perf: optimize broadcast hash join with CollectLeft mode and decompression caching

Use PartitionMode::CollectLeft instead of Partitioned for broadcast hash
joins so DataFusion can optimize hash table construction for the broadcast
side. Also cache decompressed broadcast data at executor level to avoid
repeated LZ4 decompression across tasks.

Author: andygrove

native/core/src/execution/planner.rs

@@ -1663,14 +1663,20 @@ impl PhysicalPlanner {
                 let left = Arc::clone(&join_params.left.native_plan);
                 let right = Arc::clone(&join_params.right.native_plan);
 
+                let partition_mode = if join.is_broadcast {
+                    PartitionMode::CollectLeft
+                } else {
+                    PartitionMode::Partitioned
+                };
+
                 let hash_join = Arc::new(HashJoinExec::try_new(
                     left,
                     right,
                     join_params.join_on,
                     join_params.join_filter,
                     &join_params.join_type,
                     None,
-                    PartitionMode::Partitioned,
+                    partition_mode,
                     // null doesn't equal to null in Spark join key. If the join key is
                     // `EqualNullSafe`, Spark will rewrite it during planning.
                     NullEquality::NullEqualsNothing,
@@ -1688,7 +1694,7 @@ impl PhysicalPlanner {
                     ))
                 } else {
                     let swapped_hash_join =
-                        hash_join.as_ref().swap_inputs(PartitionMode::Partitioned)?;
+                        hash_join.as_ref().swap_inputs(partition_mode)?;
 
                     let mut additional_native_plans = vec![];
                     if swapped_hash_join.as_any().is::<ProjectionExec>() {
@@ -3905,6 +3911,7 @@ mod tests {
                 join_type: 0,
                 condition: None,
                 build_side: 0,
+                is_broadcast: false,
             })),
         };
 

native/proto/src/proto/operator.proto

@@ -334,6 +334,7 @@ message HashJoin {
   JoinType join_type = 3;
   optional spark.spark_expression.Expr condition = 4;
   BuildSide build_side = 5;
+  bool is_broadcast = 6;
 }
 
 message SortMergeJoin {

spark/src/main/scala/org/apache/spark/sql/comet/CometBroadcastExchangeExec.scala

@@ -19,19 +19,23 @@
 
 package org.apache.spark.sql.comet
 
+import java.io.{ByteArrayInputStream, ByteArrayOutputStream, DataInputStream}
+import java.nio.channels.Channels
 import java.util.UUID
-import java.util.concurrent.{Future, TimeoutException, TimeUnit}
+import java.util.concurrent.{ConcurrentHashMap, Future, TimeoutException, TimeUnit}
 
 import scala.concurrent.{ExecutionContext, Promise}
 import scala.concurrent.duration.NANOSECONDS
 import scala.util.control.NonFatal
 
-import org.apache.spark.{broadcast, Partition, SparkContext, SparkException, TaskContext}
+import org.apache.spark.{broadcast, Partition, SparkContext, SparkEnv, SparkException, TaskContext}
+import org.apache.spark.io.CompressionCodec
 import org.apache.spark.rdd.RDD
 import org.apache.spark.sql.catalyst.InternalRow
 import org.apache.spark.sql.catalyst.expressions.Attribute
 import org.apache.spark.sql.catalyst.plans.logical.Statistics
 import org.apache.spark.sql.catalyst.plans.physical.{BroadcastMode, BroadcastPartitioning, Partitioning}
+import org.apache.spark.sql.comet.execution.arrow.ArrowReaderIterator
 import org.apache.spark.sql.comet.util.Utils
 import org.apache.spark.sql.errors.QueryExecutionErrors
 import org.apache.spark.sql.execution.{ColumnarToRowExec, SparkPlan, SQLExecution}
@@ -311,8 +315,46 @@ class CometBatchRDD(
 
   override def compute(split: Partition, context: TaskContext): Iterator[ColumnarBatch] = {
     val partition = split.asInstanceOf[CometBatchPartition]
-    partition.value.value.toIterator
-      .flatMap(Utils.decodeBatches(_, this.getClass.getSimpleName))
+    val broadcastId = partition.value.id
+    val decompressedBytes = CometBatchRDD.decompressedCache.computeIfAbsent(
+      broadcastId,
+      _ => {
+        val codec = CompressionCodec.createCodec(SparkEnv.get.conf)
+        partition.value.value.map { chunkedBuffer =>
+          val cbbis = chunkedBuffer.toInputStream()
+          val ins = codec.compressedInputStream(cbbis)
+          val baos = new ByteArrayOutputStream()
+          val buf = new Array[Byte](8192)
+          var n = ins.read(buf)
+          while (n != -1) {
+            baos.write(buf, 0, n)
+            n = ins.read(buf)
+          }
+          ins.close()
+          baos.toByteArray
+        }
+      })
+    decompressedBytes.iterator.flatMap { bytes =>
+      new ArrowReaderIterator(
+        Channels.newChannel(new ByteArrayInputStream(bytes)),
+        this.getClass.getSimpleName)
+    }
+  }
+}
+
+object CometBatchRDD {
+
+  /**
+   * Executor-level cache of decompressed broadcast data keyed by broadcast ID. This avoids
+   * repeated LZ4 decompression when multiple tasks on the same executor process the same
+   * broadcast relation. Each entry stores decompressed Arrow IPC byte arrays.
+   */
+  private[comet] val decompressedCache =
+    new ConcurrentHashMap[Long, Array[Array[Byte]]]()
+
+  /** Invalidate cached decompressed data for a broadcast. */
+  def invalidateCache(broadcastId: Long): Unit = {
+    decompressedCache.remove(broadcastId)
   }
 }
 

spark/src/main/scala/org/apache/spark/sql/comet/operators.scala

@@ -1674,6 +1674,7 @@ trait CometHashJoin {
         .addAllRightJoinKeys(rightKeys.map(_.get).asJava)
         .setBuildSide(if (join.buildSide == BuildLeft) OperatorOuterClass.BuildSide.BuildLeft
         else OperatorOuterClass.BuildSide.BuildRight)
+        .setIsBroadcast(join.isInstanceOf[BroadcastHashJoinExec])
       condition.foreach(joinBuilder.setCondition)
       Some(builder.setHashJoin(joinBuilder).build())
     } else {

spark/src/test/scala/org/apache/spark/sql/benchmark/CometBroadcastHashJoinBenchmark.scala

@@ -0,0 +1,129 @@
+/*
+ * 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.benchmark
+
+import org.apache.spark.SparkConf
+import org.apache.spark.benchmark.Benchmark
+import org.apache.spark.sql.SparkSession
+import org.apache.spark.sql.internal.SQLConf
+
+import org.apache.comet.{CometConf, CometSparkSessionExtensions}
+
+/**
+ * Benchmark to measure Comet broadcast hash join performance. To run this benchmark:
+ * `SPARK_GENERATE_BENCHMARK_FILES=1 make
+ * benchmark-org.apache.spark.sql.benchmark.CometBroadcastHashJoinBenchmark` Results will be
+ * written to "spark/benchmarks/CometBroadcastHashJoinBenchmark-**results.txt".
+ */
+object CometBroadcastHashJoinBenchmark extends CometBenchmarkBase {
+  override def getSparkSession: SparkSession = {
+    val conf = new SparkConf()
+      .setAppName("CometBroadcastHashJoinBenchmark")
+      .set("spark.master", "local[5]")
+      .setIfMissing("spark.driver.memory", "3g")
+      .setIfMissing("spark.executor.memory", "3g")
+      .set("spark.executor.memoryOverhead", "10g")
+
+    val sparkSession = SparkSession.builder
+      .config(conf)
+      .withExtensions(new CometSparkSessionExtensions)
+      .getOrCreate()
+
+    sparkSession.conf.set(SQLConf.PARQUET_VECTORIZED_READER_ENABLED.key, "true")
+    sparkSession.conf.set(SQLConf.WHOLESTAGE_CODEGEN_ENABLED.key, "true")
+    sparkSession.conf.set(CometConf.COMET_ENABLED.key, "false")
+    sparkSession.conf.set(CometConf.COMET_EXEC_ENABLED.key, "false")
+    sparkSession.conf.set("parquet.enable.dictionary", "false")
+
+    sparkSession
+  }
+
+  def broadcastHashJoinBenchmark(
+      streamedRows: Int,
+      broadcastRows: Int,
+      joinType: String): Unit = {
+    val benchmark = new Benchmark(
+      s"Broadcast Hash Join ($joinType, stream=$streamedRows, broadcast=$broadcastRows)",
+      streamedRows,
+      output = output)
+
+    withTempPath { dir =>
+      import spark.implicits._
+
+      // Create streamed (large) table
+      val streamedDir = dir.getCanonicalPath + "/streamed"
+      spark
+        .range(streamedRows)
+        .select(($"id" % broadcastRows).as("key"), $"id".as("value"))
+        .write
+        .mode("overwrite")
+        .parquet(streamedDir)
+
+      // Create broadcast (small) table
+      val broadcastDir = dir.getCanonicalPath + "/broadcast"
+      spark
+        .range(broadcastRows)
+        .select($"id".as("key"), ($"id" * 10).as("payload"))
+        .write
+        .mode("overwrite")
+        .parquet(broadcastDir)
+
+      spark.read.parquet(streamedDir).createOrReplaceTempView("streamed")
+      spark.read.parquet(broadcastDir).createOrReplaceTempView("broadcast")
+
+      val query =
+        s"SELECT /*+ BROADCAST(broadcast) */ s.value, b.payload " +
+          s"FROM streamed s $joinType JOIN broadcast b ON s.key = b.key"
+
+      withTempTable("streamed", "broadcast") {
+        benchmark.addCase("Spark") { _ =>
+          withSQLConf(
+            CometConf.COMET_ENABLED.key -> "false",
+            SQLConf.AUTO_BROADCASTJOIN_THRESHOLD.key -> s"${256 * 1024 * 1024}") {
+            spark.sql(query).noop()
+          }
+        }
+
+        benchmark.addCase("Comet (Scan + Exec)") { _ =>
+          withSQLConf(
+            CometConf.COMET_ENABLED.key -> "true",
+            CometConf.COMET_EXEC_ENABLED.key -> "true",
+            SQLConf.AUTO_BROADCASTJOIN_THRESHOLD.key -> s"${256 * 1024 * 1024}") {
+            spark.sql(query).noop()
+          }
+        }
+
+        benchmark.run()
+      }
+    }
+  }
+
+  override def runCometBenchmark(mainArgs: Array[String]): Unit = {
+    val streamedRows = 2 * 1024 * 1024
+    val broadcastRows = 1000
+
+    for (joinType <- Seq("INNER", "LEFT", "RIGHT")) {
+      broadcastHashJoinBenchmark(streamedRows, broadcastRows, joinType)
+    }
+
+    // Test with larger broadcast table
+    broadcastHashJoinBenchmark(streamedRows, 10000, "INNER")
+  }
+}