Skip to content

Commit

Permalink
[SPARK-38570][SQL] Incorrect DynamicPartitionPruning caused by Literal
Browse files Browse the repository at this point in the history 
### What changes were proposed in this pull request?

The return value of Literal.references is an empty AttributeSet, so Literal is mistaken for a partition column.

For example, the sql in the test case will generate such a physical plan when the adaptive is closed:
```text
*(4) Project [store_id#5281, date_id#5283, state_province#5292]
+- *(4) BroadcastHashJoin [store_id#5281], [store_id#5291], Inner, BuildRight, false
   :- Union
   :  :- *(1) Project [4 AS store_id#5281, date_id#5283]
   :  :  +- *(1) Filter ((isnotnull(date_id#5283) AND (date_id#5283 >= 1300)) AND dynamicpruningexpression(4 IN dynamicpruning#5300))
   :  :     :  +- ReusedSubquery SubqueryBroadcast dynamicpruning#5300, 0, [store_id#5291], [id=#336]
   :  :     +- *(1) ColumnarToRow
   :  :        +- FileScan parquet default.fact_sk[date_id#5283,store_id#5286] Batched: true, DataFilters: [isnotnull(date_id#5283), (date_id#5283 >= 1300)], Format: Parquet, Location: CatalogFileIndex(1 paths)[file:/Users/dongdongzhang/code/study/spark/spark-warehouse/org.apache.s..., PartitionFilters: [dynamicpruningexpression(4 IN dynamicpruning#5300)], PushedFilters: [IsNotNull(date_id), GreaterThanOrEqual(date_id,1300)], ReadSchema: struct<date_id:int>
   :  :              +- SubqueryBroadcast dynamicpruning#5300, 0, [store_id#5291], [id=#336]
   :  :                 +- BroadcastExchange HashedRelationBroadcastMode(List(cast(input[0, int, true] as bigint)),false), [id=#335]
   :  :                    +- *(1) Project [store_id#5291, state_province#5292]
   :  :                       +- *(1) Filter (((isnotnull(country#5293) AND (country#5293 = US)) AND ((store_id#5291 <=> 4) OR (store_id#5291 <=> 5))) AND isnotnull(store_id#5291))
   :  :                          +- *(1) ColumnarToRow
   :  :                             +- FileScan parquet default.dim_store[store_id#5291,state_province#5292,country#5293] Batched: true, DataFilters: [isnotnull(country#5293), (country#5293 = US), ((store_id#5291 <=> 4) OR (store_id#5291 <=> 5)), ..., Format: Parquet, Location: InMemoryFileIndex(1 paths)[file:/Users/dongdongzhang/code/study/spark/spark-warehouse/org.apache...., PartitionFilters: [], PushedFilters: [IsNotNull(country), EqualTo(country,US), Or(EqualNullSafe(store_id,4),EqualNullSafe(store_id,5))..., ReadSchema: struct<store_id:int,state_province:string,country:string>
   :  +- *(2) Project [5 AS store_id#5282, date_id#5287]
   :     +- *(2) Filter ((isnotnull(date_id#5287) AND (date_id#5287 <= 1000)) AND dynamicpruningexpression(5 IN dynamicpruning#5300))
   :        :  +- ReusedSubquery SubqueryBroadcast dynamicpruning#5300, 0, [store_id#5291], [id=#336]
   :        +- *(2) ColumnarToRow
   :           +- FileScan parquet default.fact_stats[date_id#5287,store_id#5290] Batched: true, DataFilters: [isnotnull(date_id#5287), (date_id#5287 <= 1000)], Format: Parquet, Location: CatalogFileIndex(1 paths)[file:/Users/dongdongzhang/code/study/spark/spark-warehouse/org.apache.s..., PartitionFilters: [dynamicpruningexpression(5 IN dynamicpruning#5300)], PushedFilters: [IsNotNull(date_id), LessThanOrEqual(date_id,1000)], ReadSchema: struct<date_id:int>
   :                 +- ReusedSubquery SubqueryBroadcast dynamicpruning#5300, 0, [store_id#5291], [id=#336]
   +- ReusedExchange [store_id#5291, state_province#5292], BroadcastExchange HashedRelationBroadcastMode(List(cast(input[0, int, true] as bigint)),false), [id=#335]
```

after this pr:
```text
*(4) Project [store_id#5281, date_id#5283, state_province#5292]
+- *(4) BroadcastHashJoin [store_id#5281], [store_id#5291], Inner, BuildRight, false
   :- Union
   :  :- *(1) Project [4 AS store_id#5281, date_id#5283]
   :  :  +- *(1) Filter (isnotnull(date_id#5283) AND (date_id#5283 >= 1300))
   :  :     +- *(1) ColumnarToRow
   :  :        +- FileScan parquet default.fact_sk[date_id#5283,store_id#5286] Batched: true, DataFilters: [isnotnull(date_id#5283), (date_id#5283 >= 1300)], Format: Parquet, Location: CatalogFileIndex(1 paths)[file:/Users/dongdongzhang/code/study/spark/spark-warehouse/org.apache.s..., PartitionFilters: [], PushedFilters: [IsNotNull(date_id), GreaterThanOrEqual(date_id,1300)], ReadSchema: struct<date_id:int>
   :  +- *(2) Project [5 AS store_id#5282, date_id#5287]
   :     +- *(2) Filter (isnotnull(date_id#5287) AND (date_id#5287 <= 1000))
   :        +- *(2) ColumnarToRow
   :           +- FileScan parquet default.fact_stats[date_id#5287,store_id#5290] Batched: true, DataFilters: [isnotnull(date_id#5287), (date_id#5287 <= 1000)], Format: Parquet, Location: CatalogFileIndex(1 paths)[file:/Users/dongdongzhang/code/study/spark/spark-warehouse/org.apache.s..., PartitionFilters: [], PushedFilters: [IsNotNull(date_id), LessThanOrEqual(date_id,1000)], ReadSchema: struct<date_id:int>
   +- BroadcastExchange HashedRelationBroadcastMode(List(cast(input[0, int, true] as bigint)),false), [id=#326]
      +- *(3) Project [store_id#5291, state_province#5292]
         +- *(3) Filter (((isnotnull(country#5293) AND (country#5293 = US)) AND ((store_id#5291 <=> 4) OR (store_id#5291 <=> 5))) AND isnotnull(store_id#5291))
            +- *(3) ColumnarToRow
               +- FileScan parquet default.dim_store[store_id#5291,state_province#5292,country#5293] Batched: true, DataFilters: [isnotnull(country#5293), (country#5293 = US), ((store_id#5291 <=> 4) OR (store_id#5291 <=> 5)), ..., Format: Parquet, Location: InMemoryFileIndex(1 paths)[file:/Users/dongdongzhang/code/study/spark/spark-warehouse/org.apache...., PartitionFilters: [], PushedFilters: [IsNotNull(country), EqualTo(country,US), Or(EqualNullSafe(store_id,4),EqualNullSafe(store_id,5))..., ReadSchema: struct<store_id:int,state_province:string,country:string>
```

### Why are the changes needed?
Execution performance improvement

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?
Added unit test

Closes #35878 from mcdull-zhang/literal_dynamic_partition.

Lead-authored-by: mcdull-zhang <[email protected]>
Co-authored-by: mcdull_zhang <[email protected]>
Signed-off-by: Yuming Wang <[email protected]>
  • Loading branch information
@mcdull-zhang @wangyum
mcdull-zhang authored and wangyum committed Mar 25, 2022
1 parent de960a5 commit 4c51851
Show file tree
Hide file tree
Showing 2 changed files with 29 additions and 0 deletions.
1 change: 1 addition & 0 deletions sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/predicates.scala
View file
Original file line number Diff line number Diff line change
Expand Up @@ -128,6 +128,7 @@ trait PredicateHelper extends AliasHelper with Logging {
def findExpressionAndTrackLineageDown(
exp: Expression,
plan: LogicalPlan): Option[(Expression, LogicalPlan)] = {
if (exp.references.isEmpty) return None

plan match {
case p: Project =>
Expand Down
28 changes: 28 additions & 0 deletions sql/core/src/test/scala/org/apache/spark/sql/DynamicPartitionPruningSuite.scala
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1528,6 +1528,34 @@ abstract class DynamicPartitionPruningSuiteBase
}
}
}

test("SPARK-38570: Fix incorrect DynamicPartitionPruning caused by Literal") {
withSQLConf(SQLConf.DYNAMIC_PARTITION_PRUNING_ENABLED.key -> "true") {
val df = sql(
"""
|SELECT f.store_id,
| f.date_id,
| s.state_province
|FROM (SELECT 4 AS store_id,
| date_id,
| product_id
| FROM fact_sk
| WHERE date_id >= 1300
| UNION ALL
| SELECT 5 AS store_id,
| date_id,
| product_id
| FROM fact_stats
| WHERE date_id <= 1000) f
|JOIN dim_store s
|ON f.store_id = s.store_id
|WHERE s.country = 'US'
|""".stripMargin)

checkPartitionPruningPredicate(df, withSubquery = false, withBroadcast = false)
checkAnswer(df, Row(4, 1300, "California") :: Row(5, 1000, "Texas") :: Nil)
}
}
}

abstract class DynamicPartitionPruningDataSourceSuiteBase
Expand Down

0 comments on commit 4c51851

Please sign in to comment.