[SC-3623][BRANCH-2.1] Dynamic partition pruning

## What changes were proposed in this pull request? This PR ports databricks/runtime#31 over to DB Spark branch-2.1. This adds dynamic partition pruning (see the original PR for more details on the feature). This was non-trivial to port because the read path has changed significantly in Spark 2.1. We only support partition pruning for `HadoopFsRelation`. This relation is exclusively read using the `FileSourceScanExec` and so I have only implemented dynamic partition pruning for this scan operator (in 2.0 we support dynamic partition pruning for both `RowScanExec` and `FileSourceScanExec`). ## How was this patch tested? Added a test to `SQLQuerySuite`. Author: Herman van Hovell <[email protected]> Author: Davies Liu <[email protected]> Closes apache#131 from hvanhovell/dynamic_partition_pruning.
ericl · Dec 8, 2016 · 53e0bb3 · 53e0bb3
1 parent 27e7e3d
commit 53e0bb3
Show file tree

Hide file tree

Showing 8 changed files with 185 additions and 10 deletions.
diff --git a/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/optimizer/subquery.scala b/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/optimizer/subquery.scala
@@ -54,6 +54,9 @@ object RewritePredicateSubquery extends Rule[LogicalPlan] with PredicateHelper {
 
       // Filter the plan by applying left semi and left anti joins.
       withSubquery.foldLeft(newFilter) {
+        case (p, PredicateSubquery(_, Seq(e: Expression), _, _)) if !e.isInstanceOf[Predicate] =>
+          // This predicate subquery is inserted by PartitionPruning rule, should not be rewritten.
+          p
         case (p, PredicateSubquery(sub, conditions, _, _)) =>
           val (joinCond, outerPlan) = rewriteExistentialExpr(conditions, p)
           Join(outerPlan, sub, LeftSemi, joinCond)

diff --git a/sql/core/src/main/scala/org/apache/spark/sql/execution/DataSourceScanExec.scala b/sql/core/src/main/scala/org/apache/spark/sql/execution/DataSourceScanExec.scala
@@ -156,7 +156,30 @@ case class FileSourceScanExec(
     false
   }
 
-  @transient private lazy val selectedPartitions = relation.location.listFiles(partitionFilters)
+  private def isDynamicPartitionFilter(e: Expression): Boolean =
+    e.find(_.isInstanceOf[PlanExpression[_]]).isDefined
+
+  @transient private lazy val selectedPartitions =
+    relation.location.listFiles(partitionFilters.filterNot(isDynamicPartitionFilter))
+
+  // We can only determine the actual partitions at runtime when a dynamic partition filter is
+  // present. This is because such a filter relies on information that is only available at run
+  // time (for instance the keys used in the other side of a join).
+  @transient private lazy val dynamicallySelectedPartitions = {
+    val dynamicPartitionFilters = partitionFilters.filter(isDynamicPartitionFilter)
+    if (dynamicPartitionFilters.nonEmpty) {
+      val predicate = dynamicPartitionFilters.reduce(And)
+      val partitionColumns = relation.partitionSchema
+      val boundPredicate = newPredicate(predicate.transform {
+        case a: AttributeReference =>
+          val index = partitionColumns.indexWhere(a.name == _.name)
+          BoundReference(index, partitionColumns(index).dataType, nullable = true)
+      }, Nil)
+      selectedPartitions.filter(p => boundPredicate.eval(p.values))
+    } else {
+      selectedPartitions
+    }
+  }
 
   override val (outputPartitioning, outputOrdering): (Partitioning, Seq[SortOrder]) = {
     val bucketSpec = if (relation.sparkSession.sessionState.conf.bucketingEnabled) {
@@ -261,9 +284,9 @@ case class FileSourceScanExec(
 
     relation.bucketSpec match {
       case Some(bucketing) if relation.sparkSession.sessionState.conf.bucketingEnabled =>
-        createBucketedReadRDD(bucketing, readFile, selectedPartitions, relation)
+        createBucketedReadRDD(bucketing, readFile, dynamicallySelectedPartitions, relation)
       case _ =>
-        createNonBucketedReadRDD(readFile, selectedPartitions, relation)
+        createNonBucketedReadRDD(readFile, dynamicallySelectedPartitions, relation)
     }
   }
 

diff --git a/sql/core/src/main/scala/org/apache/spark/sql/execution/SparkOptimizer.scala b/sql/core/src/main/scala/org/apache/spark/sql/execution/SparkOptimizer.scala
@@ -19,11 +19,13 @@ package org.apache.spark.sql.execution
 
 import org.apache.spark.sql.ExperimentalMethods
 import org.apache.spark.sql.catalyst.catalog.SessionCatalog
-import org.apache.spark.sql.catalyst.optimizer.Optimizer
-import org.apache.spark.sql.catalyst.plans.logical.LogicalPlan
+import org.apache.spark.sql.catalyst.expressions._
+import org.apache.spark.sql.catalyst.optimizer.{CombineFilters, Optimizer, PushDownPredicate, PushPredicateThroughJoin}
+import org.apache.spark.sql.catalyst.plans.{Inner, LeftOuter, LeftSemi, RightOuter}
+import org.apache.spark.sql.catalyst.plans.logical._
 import org.apache.spark.sql.catalyst.rules.Rule
 import org.apache.spark.sql.execution.closure.TranslateClosureOptimizerRule
-import org.apache.spark.sql.execution.datasources.PruneFileSourcePartitions
+import org.apache.spark.sql.execution.datasources.{HadoopFsRelation, LogicalRelation, PruneFileSourcePartitions}
 import org.apache.spark.sql.execution.python.ExtractPythonUDFFromAggregate
 import org.apache.spark.sql.internal.SQLConf
 
@@ -43,10 +45,109 @@ class SparkOptimizer(
       // Java closure to Catalyst expressions
       Batch("Translate Closure", Once, new TranslateClosureOptimizerRule(conf))) ++
       defaultOptimizers :+
+      Batch("PartitionPruning", Once,
+        PartitionPruning(conf),
+        OptimizeSubqueries) :+
+      Batch("Pushdown pruning subquery", fixedPoint,
+        PushPredicateThroughJoin,
+        PushDownPredicate,
+        CombineFilters) :+
       Batch("Optimize Metadata Only Query", Once, OptimizeMetadataOnlyQuery(catalog, conf)) :+
       Batch("Extract Python UDF from Aggregate", Once, ExtractPythonUDFFromAggregate) :+
       Batch("Prune File Source Table Partitions", Once, PruneFileSourcePartitions) :+
       Batch("User Provided Optimizers", fixedPoint,
         experimentalMethods.extraOptimizations ++ extraOptimizationRules: _*)
   }
 }
+
+/**
+ * Inserts a predicate for partitioned table when partition column is used as join key.
+ */
+case class PartitionPruning(conf: SQLConf) extends Rule[LogicalPlan] with PredicateHelper {
+
+  /**
+   * Returns whether an attribute is a partition column or not.
+   */
+  private def isPartitioned(a: Expression, plan: LogicalPlan): Boolean = {
+    plan.foreach {
+      case l: LogicalRelation if a.references.subsetOf(l.outputSet) =>
+        l.relation match {
+          case fs: HadoopFsRelation =>
+            val partitionColumns = AttributeSet(
+              l.resolve(fs.partitionSchema, fs.sparkSession.sessionState.analyzer.resolver))
+            if (a.references.subsetOf(partitionColumns)) {
+              return true
+            }
+          case _ =>
+        }
+      case _ =>
+    }
+    false
+  }
+
+  private def insertPredicate(
+      partitionedPlan: LogicalPlan,
+      partitioned: Expression,
+      otherPlan: LogicalPlan,
+      value: Expression): LogicalPlan = {
+    val alias = value match {
+      case a: Attribute => a
+      case o => Alias(o, o.toString)()
+    }
+    Filter(
+      PredicateSubquery(Aggregate(Seq(alias), Seq(alias), otherPlan), Seq(partitioned)),
+      partitionedPlan)
+  }
+
+  def apply(plan: LogicalPlan): LogicalPlan = {
+    if (!conf.partitionPruning) {
+      return plan
+    }
+    plan transformUp {
+      case join @ Join(left, right, joinType, Some(condition)) =>
+        var newLeft = left
+        var newRight = right
+        splitConjunctivePredicates(condition).foreach {
+          case e @ EqualTo(a: Expression, b: Expression) =>
+            // they should come from different sides, otherwise should be pushed down
+            val (l, r) = if (a.references.subsetOf(left.outputSet) &&
+              b.references.subsetOf(right.outputSet)) {
+              a -> b
+            } else {
+              b -> a
+            }
+            if (isPartitioned(l, left) && hasHighlySelectivePredicate(right) &&
+              (joinType == Inner || joinType == LeftSemi || joinType == RightOuter) &&
+              r.references.subsetOf(right.outputSet)) {
+              newLeft = insertPredicate(newLeft, l, right, r)
+            } else if (isPartitioned(r, right) && hasHighlySelectivePredicate(left) &&
+              (joinType == Inner || joinType == LeftOuter) &&
+              l.references.subsetOf(left.outputSet)) {
+              newRight = insertPredicate(newRight, r, left, l)
+            }
+          case _ =>
+        }
+        Join(newLeft, newRight, joinType, Some(condition))
+    }
+  }
+
+  /**
+   * Returns whether an expression is highly selective or not.
+   */
+  def isHighlySelective(e: Expression): Boolean = e match {
+    case Not(expr) => isHighlySelective(expr)
+    case And(l, r) => isHighlySelective(l) || isHighlySelective(r)
+    case Or(l, r) => isHighlySelective(l) && isHighlySelective(r)
+    case _: BinaryComparison => true
+    case _: In | _: InSet => true
+    case _: StringPredicate => true
+    case _ => false
+  }
+
+  def hasHighlySelectivePredicate(plan: LogicalPlan): Boolean = {
+    plan.find {
+      case f: Filter => isHighlySelective(f.condition)
+      case _ => false
+    }.isDefined
+  }
+}
diff --git a/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/FileSourceStrategy.scala b/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/FileSourceStrategy.scala
@@ -67,7 +67,7 @@ object FileSourceStrategy extends Strategy with Logging {
       val normalizedFilters = filters.map { e =>
         e transform {
           case a: AttributeReference =>
-            a.withName(l.output.find(_.semanticEquals(a)).get.name)
+            a.withName(l.output.find(_.semanticEquals(a)).getOrElse(a).name)
         }
       }
 

diff --git a/...src/main/scala/org/apache/spark/sql/execution/datasources/PruneFileSourcePartitions.scala b/...src/main/scala/org/apache/spark/sql/execution/datasources/PruneFileSourcePartitions.scala
@@ -52,8 +52,9 @@ private[sql] object PruneFileSourcePartitions extends Rule[LogicalPlan] {
         logicalRelation.resolve(
           partitionSchema, sparkSession.sessionState.analyzer.resolver)
       val partitionSet = AttributeSet(partitionColumns)
-      val partitionKeyFilters =
-        ExpressionSet(normalizedFilters.filter(_.references.subsetOf(partitionSet)))
+      val partitionKeyFilters = ExpressionSet(normalizedFilters.filter { f =>
+        f.references.subsetOf(partitionSet) && f.find(_.isInstanceOf[SubqueryExpression]).isEmpty
+      })
 
       if (partitionKeyFilters.nonEmpty) {
         val prunedFileIndex = catalogFileIndex.filterPartitions(partitionKeyFilters.toSeq)

diff --git a/sql/core/src/main/scala/org/apache/spark/sql/internal/SQLConf.scala b/sql/core/src/main/scala/org/apache/spark/sql/internal/SQLConf.scala
@@ -417,6 +417,12 @@ object SQLConf {
     .booleanConf
     .createWithDefault(true)
 
+  val PARTITION_PRUNING = SQLConfigBuilder("spark.sql.dynamicPartitionPruning")
+    .internal()
+    .doc("When true, we will generate predicate for partition column when it's used as join key")
+    .booleanConf
+    .createWithDefault(true)
+
   val WHOLESTAGE_CODEGEN_ENABLED = SQLConfigBuilder("spark.sql.codegen.wholeStage")
     .internal()
     .doc("When true, the whole stage (of multiple operators) will be compiled into single java" +
@@ -790,6 +796,8 @@ private[sql] class SQLConf extends Serializable with CatalystConf with Logging {
 
   def enableTwoLevelAggMap: Boolean = getConf(ENABLE_TWOLEVEL_AGG_MAP)
 
+  def partitionPruning: Boolean = getConf(PARTITION_PRUNING)
+
   def variableSubstituteEnabled: Boolean = getConf(VARIABLE_SUBSTITUTE_ENABLED)
 
   def variableSubstituteDepth: Int = getConf(VARIABLE_SUBSTITUTE_DEPTH)

diff --git a/sql/core/src/test/scala/org/apache/spark/sql/SQLQuerySuite.scala b/sql/core/src/test/scala/org/apache/spark/sql/SQLQuerySuite.scala
@@ -22,8 +22,9 @@ import java.math.MathContext
 import java.sql.Timestamp
 
 import org.apache.spark.{AccumulatorSuite, SparkException}
+import org.apache.spark.sql.catalyst.expressions.PlanExpression
 import org.apache.spark.sql.catalyst.util.StringUtils
-import org.apache.spark.sql.execution.aggregate
+import org.apache.spark.sql.execution.{aggregate, FileSourceScanExec}
 import org.apache.spark.sql.execution.joins.{BroadcastHashJoinExec, CartesianProductExec, SortMergeJoinExec}
 import org.apache.spark.sql.functions._
 import org.apache.spark.sql.internal.SQLConf
@@ -2086,6 +2087,20 @@ class SQLQuerySuite extends QueryTest with SharedSQLContext {
     }
   }
 
+  test("dynamic partition pruning") {
+    withTempDir { dir =>
+      val df = spark.range(100).selectExpr("id", "id as k")
+      df.write.mode("overwrite").partitionBy("k").parquet(dir.toString)
+      val df2 = spark.read.parquet(dir.toString).join(df.filter("id < 2"), "k")
+      assert(df2.queryExecution.executedPlan.find {
+        case s: FileSourceScanExec =>
+          s.partitionFilters.exists(_.find(_.isInstanceOf[PlanExpression[_]]).isDefined)
+        case o => false
+      }.isDefined, "Parquet scan should have partition predicate")
+      checkAnswer(df2, Row(0, 0, 0) :: Row(1, 1, 1) :: Nil)
+    }
+  }
+
   test("SPARK-14986: Outer lateral view with empty generate expression") {
     checkAnswer(
       sql("select nil from (select 1 as x ) x lateral view outer explode(array()) n as nil"),

diff --git a/sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/SQLQuerySuite.scala b/sql/hive/src/test/scala/org/apache/spark/sql/hive/execution/SQLQuerySuite.scala
@@ -2011,6 +2011,30 @@ class SQLQuerySuite extends QueryTest with SQLTestUtils with TestHiveSingleton {
     }
   }
 
+  test("dynamic partition pruning") {
+    withTable("df1", "df2") {
+      spark.range(100)
+        .select($"id", $"id".as("k"))
+        .write
+        .partitionBy("k")
+        .format("parquet")
+        .mode("overwrite")
+        .saveAsTable("df1")
+
+      spark.range(100)
+        .select($"id", $"id".as("k"))
+        .write
+        .partitionBy("k")
+        .format("parquet")
+        .mode("overwrite")
+        .saveAsTable("df2")
+
+      checkAnswer(
+        sql("select df1.id, df2.k from df1 join df2 on df1.k = df2.k and df2.id < 2"),
+        Row(0, 0) :: Row(1, 1) :: Nil)
+    }
+  }
+
   def testCommandAvailable(command: String): Boolean = {
     val attempt = Try(Process(command).run(ProcessLogger(_ => ())).exitValue())
     attempt.isSuccess && attempt.get == 0