
/* 004 */
/* 005 */ // codegenStageId=5
/* 006 */ final class GeneratedIteratorForCodegenStage5 extends org.apache.spark.sql.execution.BufferedRowIterator {
/* 007 */ private Object[] references;
/* 008 */ private scala.collection.Iterator[] inputs;
/* 009 */ private scala.collection.Iterator smj_streamedInput_0;
/* 010 */ private scala.collection.Iterator smj_bufferedInput_0;
/* 011 */ private InternalRow smj_streamedRow_0;
/* 012 */ private InternalRow smj_bufferedRow_0;
/* 013 */ private long smj_value_2;
/* 014 */ private org.apache.spark.sql.execution.ExternalAppendOnlyUnsafeRowArray smj_matches_0;
/* 015 */ private long smj_value_3;
/* 016 */ private org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter[] smj_mutableStateArray_0 = new org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter[2];
/* 017 */
/* 018 */ public GeneratedIteratorForCodegenStage5(Object[] references) {
/* 019 */ this.references = references;
/* 020 */ }
/* 021 */
/* 022 */ public void init(int index, scala.collection.Iterator[] inputs) {
/* 023 */ partitionIndex = index;
/* 024 */ this.inputs = inputs;
/* 025 */ smj_streamedInput_0 = inputs[0];
/* 026 */ smj_bufferedInput_0 = inputs[1];
/* 027 */
/* 028 */ smj_matches_0 = new org.apache.spark.sql.execution.ExternalAppendOnlyUnsafeRowArray(1, 2147483647);
/* 029 */ smj_mutableStateArray_0[0] = new org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter(1, 0);
/* 030 */ smj_mutableStateArray_0[1] = new org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter(1, 0);
/* 031 */
/* 032 */ }
/* 033 */
/* 034 */ private boolean findNextJoinRows(
/* 035 */ scala.collection.Iterator streamedIter,
/* 036 */ scala.collection.Iterator bufferedIter) {
/* 037 */ smj_streamedRow_0 = null;
/* 038 */ int comp = 0;
/* 039 */ while (smj_streamedRow_0 == null) {
/* 040 */ if (!streamedIter.hasNext()) return false;
/* 041 */ smj_streamedRow_0 = (InternalRow) streamedIter.next();
/* 042 */ long smj_value_0 = smj_streamedRow_0.getLong(0);
/* 043 */ if (false) {
/* 044 */ if (!smj_matches_0.isEmpty()) {
/* 045 */ smj_matches_0.clear();
/* 046 */ }
/* 047 */ return false;
/* 048 */
/* 049 */ }
/* 050 */ if (!smj_matches_0.isEmpty()) {
/* 051 */ comp = 0;
/* 052 */ if (comp == 0) {
/* 053 */ comp = (smj_value_0 > smj_value_3 ? 1 : smj_value_0 < smj_value_3 ? -1 : 0);
/* 054 */ }
/* 055 */
/* 056 */ if (comp == 0) {
/* 057 */ return true;
/* 058 */ }
/* 059 */ smj_matches_0.clear();
/* 060 */ }
/* 061 */
/* 062 */ do {
/* 063 */ if (smj_bufferedRow_0 == null) {
/* 064 */ if (!bufferedIter.hasNext()) {
/* 065 */ smj_value_3 = smj_value_0;
/* 066 */ return !smj_matches_0.isEmpty();
/* 067 */ }
/* 068 */ smj_bufferedRow_0 = (InternalRow) bufferedIter.next();
/* 069 */ long smj_value_1 = smj_bufferedRow_0.getLong(0);
/* 070 */ if (false) {
/* 071 */ smj_bufferedRow_0 = null;
/* 072 */ continue;
/* 073 */ }
/* 074 */ smj_value_2 = smj_value_1;
/* 075 */ }
/* 076 */
/* 077 */ comp = 0;
/* 078 */ if (comp == 0) {
/* 079 */ comp = (smj_value_0 > smj_value_2 ? 1 : smj_value_0 < smj_value_2 ? -1 : 0);
/* 080 */ }
/* 081 */
/* 082 */ if (comp > 0) {
/* 083 */ smj_bufferedRow_0 = null;
/* 084 */ } else if (comp < 0) {
/* 085 */ if (!smj_matches_0.isEmpty()) {
/* 086 */ smj_value_3 = smj_value_0;
/* 087 */ return true;
/* 088 */ } else {
/* 089 */ return false;
/* 090 */ }
/* 091 */ } else {
/* 092 */ if (smj_matches_0.isEmpty()) {
/* 093 */ smj_matches_0.add((UnsafeRow) smj_bufferedRow_0);
/* 094 */ }
/* 095 */
/* 096 */ smj_bufferedRow_0 = null;
/* 097 */ }
/* 098 */ } while (smj_streamedRow_0 != null);
/* 099 */ }
/* 100 */ return false; // unreachable
/* 101 */ }
/* 102 */
/* 103 */ protected void processNext() throws java.io.IOException {
/* 104 */ while (smj_streamedInput_0.hasNext()) {
/* 105 */ findNextJoinRows(smj_streamedInput_0, smj_bufferedInput_0);
/* 106 */
/* 107 */ long smj_value_4 = -1L;
/* 108 */ smj_value_4 = smj_streamedRow_0.getLong(0);
/* 109 */ scala.collection.Iterator<UnsafeRow> smj_iterator_0 = smj_matches_0.generateIterator();
/* 110 */
/* 111 */ boolean wholestagecodegen_hasOutputRow_0 = false;
/* 112 */
/* 113 */ while (!wholestagecodegen_hasOutputRow_0 && smj_iterator_0.hasNext()) {
/* 114 */ InternalRow smj_bufferedRow_1 = (InternalRow) smj_iterator_0.next();
/* 115 */
/* 116 */ wholestagecodegen_hasOutputRow_0 = true;
/* 117 */ }
/* 118 */
/* 119 */ if (!wholestagecodegen_hasOutputRow_0) {
/* 120 */ // load all values of streamed row, because the values not in join condition are not
/* 121 */ // loaded yet.
/* 122 */
/* 123 */ ((org.apache.spark.sql.execution.metric.SQLMetric) references[0] /* numOutputRows */).add(1);
/* 124 */
/* 125 */ // common sub-expressions
/* 126 */
/* 127 */ smj_mutableStateArray_0[1].reset();
/* 128 */
/* 129 */ smj_mutableStateArray_0[1].write(0, smj_value_4);
/* 130 */ append((smj_mutableStateArray_0[1].getRow()).copy());
/* 131 */
/* 132 */ }
/* 133 */ if (shouldStop()) return;
/* 134 */ }
/* 135 */ ((org.apache.spark.sql.execution.joins.SortMergeJoinExec) references[1] /* plan */).cleanupResources();
/* 136 */ }
/* 137 */
/* 138 */ }
```
### Why are the changes needed?
Improve the query CPU performance.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
Added unit test in `WholeStageCodegenSuite.scala`, and existed unit test in `ExistenceJoinSuite.scala`.
Closes apache#32547 from c21/smj-left-anti.
Authored-by: Cheng Su <chengsu@fb.com>
Signed-off-by: Takeshi Yamamuro <yamamuro@apache.org>
HyukjinKwon
pushed a commit
that referenced
this pull request
Aug 5, 2022
### What changes were proposed in this pull request?
Optimize the TransposeWindow rule to extend applicable cases and optimize time complexity.
TransposeWindow rule will try to eliminate unnecessary shuffle:
but the function compatiblePartitions will only take the first n elements of the window2 partition sequence, for some cases, this will not take effect, like the case below:
val df = spark.range(10).selectExpr("id AS a", "id AS b", "id AS c", "id AS d")
df.selectExpr(
"sum(`d`) OVER(PARTITION BY `b`,`a`) as e",
"sum(`c`) OVER(PARTITION BY `a`) as f"
).explain
Current plan
== Physical Plan ==
*(5) Project [e#10L, f#11L]
+- Window [sum(c#4L) windowspecdefinition(a#2L, specifiedwindowframe(RowFrame, unboundedpreceding$(), unboundedfollowing$())) AS f#11L], [a#2L]
+- *(4) Sort [a#2L ASC NULLS FIRST], false, 0
+- Exchange hashpartitioning(a#2L, 200), true, [id=#41]
+- *(3) Project [a#2L, c#4L, e#10L]
+- Window [sum(d#5L) windowspecdefinition(b#3L, a#2L, specifiedwindowframe(RowFrame, unboundedpreceding$(), unboundedfollowing$())) AS e#10L], [b#3L, a#2L]
+- *(2) Sort [b#3L ASC NULLS FIRST, a#2L ASC NULLS FIRST], false, 0
+- Exchange hashpartitioning(b#3L, a#2L, 200), true, [id=#33]
+- *(1) Project [id#0L AS d#5L, id#0L AS b#3L, id#0L AS a#2L, id#0L AS c#4L]
+- *(1) Range (0, 10, step=1, splits=10)
Expected plan:
== Physical Plan ==
*(4) Project [e#924L, f#925L]
+- Window [sum(d#43L) windowspecdefinition(b#41L, a#40L, specifiedwindowframe(RowFrame, unboundedpreceding$(), unboundedfollowing$())) AS e#924L], [b#41L, a#40L]
+- *(3) Sort [b#41L ASC NULLS FIRST, a#40L ASC NULLS FIRST], false, 0
+- *(3) Project [d#43L, b#41L, a#40L, f#925L]
+- Window [sum(c#42L) windowspecdefinition(a#40L, specifiedwindowframe(RowFrame, unboundedpreceding$(), unboundedfollowing$())) AS f#925L], [a#40L]
+- *(2) Sort [a#40L ASC NULLS FIRST], false, 0
+- Exchange hashpartitioning(a#40L, 200), true, [id=apache#282]
+- *(1) Project [id#38L AS d#43L, id#38L AS b#41L, id#38L AS a#40L, id#38L AS c#42L]
+- *(1) Range (0, 10, step=1, splits=10)
Also the permutations method has a O(n!) time complexity, which is very expensive when there are many partition columns, we could try to optimize it.
### Why are the changes needed?
We could apply the rule for more cases, which could improve the execution performance by eliminate unnecessary shuffle, and by reducing the time complexity from O(n!) to O(n2), the performance for the rule itself could improve
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
UT
Closes apache#35334 from constzhou/SPARK-38034_optimize_transpose_window_rule.
Authored-by: xzhou <15210830305@163.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
HyukjinKwon
pushed a commit
that referenced
this pull request
Aug 29, 2022
### What changes were proposed in this pull request?
Optimize the TransposeWindow rule to extend applicable cases and optimize time complexity.
TransposeWindow rule will try to eliminate unnecessary shuffle:
but the function compatiblePartitions will only take the first n elements of the window2 partition sequence, for some cases, this will not take effect, like the case below:
val df = spark.range(10).selectExpr("id AS a", "id AS b", "id AS c", "id AS d")
df.selectExpr(
"sum(`d`) OVER(PARTITION BY `b`,`a`) as e",
"sum(`c`) OVER(PARTITION BY `a`) as f"
).explain
Current plan
== Physical Plan ==
*(5) Project [e#10L, f#11L]
+- Window [sum(c#4L) windowspecdefinition(a#2L, specifiedwindowframe(RowFrame, unboundedpreceding$(), unboundedfollowing$())) AS f#11L], [a#2L]
+- *(4) Sort [a#2L ASC NULLS FIRST], false, 0
+- Exchange hashpartitioning(a#2L, 200), true, [id=#41]
+- *(3) Project [a#2L, c#4L, e#10L]
+- Window [sum(d#5L) windowspecdefinition(b#3L, a#2L, specifiedwindowframe(RowFrame, unboundedpreceding$(), unboundedfollowing$())) AS e#10L], [b#3L, a#2L]
+- *(2) Sort [b#3L ASC NULLS FIRST, a#2L ASC NULLS FIRST], false, 0
+- Exchange hashpartitioning(b#3L, a#2L, 200), true, [id=#33]
+- *(1) Project [id#0L AS d#5L, id#0L AS b#3L, id#0L AS a#2L, id#0L AS c#4L]
+- *(1) Range (0, 10, step=1, splits=10)
Expected plan:
== Physical Plan ==
*(4) Project [e#924L, f#925L]
+- Window [sum(d#43L) windowspecdefinition(b#41L, a#40L, specifiedwindowframe(RowFrame, unboundedpreceding$(), unboundedfollowing$())) AS e#924L], [b#41L, a#40L]
+- *(3) Sort [b#41L ASC NULLS FIRST, a#40L ASC NULLS FIRST], false, 0
+- *(3) Project [d#43L, b#41L, a#40L, f#925L]
+- Window [sum(c#42L) windowspecdefinition(a#40L, specifiedwindowframe(RowFrame, unboundedpreceding$(), unboundedfollowing$())) AS f#925L], [a#40L]
+- *(2) Sort [a#40L ASC NULLS FIRST], false, 0
+- Exchange hashpartitioning(a#40L, 200), true, [id=apache#282]
+- *(1) Project [id#38L AS d#43L, id#38L AS b#41L, id#38L AS a#40L, id#38L AS c#42L]
+- *(1) Range (0, 10, step=1, splits=10)
Also the permutations method has a O(n!) time complexity, which is very expensive when there are many partition columns, we could try to optimize it.
### Why are the changes needed?
We could apply the rule for more cases, which could improve the execution performance by eliminate unnecessary shuffle, and by reducing the time complexity from O(n!) to O(n2), the performance for the rule itself could improve
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
UT
Closes apache#35334 from constzhou/SPARK-38034_optimize_transpose_window_rule.
Authored-by: xzhou <15210830305@163.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
(cherry picked from commit 0cc331d)
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.