/
Optimizer.scala
1304 lines (1180 loc) · 54.6 KB
/
Optimizer.scala
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
/*
* Licensed to the Apache Software Foundation (ASF) under one or more
* contributor license agreements. See the NOTICE file distributed with
* this work for additional information regarding copyright ownership.
* The ASF licenses this file to You under the Apache License, Version 2.0
* (the "License"); you may not use this file except in compliance with
* the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package org.apache.spark.sql.catalyst.optimizer
import scala.collection.mutable
import org.apache.spark.sql.AnalysisException
import org.apache.spark.sql.catalyst.analysis._
import org.apache.spark.sql.catalyst.catalog.{InMemoryCatalog, SessionCatalog}
import org.apache.spark.sql.catalyst.expressions._
import org.apache.spark.sql.catalyst.expressions.aggregate._
import org.apache.spark.sql.catalyst.plans._
import org.apache.spark.sql.catalyst.plans.logical._
import org.apache.spark.sql.catalyst.rules._
import org.apache.spark.sql.internal.SQLConf
import org.apache.spark.sql.types._
import org.apache.spark.util.Utils
/**
* Abstract class all optimizers should inherit of, contains the standard batches (extending
* Optimizers can override this.
*/
abstract class Optimizer(sessionCatalog: SessionCatalog)
extends RuleExecutor[LogicalPlan] {
// Check for structural integrity of the plan in test mode. Currently we only check if a plan is
// still resolved after the execution of each rule.
override protected def isPlanIntegral(plan: LogicalPlan): Boolean = {
!Utils.isTesting || plan.resolved
}
protected def fixedPoint = FixedPoint(SQLConf.get.optimizerMaxIterations)
def batches: Seq[Batch] = {
val operatorOptimizationRuleSet =
Seq(
// Operator push down
PushProjectionThroughUnion,
ReorderJoin,
EliminateOuterJoin,
PushPredicateThroughJoin,
PushDownPredicate,
LimitPushDown,
ColumnPruning,
InferFiltersFromConstraints,
// Operator combine
CollapseRepartition,
CollapseProject,
CollapseWindow,
CombineFilters,
CombineLimits,
CombineUnions,
// Constant folding and strength reduction
NullPropagation,
ConstantPropagation,
FoldablePropagation,
OptimizeIn,
ConstantFolding,
ReorderAssociativeOperator,
LikeSimplification,
BooleanSimplification,
SimplifyConditionals,
RemoveDispensableExpressions,
SimplifyBinaryComparison,
PruneFilters,
EliminateSorts,
SimplifyCasts,
SimplifyCaseConversionExpressions,
RewriteCorrelatedScalarSubquery,
EliminateSerialization,
RemoveRedundantAliases,
RemoveRedundantProject,
SimplifyCreateStructOps,
SimplifyCreateArrayOps,
SimplifyCreateMapOps,
CombineConcats) ++
extendedOperatorOptimizationRules
val operatorOptimizationBatch: Seq[Batch] = {
val rulesWithoutInferFiltersFromConstraints =
operatorOptimizationRuleSet.filterNot(_ == InferFiltersFromConstraints)
Batch("Operator Optimization before Inferring Filters", fixedPoint,
rulesWithoutInferFiltersFromConstraints: _*) ::
Batch("Infer Filters", Once,
InferFiltersFromConstraints) ::
Batch("Operator Optimization after Inferring Filters", fixedPoint,
rulesWithoutInferFiltersFromConstraints: _*) :: Nil
}
(Batch("Eliminate Distinct", Once, EliminateDistinct) ::
// Technically some of the rules in Finish Analysis are not optimizer rules and belong more
// in the analyzer, because they are needed for correctness (e.g. ComputeCurrentTime).
// However, because we also use the analyzer to canonicalized queries (for view definition),
// we do not eliminate subqueries or compute current time in the analyzer.
Batch("Finish Analysis", Once,
EliminateSubqueryAliases,
EliminateView,
ReplaceExpressions,
ComputeCurrentTime,
GetCurrentDatabase(sessionCatalog),
RewriteDistinctAggregates,
ReplaceDeduplicateWithAggregate) ::
//////////////////////////////////////////////////////////////////////////////////////////
// Optimizer rules start here
//////////////////////////////////////////////////////////////////////////////////////////
// - Do the first call of CombineUnions before starting the major Optimizer rules,
// since it can reduce the number of iteration and the other rules could add/move
// extra operators between two adjacent Union operators.
// - Call CombineUnions again in Batch("Operator Optimizations"),
// since the other rules might make two separate Unions operators adjacent.
Batch("Union", Once,
CombineUnions) ::
Batch("Pullup Correlated Expressions", Once,
PullupCorrelatedPredicates) ::
Batch("Subquery", Once,
OptimizeSubqueries) ::
Batch("Replace Operators", fixedPoint,
ReplaceIntersectWithSemiJoin,
ReplaceExceptWithFilter,
ReplaceExceptWithAntiJoin,
ReplaceDistinctWithAggregate) ::
Batch("Aggregate", fixedPoint,
RemoveLiteralFromGroupExpressions,
RemoveRepetitionFromGroupExpressions) :: Nil ++
operatorOptimizationBatch) :+
Batch("Join Reorder", Once,
CostBasedJoinReorder) :+
Batch("Decimal Optimizations", fixedPoint,
DecimalAggregates) :+
Batch("Object Expressions Optimization", fixedPoint,
EliminateMapObjects,
CombineTypedFilters) :+
Batch("LocalRelation", fixedPoint,
ConvertToLocalRelation,
PropagateEmptyRelation) :+
// The following batch should be executed after batch "Join Reorder" and "LocalRelation".
Batch("Check Cartesian Products", Once,
CheckCartesianProducts) :+
Batch("RewriteSubquery", Once,
RewritePredicateSubquery,
ColumnPruning,
CollapseProject,
RemoveRedundantProject)
}
/**
* Optimize all the subqueries inside expression.
*/
object OptimizeSubqueries extends Rule[LogicalPlan] {
def apply(plan: LogicalPlan): LogicalPlan = plan transformAllExpressions {
case s: SubqueryExpression =>
val Subquery(newPlan) = Optimizer.this.execute(Subquery(s.plan))
s.withNewPlan(newPlan)
}
}
/**
* Override to provide additional rules for the operator optimization batch.
*/
def extendedOperatorOptimizationRules: Seq[Rule[LogicalPlan]] = Nil
}
/**
* Remove useless DISTINCT for MAX and MIN.
* This rule should be applied before RewriteDistinctAggregates.
*/
object EliminateDistinct extends Rule[LogicalPlan] {
override def apply(plan: LogicalPlan): LogicalPlan = plan transformExpressions {
case ae: AggregateExpression if ae.isDistinct =>
ae.aggregateFunction match {
case _: Max | _: Min => ae.copy(isDistinct = false)
case _ => ae
}
}
}
/**
* An optimizer used in test code.
*
* To ensure extendability, we leave the standard rules in the abstract optimizer rules, while
* specific rules go to the subclasses
*/
object SimpleTestOptimizer extends SimpleTestOptimizer
class SimpleTestOptimizer extends Optimizer(
new SessionCatalog(
new InMemoryCatalog,
EmptyFunctionRegistry,
new SQLConf().copy(SQLConf.CASE_SENSITIVE -> true)))
/**
* Remove redundant aliases from a query plan. A redundant alias is an alias that does not change
* the name or metadata of a column, and does not deduplicate it.
*/
object RemoveRedundantAliases extends Rule[LogicalPlan] {
/**
* Create an attribute mapping from the old to the new attributes. This function will only
* return the attribute pairs that have changed.
*/
private def createAttributeMapping(current: LogicalPlan, next: LogicalPlan)
: Seq[(Attribute, Attribute)] = {
current.output.zip(next.output).filterNot {
case (a1, a2) => a1.semanticEquals(a2)
}
}
/**
* Remove the top-level alias from an expression when it is redundant.
*/
private def removeRedundantAlias(e: Expression, blacklist: AttributeSet): Expression = e match {
// Alias with metadata can not be stripped, or the metadata will be lost.
// If the alias name is different from attribute name, we can't strip it either, or we
// may accidentally change the output schema name of the root plan.
case a @ Alias(attr: Attribute, name)
if a.metadata == Metadata.empty &&
name == attr.name &&
!blacklist.contains(attr) &&
!blacklist.contains(a) =>
attr
case a => a
}
/**
* Remove redundant alias expression from a LogicalPlan and its subtree. A blacklist is used to
* prevent the removal of seemingly redundant aliases used to deduplicate the input for a (self)
* join or to prevent the removal of top-level subquery attributes.
*/
private def removeRedundantAliases(plan: LogicalPlan, blacklist: AttributeSet): LogicalPlan = {
plan match {
// We want to keep the same output attributes for subqueries. This means we cannot remove
// the aliases that produce these attributes
case Subquery(child) =>
Subquery(removeRedundantAliases(child, blacklist ++ child.outputSet))
// A join has to be treated differently, because the left and the right side of the join are
// not allowed to use the same attributes. We use a blacklist to prevent us from creating a
// situation in which this happens; the rule will only remove an alias if its child
// attribute is not on the black list.
case Join(left, right, joinType, condition) =>
val newLeft = removeRedundantAliases(left, blacklist ++ right.outputSet)
val newRight = removeRedundantAliases(right, blacklist ++ newLeft.outputSet)
val mapping = AttributeMap(
createAttributeMapping(left, newLeft) ++
createAttributeMapping(right, newRight))
val newCondition = condition.map(_.transform {
case a: Attribute => mapping.getOrElse(a, a)
})
Join(newLeft, newRight, joinType, newCondition)
case _ =>
// Remove redundant aliases in the subtree(s).
val currentNextAttrPairs = mutable.Buffer.empty[(Attribute, Attribute)]
val newNode = plan.mapChildren { child =>
val newChild = removeRedundantAliases(child, blacklist)
currentNextAttrPairs ++= createAttributeMapping(child, newChild)
newChild
}
// Create the attribute mapping. Note that the currentNextAttrPairs can contain duplicate
// keys in case of Union (this is caused by the PushProjectionThroughUnion rule); in this
// case we use the the first mapping (which should be provided by the first child).
val mapping = AttributeMap(currentNextAttrPairs)
// Create a an expression cleaning function for nodes that can actually produce redundant
// aliases, use identity otherwise.
val clean: Expression => Expression = plan match {
case _: Project => removeRedundantAlias(_, blacklist)
case _: Aggregate => removeRedundantAlias(_, blacklist)
case _: Window => removeRedundantAlias(_, blacklist)
case _ => identity[Expression]
}
// Transform the expressions.
newNode.mapExpressions { expr =>
clean(expr.transform {
case a: Attribute => mapping.getOrElse(a, a)
})
}
}
}
def apply(plan: LogicalPlan): LogicalPlan = removeRedundantAliases(plan, AttributeSet.empty)
}
/**
* Remove projections from the query plan that do not make any modifications.
*/
object RemoveRedundantProject extends Rule[LogicalPlan] {
def apply(plan: LogicalPlan): LogicalPlan = plan transform {
case p @ Project(_, child) if p.output == child.output => child
}
}
/**
* Pushes down [[LocalLimit]] beneath UNION ALL and beneath the streamed inputs of outer joins.
*/
object LimitPushDown extends Rule[LogicalPlan] {
private def stripGlobalLimitIfPresent(plan: LogicalPlan): LogicalPlan = {
plan match {
case GlobalLimit(_, child) => child
case _ => plan
}
}
private def maybePushLocalLimit(limitExp: Expression, plan: LogicalPlan): LogicalPlan = {
(limitExp, plan.maxRowsPerPartition) match {
case (IntegerLiteral(newLimit), Some(childMaxRows)) if newLimit < childMaxRows =>
// If the child has a cap on max rows per partition and the cap is larger than
// the new limit, put a new LocalLimit there.
LocalLimit(limitExp, stripGlobalLimitIfPresent(plan))
case (_, None) =>
// If the child has no cap, put the new LocalLimit.
LocalLimit(limitExp, stripGlobalLimitIfPresent(plan))
case _ =>
// Otherwise, don't put a new LocalLimit.
plan
}
}
def apply(plan: LogicalPlan): LogicalPlan = plan transform {
// Adding extra Limits below UNION ALL for children which are not Limit or do not have Limit
// descendants whose maxRow is larger. This heuristic is valid assuming there does not exist any
// Limit push-down rule that is unable to infer the value of maxRows.
// Note: right now Union means UNION ALL, which does not de-duplicate rows, so it is safe to
// pushdown Limit through it. Once we add UNION DISTINCT, however, we will not be able to
// pushdown Limit.
case LocalLimit(exp, Union(children)) =>
LocalLimit(exp, Union(children.map(maybePushLocalLimit(exp, _))))
// Add extra limits below OUTER JOIN. For LEFT OUTER and RIGHT OUTER JOIN we push limits to
// the left and right sides, respectively. It's not safe to push limits below FULL OUTER
// JOIN in the general case without a more invasive rewrite.
// We also need to ensure that this limit pushdown rule will not eventually introduce limits
// on both sides if it is applied multiple times. Therefore:
// - If one side is already limited, stack another limit on top if the new limit is smaller.
// The redundant limit will be collapsed by the CombineLimits rule.
// - If neither side is limited, limit the side that is estimated to be bigger.
case LocalLimit(exp, join @ Join(left, right, joinType, _)) =>
val newJoin = joinType match {
case RightOuter => join.copy(right = maybePushLocalLimit(exp, right))
case LeftOuter => join.copy(left = maybePushLocalLimit(exp, left))
case _ => join
}
LocalLimit(exp, newJoin)
}
}
/**
* Pushes Project operator to both sides of a Union operator.
* Operations that are safe to pushdown are listed as follows.
* Union:
* Right now, Union means UNION ALL, which does not de-duplicate rows. So, it is
* safe to pushdown Filters and Projections through it. Filter pushdown is handled by another
* rule PushDownPredicate. Once we add UNION DISTINCT, we will not be able to pushdown Projections.
*/
object PushProjectionThroughUnion extends Rule[LogicalPlan] with PredicateHelper {
/**
* Maps Attributes from the left side to the corresponding Attribute on the right side.
*/
private def buildRewrites(left: LogicalPlan, right: LogicalPlan): AttributeMap[Attribute] = {
assert(left.output.size == right.output.size)
AttributeMap(left.output.zip(right.output))
}
/**
* Rewrites an expression so that it can be pushed to the right side of a
* Union or Except operator. This method relies on the fact that the output attributes
* of a union/intersect/except are always equal to the left child's output.
*/
private def pushToRight[A <: Expression](e: A, rewrites: AttributeMap[Attribute]) = {
val result = e transform {
case a: Attribute => rewrites(a)
}
// We must promise the compiler that we did not discard the names in the case of project
// expressions. This is safe since the only transformation is from Attribute => Attribute.
result.asInstanceOf[A]
}
def apply(plan: LogicalPlan): LogicalPlan = plan transform {
// Push down deterministic projection through UNION ALL
case p @ Project(projectList, Union(children)) =>
assert(children.nonEmpty)
if (projectList.forall(_.deterministic)) {
val newFirstChild = Project(projectList, children.head)
val newOtherChildren = children.tail.map { child =>
val rewrites = buildRewrites(children.head, child)
Project(projectList.map(pushToRight(_, rewrites)), child)
}
Union(newFirstChild +: newOtherChildren)
} else {
p
}
}
}
/**
* Attempts to eliminate the reading of unneeded columns from the query plan.
*
* Since adding Project before Filter conflicts with PushPredicatesThroughProject, this rule will
* remove the Project p2 in the following pattern:
*
* p1 @ Project(_, Filter(_, p2 @ Project(_, child))) if p2.outputSet.subsetOf(p2.inputSet)
*
* p2 is usually inserted by this rule and useless, p1 could prune the columns anyway.
*/
object ColumnPruning extends Rule[LogicalPlan] {
private def sameOutput(output1: Seq[Attribute], output2: Seq[Attribute]): Boolean =
output1.size == output2.size &&
output1.zip(output2).forall(pair => pair._1.semanticEquals(pair._2))
def apply(plan: LogicalPlan): LogicalPlan = removeProjectBeforeFilter(plan transform {
// Prunes the unused columns from project list of Project/Aggregate/Expand
case p @ Project(_, p2: Project) if (p2.outputSet -- p.references).nonEmpty =>
p.copy(child = p2.copy(projectList = p2.projectList.filter(p.references.contains)))
case p @ Project(_, a: Aggregate) if (a.outputSet -- p.references).nonEmpty =>
p.copy(
child = a.copy(aggregateExpressions = a.aggregateExpressions.filter(p.references.contains)))
case a @ Project(_, e @ Expand(_, _, grandChild)) if (e.outputSet -- a.references).nonEmpty =>
val newOutput = e.output.filter(a.references.contains(_))
val newProjects = e.projections.map { proj =>
proj.zip(e.output).filter { case (_, a) =>
newOutput.contains(a)
}.unzip._1
}
a.copy(child = Expand(newProjects, newOutput, grandChild))
// Prunes the unused columns from child of `DeserializeToObject`
case d @ DeserializeToObject(_, _, child) if (child.outputSet -- d.references).nonEmpty =>
d.copy(child = prunedChild(child, d.references))
// Prunes the unused columns from child of Aggregate/Expand/Generate
case a @ Aggregate(_, _, child) if (child.outputSet -- a.references).nonEmpty =>
a.copy(child = prunedChild(child, a.references))
case f @ FlatMapGroupsInPandas(_, _, _, child) if (child.outputSet -- f.references).nonEmpty =>
f.copy(child = prunedChild(child, f.references))
case e @ Expand(_, _, child) if (child.outputSet -- e.references).nonEmpty =>
e.copy(child = prunedChild(child, e.references))
case g: Generate if !g.join && (g.child.outputSet -- g.references).nonEmpty =>
g.copy(child = prunedChild(g.child, g.references))
// Turn off `join` for Generate if no column from it's child is used
case p @ Project(_, g: Generate) if g.join && p.references.subsetOf(g.generatedSet) =>
p.copy(child = g.copy(join = false))
// Eliminate unneeded attributes from right side of a Left Existence Join.
case j @ Join(_, right, LeftExistence(_), _) =>
j.copy(right = prunedChild(right, j.references))
// all the columns will be used to compare, so we can't prune them
case p @ Project(_, _: SetOperation) => p
case p @ Project(_, _: Distinct) => p
// Eliminate unneeded attributes from children of Union.
case p @ Project(_, u: Union) =>
if ((u.outputSet -- p.references).nonEmpty) {
val firstChild = u.children.head
val newOutput = prunedChild(firstChild, p.references).output
// pruning the columns of all children based on the pruned first child.
val newChildren = u.children.map { p =>
val selected = p.output.zipWithIndex.filter { case (a, i) =>
newOutput.contains(firstChild.output(i))
}.map(_._1)
Project(selected, p)
}
p.copy(child = u.withNewChildren(newChildren))
} else {
p
}
// Prune unnecessary window expressions
case p @ Project(_, w: Window) if (w.windowOutputSet -- p.references).nonEmpty =>
p.copy(child = w.copy(
windowExpressions = w.windowExpressions.filter(p.references.contains)))
// Eliminate no-op Window
case w: Window if w.windowExpressions.isEmpty => w.child
// Eliminate no-op Projects
case p @ Project(_, child) if sameOutput(child.output, p.output) => child
// Can't prune the columns on LeafNode
case p @ Project(_, _: LeafNode) => p
// for all other logical plans that inherits the output from it's children
case p @ Project(_, child) =>
val required = child.references ++ p.references
if ((child.inputSet -- required).nonEmpty) {
val newChildren = child.children.map(c => prunedChild(c, required))
p.copy(child = child.withNewChildren(newChildren))
} else {
p
}
})
/** Applies a projection only when the child is producing unnecessary attributes */
private def prunedChild(c: LogicalPlan, allReferences: AttributeSet) =
if ((c.outputSet -- allReferences.filter(c.outputSet.contains)).nonEmpty) {
Project(c.output.filter(allReferences.contains), c)
} else {
c
}
/**
* The Project before Filter is not necessary but conflict with PushPredicatesThroughProject,
* so remove it.
*/
private def removeProjectBeforeFilter(plan: LogicalPlan): LogicalPlan = plan transform {
case p1 @ Project(_, f @ Filter(_, p2 @ Project(_, child)))
if p2.outputSet.subsetOf(child.outputSet) =>
p1.copy(child = f.copy(child = child))
}
}
/**
* Combines two adjacent [[Project]] operators into one and perform alias substitution,
* merging the expressions into one single expression.
*/
object CollapseProject extends Rule[LogicalPlan] {
def apply(plan: LogicalPlan): LogicalPlan = plan transformUp {
case p1 @ Project(_, p2: Project) =>
if (haveCommonNonDeterministicOutput(p1.projectList, p2.projectList)) {
p1
} else {
p2.copy(projectList = buildCleanedProjectList(p1.projectList, p2.projectList))
}
case p @ Project(_, agg: Aggregate) =>
if (haveCommonNonDeterministicOutput(p.projectList, agg.aggregateExpressions)) {
p
} else {
agg.copy(aggregateExpressions = buildCleanedProjectList(
p.projectList, agg.aggregateExpressions))
}
}
private def collectAliases(projectList: Seq[NamedExpression]): AttributeMap[Alias] = {
AttributeMap(projectList.collect {
case a: Alias => a.toAttribute -> a
})
}
private def haveCommonNonDeterministicOutput(
upper: Seq[NamedExpression], lower: Seq[NamedExpression]): Boolean = {
// Create a map of Aliases to their values from the lower projection.
// e.g., 'SELECT ... FROM (SELECT a + b AS c, d ...)' produces Map(c -> Alias(a + b, c)).
val aliases = collectAliases(lower)
// Collapse upper and lower Projects if and only if their overlapped expressions are all
// deterministic.
upper.exists(_.collect {
case a: Attribute if aliases.contains(a) => aliases(a).child
}.exists(!_.deterministic))
}
private def buildCleanedProjectList(
upper: Seq[NamedExpression],
lower: Seq[NamedExpression]): Seq[NamedExpression] = {
// Create a map of Aliases to their values from the lower projection.
// e.g., 'SELECT ... FROM (SELECT a + b AS c, d ...)' produces Map(c -> Alias(a + b, c)).
val aliases = collectAliases(lower)
// Substitute any attributes that are produced by the lower projection, so that we safely
// eliminate it.
// e.g., 'SELECT c + 1 FROM (SELECT a + b AS C ...' produces 'SELECT a + b + 1 ...'
// Use transformUp to prevent infinite recursion.
val rewrittenUpper = upper.map(_.transformUp {
case a: Attribute => aliases.getOrElse(a, a)
})
// collapse upper and lower Projects may introduce unnecessary Aliases, trim them here.
rewrittenUpper.map { p =>
CleanupAliases.trimNonTopLevelAliases(p).asInstanceOf[NamedExpression]
}
}
}
/**
* Combines adjacent [[RepartitionOperation]] operators
*/
object CollapseRepartition extends Rule[LogicalPlan] {
def apply(plan: LogicalPlan): LogicalPlan = plan transformUp {
// Case 1: When a Repartition has a child of Repartition or RepartitionByExpression,
// 1) When the top node does not enable the shuffle (i.e., coalesce API), but the child
// enables the shuffle. Returns the child node if the last numPartitions is bigger;
// otherwise, keep unchanged.
// 2) In the other cases, returns the top node with the child's child
case r @ Repartition(_, _, child: RepartitionOperation) => (r.shuffle, child.shuffle) match {
case (false, true) => if (r.numPartitions >= child.numPartitions) child else r
case _ => r.copy(child = child.child)
}
// Case 2: When a RepartitionByExpression has a child of Repartition or RepartitionByExpression
// we can remove the child.
case r @ RepartitionByExpression(_, child: RepartitionOperation, _) =>
r.copy(child = child.child)
}
}
/**
* Collapse Adjacent Window Expression.
* - If the partition specs and order specs are the same and the window expression are
* independent, collapse into the parent.
*/
object CollapseWindow extends Rule[LogicalPlan] {
def apply(plan: LogicalPlan): LogicalPlan = plan transformUp {
case w1 @ Window(we1, ps1, os1, w2 @ Window(we2, ps2, os2, grandChild))
if ps1 == ps2 && os1 == os2 && w1.references.intersect(w2.windowOutputSet).isEmpty =>
w1.copy(windowExpressions = we2 ++ we1, child = grandChild)
}
}
/**
* Generate a list of additional filters from an operator's existing constraint but remove those
* that are either already part of the operator's condition or are part of the operator's child
* constraints. These filters are currently inserted to the existing conditions in the Filter
* operators and on either side of Join operators.
*
* Note: While this optimization is applicable to all types of join, it primarily benefits Inner and
* LeftSemi joins.
*/
object InferFiltersFromConstraints extends Rule[LogicalPlan] with PredicateHelper {
def apply(plan: LogicalPlan): LogicalPlan = {
if (SQLConf.get.constraintPropagationEnabled) {
inferFilters(plan)
} else {
plan
}
}
private def inferFilters(plan: LogicalPlan): LogicalPlan = plan transform {
case filter @ Filter(condition, child) =>
val newFilters = filter.constraints --
(child.constraints ++ splitConjunctivePredicates(condition))
if (newFilters.nonEmpty) {
Filter(And(newFilters.reduce(And), condition), child)
} else {
filter
}
case join @ Join(left, right, joinType, conditionOpt) =>
// Only consider constraints that can be pushed down completely to either the left or the
// right child
val constraints = join.constraints.filter { c =>
c.references.subsetOf(left.outputSet) || c.references.subsetOf(right.outputSet)
}
// Remove those constraints that are already enforced by either the left or the right child
val additionalConstraints = constraints -- (left.constraints ++ right.constraints)
val newConditionOpt = conditionOpt match {
case Some(condition) =>
val newFilters = additionalConstraints -- splitConjunctivePredicates(condition)
if (newFilters.nonEmpty) Option(And(newFilters.reduce(And), condition)) else None
case None =>
additionalConstraints.reduceOption(And)
}
if (newConditionOpt.isDefined) Join(left, right, joinType, newConditionOpt) else join
}
}
/**
* Combines all adjacent [[Union]] operators into a single [[Union]].
*/
object CombineUnions extends Rule[LogicalPlan] {
def apply(plan: LogicalPlan): LogicalPlan = plan transformDown {
case u: Union => flattenUnion(u, false)
case Distinct(u: Union) => Distinct(flattenUnion(u, true))
}
private def flattenUnion(union: Union, flattenDistinct: Boolean): Union = {
val stack = mutable.Stack[LogicalPlan](union)
val flattened = mutable.ArrayBuffer.empty[LogicalPlan]
while (stack.nonEmpty) {
stack.pop() match {
case Distinct(Union(children)) if flattenDistinct =>
stack.pushAll(children.reverse)
case Union(children) =>
stack.pushAll(children.reverse)
case child =>
flattened += child
}
}
Union(flattened)
}
}
/**
* Combines two adjacent [[Filter]] operators into one, merging the non-redundant conditions into
* one conjunctive predicate.
*/
object CombineFilters extends Rule[LogicalPlan] with PredicateHelper {
def apply(plan: LogicalPlan): LogicalPlan = plan transform {
// The query execution/optimization does not guarantee the expressions are evaluated in order.
// We only can combine them if and only if both are deterministic.
case Filter(fc, nf @ Filter(nc, grandChild)) if fc.deterministic && nc.deterministic =>
(ExpressionSet(splitConjunctivePredicates(fc)) --
ExpressionSet(splitConjunctivePredicates(nc))).reduceOption(And) match {
case Some(ac) =>
Filter(And(nc, ac), grandChild)
case None =>
nf
}
}
}
/**
* Removes no-op SortOrder from Sort
*/
object EliminateSorts extends Rule[LogicalPlan] {
def apply(plan: LogicalPlan): LogicalPlan = plan transform {
case s @ Sort(orders, _, child) if orders.isEmpty || orders.exists(_.child.foldable) =>
val newOrders = orders.filterNot(_.child.foldable)
if (newOrders.isEmpty) child else s.copy(order = newOrders)
}
}
/**
* Removes filters that can be evaluated trivially. This can be done through the following ways:
* 1) by eliding the filter for cases where it will always evaluate to `true`.
* 2) by substituting a dummy empty relation when the filter will always evaluate to `false`.
* 3) by eliminating the always-true conditions given the constraints on the child's output.
*/
object PruneFilters extends Rule[LogicalPlan] with PredicateHelper {
def apply(plan: LogicalPlan): LogicalPlan = plan transform {
// If the filter condition always evaluate to true, remove the filter.
case Filter(Literal(true, BooleanType), child) => child
// If the filter condition always evaluate to null or false,
// replace the input with an empty relation.
case Filter(Literal(null, _), child) =>
LocalRelation(child.output, data = Seq.empty, isStreaming = plan.isStreaming)
case Filter(Literal(false, BooleanType), child) =>
LocalRelation(child.output, data = Seq.empty, isStreaming = plan.isStreaming)
// If any deterministic condition is guaranteed to be true given the constraints on the child's
// output, remove the condition
case f @ Filter(fc, p: LogicalPlan) =>
val (prunedPredicates, remainingPredicates) =
splitConjunctivePredicates(fc).partition { cond =>
cond.deterministic && p.constraints.contains(cond)
}
if (prunedPredicates.isEmpty) {
f
} else if (remainingPredicates.isEmpty) {
p
} else {
val newCond = remainingPredicates.reduce(And)
Filter(newCond, p)
}
}
}
/**
* Pushes [[Filter]] operators through many operators iff:
* 1) the operator is deterministic
* 2) the predicate is deterministic and the operator will not change any of rows.
*
* This heuristic is valid assuming the expression evaluation cost is minimal.
*/
object PushDownPredicate extends Rule[LogicalPlan] with PredicateHelper {
def apply(plan: LogicalPlan): LogicalPlan = plan transform {
// SPARK-13473: We can't push the predicate down when the underlying projection output non-
// deterministic field(s). Non-deterministic expressions are essentially stateful. This
// implies that, for a given input row, the output are determined by the expression's initial
// state and all the input rows processed before. In another word, the order of input rows
// matters for non-deterministic expressions, while pushing down predicates changes the order.
// This also applies to Aggregate.
case Filter(condition, project @ Project(fields, grandChild))
if fields.forall(_.deterministic) && canPushThroughCondition(grandChild, condition) =>
// Create a map of Aliases to their values from the child projection.
// e.g., 'SELECT a + b AS c, d ...' produces Map(c -> a + b).
val aliasMap = AttributeMap(fields.collect {
case a: Alias => (a.toAttribute, a.child)
})
project.copy(child = Filter(replaceAlias(condition, aliasMap), grandChild))
case filter @ Filter(condition, aggregate: Aggregate)
if aggregate.aggregateExpressions.forall(_.deterministic) =>
// Find all the aliased expressions in the aggregate list that don't include any actual
// AggregateExpression, and create a map from the alias to the expression
val aliasMap = AttributeMap(aggregate.aggregateExpressions.collect {
case a: Alias if a.child.find(_.isInstanceOf[AggregateExpression]).isEmpty =>
(a.toAttribute, a.child)
})
// For each filter, expand the alias and check if the filter can be evaluated using
// attributes produced by the aggregate operator's child operator.
val (candidates, containingNonDeterministic) =
splitConjunctivePredicates(condition).span(_.deterministic)
val (pushDown, rest) = candidates.partition { cond =>
val replaced = replaceAlias(cond, aliasMap)
cond.references.nonEmpty && replaced.references.subsetOf(aggregate.child.outputSet)
}
val stayUp = rest ++ containingNonDeterministic
if (pushDown.nonEmpty) {
val pushDownPredicate = pushDown.reduce(And)
val replaced = replaceAlias(pushDownPredicate, aliasMap)
val newAggregate = aggregate.copy(child = Filter(replaced, aggregate.child))
// If there is no more filter to stay up, just eliminate the filter.
// Otherwise, create "Filter(stayUp) <- Aggregate <- Filter(pushDownPredicate)".
if (stayUp.isEmpty) newAggregate else Filter(stayUp.reduce(And), newAggregate)
} else {
filter
}
// Push [[Filter]] operators through [[Window]] operators. Parts of the predicate that can be
// pushed beneath must satisfy the following conditions:
// 1. All the expressions are part of window partitioning key. The expressions can be compound.
// 2. Deterministic.
// 3. Placed before any non-deterministic predicates.
case filter @ Filter(condition, w: Window)
if w.partitionSpec.forall(_.isInstanceOf[AttributeReference]) =>
val partitionAttrs = AttributeSet(w.partitionSpec.flatMap(_.references))
val (candidates, containingNonDeterministic) =
splitConjunctivePredicates(condition).span(_.deterministic)
val (pushDown, rest) = candidates.partition { cond =>
cond.references.subsetOf(partitionAttrs)
}
val stayUp = rest ++ containingNonDeterministic
if (pushDown.nonEmpty) {
val pushDownPredicate = pushDown.reduce(And)
val newWindow = w.copy(child = Filter(pushDownPredicate, w.child))
if (stayUp.isEmpty) newWindow else Filter(stayUp.reduce(And), newWindow)
} else {
filter
}
case filter @ Filter(condition, union: Union) =>
// Union could change the rows, so non-deterministic predicate can't be pushed down
val (pushDown, stayUp) = splitConjunctivePredicates(condition).span(_.deterministic)
if (pushDown.nonEmpty) {
val pushDownCond = pushDown.reduceLeft(And)
val output = union.output
val newGrandChildren = union.children.map { grandchild =>
val newCond = pushDownCond transform {
case e if output.exists(_.semanticEquals(e)) =>
grandchild.output(output.indexWhere(_.semanticEquals(e)))
}
assert(newCond.references.subsetOf(grandchild.outputSet))
Filter(newCond, grandchild)
}
val newUnion = union.withNewChildren(newGrandChildren)
if (stayUp.nonEmpty) {
Filter(stayUp.reduceLeft(And), newUnion)
} else {
newUnion
}
} else {
filter
}
case filter @ Filter(condition, watermark: EventTimeWatermark) =>
// We can only push deterministic predicates which don't reference the watermark attribute.
// We could in theory span() only on determinism and pull out deterministic predicates
// on the watermark separately. But it seems unnecessary and a bit confusing to not simply
// use the prefix as we do for nondeterminism in other cases.
val (pushDown, stayUp) = splitConjunctivePredicates(condition).span(
p => p.deterministic && !p.references.contains(watermark.eventTime))
if (pushDown.nonEmpty) {
val pushDownPredicate = pushDown.reduceLeft(And)
val newWatermark = watermark.copy(child = Filter(pushDownPredicate, watermark.child))
// If there is no more filter to stay up, just eliminate the filter.
// Otherwise, create "Filter(stayUp) <- watermark <- Filter(pushDownPredicate)".
if (stayUp.isEmpty) newWatermark else Filter(stayUp.reduceLeft(And), newWatermark)
} else {
filter
}
case filter @ Filter(_, u: UnaryNode)
if canPushThrough(u) && u.expressions.forall(_.deterministic) =>
pushDownPredicate(filter, u.child) { predicate =>
u.withNewChildren(Seq(Filter(predicate, u.child)))
}
}
private def canPushThrough(p: UnaryNode): Boolean = p match {
// Note that some operators (e.g. project, aggregate, union) are being handled separately
// (earlier in this rule).
case _: AppendColumns => true
case _: ResolvedHint => true
case _: Distinct => true
case _: Generate => true
case _: Pivot => true
case _: RepartitionByExpression => true
case _: Repartition => true
case _: ScriptTransformation => true
case _: Sort => true
case _ => false
}
private def pushDownPredicate(
filter: Filter,
grandchild: LogicalPlan)(insertFilter: Expression => LogicalPlan): LogicalPlan = {
// Only push down the predicates that is deterministic and all the referenced attributes
// come from grandchild.
// TODO: non-deterministic predicates could be pushed through some operators that do not change
// the rows.
val (candidates, containingNonDeterministic) =
splitConjunctivePredicates(filter.condition).span(_.deterministic)
val (pushDown, rest) = candidates.partition { cond =>
cond.references.subsetOf(grandchild.outputSet)
}
val stayUp = rest ++ containingNonDeterministic
if (pushDown.nonEmpty) {
val newChild = insertFilter(pushDown.reduceLeft(And))
if (stayUp.nonEmpty) {
Filter(stayUp.reduceLeft(And), newChild)
} else {
newChild
}
} else {
filter
}
}
/**
* Check if we can safely push a filter through a projection, by making sure that predicate
* subqueries in the condition do not contain the same attributes as the plan they are moved
* into. This can happen when the plan and predicate subquery have the same source.
*/
private def canPushThroughCondition(plan: LogicalPlan, condition: Expression): Boolean = {
val attributes = plan.outputSet
val matched = condition.find {
case s: SubqueryExpression => s.plan.outputSet.intersect(attributes).nonEmpty
case _ => false
}
matched.isEmpty
}
}
/**
* Pushes down [[Filter]] operators where the `condition` can be
* evaluated using only the attributes of the left or right side of a join. Other
* [[Filter]] conditions are moved into the `condition` of the [[Join]].
*
* And also pushes down the join filter, where the `condition` can be evaluated using only the
* attributes of the left or right side of sub query when applicable.
*
* Check https://cwiki.apache.org/confluence/display/Hive/OuterJoinBehavior for more details
*/
object PushPredicateThroughJoin extends Rule[LogicalPlan] with PredicateHelper {
/**
* Splits join condition expressions or filter predicates (on a given join's output) into three
* categories based on the attributes required to evaluate them. Note that we explicitly exclude
* on-deterministic (i.e., stateful) condition expressions in canEvaluateInLeft or
* canEvaluateInRight to prevent pushing these predicates on either side of the join.
*
* @return (canEvaluateInLeft, canEvaluateInRight, haveToEvaluateInBoth)
*/
private def split(condition: Seq[Expression], left: LogicalPlan, right: LogicalPlan) = {
// Note: In order to ensure correctness, it's important to not change the relative ordering of
// any deterministic expression that follows a non-deterministic expression. To achieve this,
// we only consider pushing down those expressions that precede the first non-deterministic
// expression in the condition.
val (pushDownCandidates, containingNonDeterministic) = condition.span(_.deterministic)
val (leftEvaluateCondition, rest) =
pushDownCandidates.partition(_.references.subsetOf(left.outputSet))
val (rightEvaluateCondition, commonCondition) =
rest.partition(expr => expr.references.subsetOf(right.outputSet))
(leftEvaluateCondition, rightEvaluateCondition, commonCondition ++ containingNonDeterministic)
}
def apply(plan: LogicalPlan): LogicalPlan = plan transform {
// push the where condition down into join filter
case f @ Filter(filterCondition, Join(left, right, joinType, joinCondition)) =>
val (leftFilterConditions, rightFilterConditions, commonFilterCondition) =
split(splitConjunctivePredicates(filterCondition), left, right)
joinType match {
case _: InnerLike =>