Check and remove unnecessary shuffle added by Hybrid Scan

Author: sezruby

src/main/scala/com/microsoft/hyperspace/index/IndexConstants.scala

@@ -47,6 +47,12 @@ object IndexConstants {
     "spark.hyperspace.index.hybridscan.maxAppendedRatio"
   val INDEX_HYBRID_SCAN_APPENDED_RATIO_THRESHOLD_DEFAULT = "0.3"
 
+  // If this config is enabled, Hybrid Scan won't be applied when it causes unnecessary shuffle
+  // to merge appended data.
+  val INDEX_HYBRID_SCAN_SHUFFLE_CHECK_ENABLED =
+    "spark.hyperspace.index.hybridscan.shuffleCheck.enabled"
+  val INDEX_HYBRID_SCAN_SHUFFLE_CHECK_ENABLED_DEFAULT = "true"
+
   // Identifier injected to HadoopFsRelation as an option if an index is applied.
   // Currently, the identifier is added to options field of HadoopFsRelation.
   // In Spark 3.0, we could utilize TreeNodeTag to mark the identifier for each plan.

src/main/scala/com/microsoft/hyperspace/index/rules/JoinIndexRule.scala

@@ -24,13 +24,17 @@ import org.apache.spark.sql.catalyst.analysis.CleanupAliases
 import org.apache.spark.sql.catalyst.expressions.{Alias, And, Attribute, AttributeReference, AttributeSet, EqualTo, Expression}
 import org.apache.spark.sql.catalyst.plans.logical.{Join, LogicalPlan, Project}
 import org.apache.spark.sql.catalyst.rules.Rule
+import org.apache.spark.sql.execution.SparkPlan
 import org.apache.spark.sql.execution.datasources.{HadoopFsRelation, LogicalRelation}
+import org.apache.spark.sql.execution.exchange.ShuffleExchangeExec
+import org.apache.spark.sql.execution.joins.SortMergeJoinExec
 
 import com.microsoft.hyperspace.{ActiveSparkSession, Hyperspace}
 import com.microsoft.hyperspace.actions.Constants
 import com.microsoft.hyperspace.index._
 import com.microsoft.hyperspace.index.rankers.JoinIndexRanker
 import com.microsoft.hyperspace.telemetry.{AppInfo, HyperspaceEventLogging, HyperspaceIndexUsageEvent}
+import com.microsoft.hyperspace.util.HyperspaceConf
 import com.microsoft.hyperspace.util.ResolverUtils._
 
 /**
@@ -68,18 +72,54 @@ object JoinIndexRule
                     right =
                       RuleUtils.transformPlanToUseIndex(spark, rIndex, r, useBucketSpec = true))
 
+              def getShuffleCnt(sparkPlan: SparkPlan): Long = {
+                sparkPlan.collect { case _: ShuffleExchangeExec => true }.length
+              }
+
+              val resultPlan =
+                if (!HyperspaceConf.hybridScanEnabled(spark) || !HyperspaceConf
+                      .hybridScanShuffleCheckEnabled(spark)) {
+                  updatedPlan
+                } else {
+                  val shuffleCntPair = spark.sessionState
+                    .executePlan(updatedPlan)
+                    .executedPlan
+                    .collect {
+                      case smj: SortMergeJoinExec =>
+                        (getShuffleCnt(smj.left), getShuffleCnt(smj.right))
+                    }
+                  assert(shuffleCntPair.length <= 1)
+                  shuffleCntPair.headOption.flatMap {
+                    // If the number of shuffle is 2, the candidate index pair cannot remove
+                    // the shuffles in both left and right for join and also Hybrid Scan causes
+                    // an additional shuffle for merging appended files. We don't apply the index
+                    // for the child with 2 shuffle nodes using JoinIndexRule.
+                    // However, the child node is still applicable for FilterIndexRule.
+                    case (leftCnt, _) if leftCnt == 2 => Some(updatedPlan.copy(left = l))
+                    case (_, rightCnt) if rightCnt == 2 => Some(updatedPlan.copy(right = r))
+                    case _ => None
+                  }.getOrElse {
+                    updatedPlan
+                  }
+                }
+
+              val appliedIndexes = resultPlan match {
+                case j: Join if j.left.equals(l) => Seq(rIndex)
+                case j: Join if j.right.equals(r) => Seq(lIndex)
+                case _ => Seq(lIndex, rIndex)
+              }
+
               logEvent(
                 HyperspaceIndexUsageEvent(
                   AppInfo(
                     sparkContext.sparkUser,
                     sparkContext.applicationId,
                     sparkContext.appName),
-                  Seq(lIndex, rIndex),
+                  appliedIndexes,
                   join.toString,
-                  updatedPlan.toString,
+                  resultPlan.toString,
                   "Join index rule applied."))
-
-              updatedPlan
+              resultPlan
           }
           .getOrElse(join)
       } catch {
@@ -325,11 +365,7 @@ object JoinIndexRule
     compatibleIndexPairs.map(
       indexPairs =>
         JoinIndexRanker
-          .rank(
-            spark,
-            leftRel,
-            rightRel,
-            indexPairs)
+          .rank(spark, leftRel, rightRel, indexPairs)
           .head)
   }
 

src/main/scala/com/microsoft/hyperspace/util/HyperspaceConf.scala

@@ -36,6 +36,14 @@ object HyperspaceConf {
     hybridScanDeletedRatioThreshold(spark) > 0.0
   }
 
+  def hybridScanShuffleCheckEnabled(spark: SparkSession): Boolean = {
+    spark.conf
+      .get(
+        IndexConstants.INDEX_HYBRID_SCAN_SHUFFLE_CHECK_ENABLED,
+        IndexConstants.INDEX_HYBRID_SCAN_SHUFFLE_CHECK_ENABLED_DEFAULT)
+      .toBoolean
+  }
+
   def optimizeFileSizeThreshold(spark: SparkSession): Long = {
     spark.conf
       .get(

src/test/scala/com/microsoft/hyperspace/index/E2EHyperspaceRulesTest.scala

@@ -20,9 +20,10 @@ import org.apache.hadoop.conf.Configuration
 import org.apache.hadoop.fs.Path
 import org.apache.spark.sql.{AnalysisException, DataFrame, QueryTest, Row}
 import org.apache.spark.sql.catalyst.plans.logical.{LogicalPlan, Project}
-import org.apache.spark.sql.execution.SortExec
+import org.apache.spark.sql.execution.{SortExec, SparkPlan => SparkPlanNode}
 import org.apache.spark.sql.execution.datasources.{HadoopFsRelation, InMemoryFileIndex, LogicalRelation}
 import org.apache.spark.sql.execution.exchange.ShuffleExchangeExec
+import org.apache.spark.sql.execution.joins.SortMergeJoinExec
 
 import com.microsoft.hyperspace.{Hyperspace, Implicits, SampleData, TestConfig, TestUtils}
 import com.microsoft.hyperspace.index.IndexConstants.{GLOBBING_PATTERN_KEY, REFRESH_MODE_INCREMENTAL, REFRESH_MODE_QUICK}
@@ -670,13 +671,95 @@ class E2EHyperspaceRulesTest extends QueryTest with HyperspaceSuite {
 
         // Refreshed index as quick mode can be applied with Hybrid Scan config.
         withSQLConf(TestConfig.HybridScanEnabled: _*) {
-            spark.disableHyperspace()
-            val dfWithHyperspaceDisabled = query()
-            val basePlan = dfWithHyperspaceDisabled.queryExecution.optimizedPlan
-            spark.enableHyperspace()
-            val dfWithHyperspaceEnabled = query()
-            assert(!basePlan.equals(dfWithHyperspaceEnabled.queryExecution.optimizedPlan))
-            checkAnswer(dfWithHyperspaceDisabled, dfWithHyperspaceEnabled)
+          spark.disableHyperspace()
+          val dfWithHyperspaceDisabled = query()
+          val basePlan = dfWithHyperspaceDisabled.queryExecution.optimizedPlan
+          spark.enableHyperspace()
+          val dfWithHyperspaceEnabled = query()
+          assert(!basePlan.equals(dfWithHyperspaceEnabled.queryExecution.optimizedPlan))
+          checkAnswer(dfWithHyperspaceDisabled, dfWithHyperspaceEnabled)
+        }
+      }
+    }
+  }
+
+  test("Verify Hybrid Scan is not applied when shuffle is not removed.") {
+    withTempPathAsString { testPath =>
+      val indexConfig = IndexConfig("indexRight", Seq("c2"), Seq("c4"))
+      val indexConfig2 = IndexConfig("indexLeft", Seq("c2"), Seq("c3"))
+      import spark.implicits._
+      SampleData.testData
+        .toDF("c1", "c2", "c3", "c4", "c5")
+        .limit(10)
+        .write
+        .parquet(testPath)
+      val df = spark.read.load(testPath)
+
+      withSQLConf(IndexConstants.INDEX_NUM_BUCKETS -> "11") {
+        // Create an index with bucket num 11.
+        hyperspace.createIndex(df, indexConfig)
+      }
+      withSQLConf(IndexConstants.INDEX_NUM_BUCKETS -> "12") {
+        // Create an index with bucket num 12.
+        hyperspace.createIndex(df, indexConfig2)
+      }
+
+      // Append to original data.
+      SampleData.testData
+        .toDF("c1", "c2", "c3", "c4", "c5")
+        .limit(3)
+        .write
+        .mode("append")
+        .parquet(testPath)
+
+      {
+        // Create a join query.
+        val df2 = spark.read.parquet(testPath)
+
+        def query(): DataFrame = {
+          df2.select("c2", "c3").join(df2.select("c2", "c4"), "c2")
+        }
+
+        val inputFiles = df.inputFiles
+        val appendedFiles = df2.inputFiles.diff(inputFiles).map(new Path(_))
+
+        spark.enableHyperspace()
+        withSQLConf(TestConfig.HybridScanEnabled: _*) {
+          def getShuffleCnt(sparkPlan: SparkPlanNode): Long = {
+            sparkPlan.collect { case _: ShuffleExchangeExec => true }.length
+          }
+          withSQLConf(IndexConstants.INDEX_HYBRID_SCAN_SHUFFLE_CHECK_ENABLED -> "false") {
+            val execPlan = query().queryExecution.executedPlan
+            val shuffleCnt = execPlan.collect {
+              case smj: SortMergeJoinExec =>
+                (getShuffleCnt(smj.left), getShuffleCnt(smj.right))
+            }.head
+            assert(shuffleCnt._1 === 1)
+            // Right child of join has 2 shuffle nodes because of Hybrid Scan for appended files.
+            assert(shuffleCnt._2 === 2)
+
+            // Verify indexes are used, and all index files are picked.
+            verifyIndexUsage(
+              query,
+              getIndexFilesPath(indexConfig.indexName, Seq(0)) ++ appendedFiles ++
+                getIndexFilesPath(indexConfig2.indexName, Seq(0)) ++ appendedFiles)
+          }
+          withSQLConf(IndexConstants.INDEX_HYBRID_SCAN_SHUFFLE_CHECK_ENABLED -> "true") {
+            val execPlan = query().queryExecution.executedPlan
+            val shuffleCnt = execPlan.collect {
+              case smj: SortMergeJoinExec =>
+                (getShuffleCnt(smj.left), getShuffleCnt(smj.right))
+            }.head
+            assert(shuffleCnt._1 === 1)
+            // One shuffle node of right child is removed with shuffle count check.
+            assert(shuffleCnt._2 === 1)
+
+            // For right child, indexRight can be still applied by FilterIndexRule.
+            verifyIndexUsage(
+              query,
+              getIndexFilesPath(indexConfig.indexName, Seq(0)) ++ appendedFiles ++
+                getIndexFilesPath(indexConfig2.indexName, Seq(0)) ++ appendedFiles)
+          }
         }
       }
     }