[SPARK-38570][SQL][3.1] Incorrect DynamicPartitionPruning caused by Literal

mcdull-zhang · wangyum · commit 789ec137c1e2 · 2022-03-26T13:27:25.000+08:00
This is a backport of #35878 to branch 3.1. 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: ```tex *(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: ```tex *(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> ``` Execution performance improvement No Added unit test Closes #35967 from mcdull-zhang/spark_38570_3.2. Authored-by: mcdull-zhang <work4dong@163.com> Signed-off-by: Yuming Wang <yumwang@ebay.com> (cherry picked from commit 8621914) Signed-off-by: Yuming Wang <yumwang@ebay.com>
diff --git a/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/predicates.scala b/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/predicates.scala
@@ -127,6 +127,7 @@ 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 =>
diff --git a/sql/core/src/test/scala/org/apache/spark/sql/DynamicPartitionPruningSuite.scala b/sql/core/src/test/scala/org/apache/spark/sql/DynamicPartitionPruningSuite.scala
@@ -1382,6 +1382,34 @@ 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)
+    }
+  }
 }
 
 class DynamicPartitionPruningSuiteAEOff extends DynamicPartitionPruningSuiteBase {
