Merge pull request NVIDIA#10698 from NVIDIA/branch-24.04

[auto-merge] branch-24.04 to branch-24.06 [skip ci] [bot]

integration_tests/src/main/python/aqe_test.py

@@ -298,3 +298,40 @@ def do_it(spark):
 
     assert_gpu_and_cpu_are_equal_collect(do_it, conf=bhj_disable_conf)
 
+
+# See https://github.com/NVIDIA/spark-rapids/issues/10645. Sometimes the exchange can provide multiple
+# batches, so we to coalesce them into a single batch for the broadcast hash join.
+@ignore_order(local=True)
+@pytest.mark.skipif(not (is_databricks_runtime()), \
+    reason="Executor side broadcast only supported on Databricks")
+def test_aqe_join_executor_broadcast_enforce_single_batch():
+    # Use a small batch to see if Databricks could send multiple batches
+    conf = copy_and_update(_adaptive_conf, { "spark.rapids.sql.batchSizeBytes": "25" })
+    def prep(spark):
+        id_gen = RepeatSeqGen(IntegerGen(nullable=False), length=250)
+        name_gen = RepeatSeqGen(["Adam", "Bob", "Cathy"], data_type=StringType())
+        school_gen = RepeatSeqGen(["School1", "School2", "School3"], data_type=StringType())
+
+        df = gen_df(spark, StructGen([('id', id_gen), ('name', name_gen)], nullable=False), length=1000)
+        df.createOrReplaceTempView("df")
+
+        df_school = gen_df(spark, StructGen([('id', id_gen), ('school', school_gen)], nullable=False), length=250)
+        df.createOrReplaceTempView("df_school")
+
+    with_cpu_session(prep)
+
+    def do_it(spark):
+        res = spark.sql(
+            """
+                select /*+ BROADCAST(df_school) */ * from df, df_school where df.id == df_school.id
+            """
+        )
+        res.explain()
+        return res
+    # Ensure this is an EXECUTOR_BROADCAST
+    assert_cpu_and_gpu_are_equal_collect_with_capture(
+        do_it, 
+        exist_classes="GpuShuffleExchangeExec,GpuBroadcastHashJoinExec",
+        non_exist_classes="GpuBroadcastExchangeExec",
+        conf=conf)
+

sql-plugin/src/main/spark330db/scala/org/apache/spark/sql/rapids/execution/GpuExecutorBroadcastHelper.scala

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2023, NVIDIA CORPORATION.
+ * Copyright (c) 2023-2024, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -21,7 +21,7 @@
 spark-rapids-shim-json-lines ***/
 package org.apache.spark.sql.rapids.execution
 
-import com.nvidia.spark.rapids.{ConcatAndConsumeAll, GpuColumnVector, GpuMetric, GpuShuffleCoalesceIterator, HostShuffleCoalesceIterator}
+import com.nvidia.spark.rapids.{ConcatAndConsumeAll, GpuCoalesceIterator, GpuColumnVector, GpuMetric, GpuShuffleCoalesceIterator, HostShuffleCoalesceIterator, NoopMetric, RequireSingleBatch}
 import com.nvidia.spark.rapids.Arm.withResource
 
 import org.apache.spark.TaskContext
@@ -40,6 +40,7 @@ import org.apache.spark.sql.vectorized.ColumnarBatch
  * which means they require a format of the data that can be used on the GPU.
  */
 object GpuExecutorBroadcastHelper {
+  import GpuMetric._
 
   // This reads the shuffle data that we have retrieved using `getShuffleRDD` from the shuffle
   // exchange. WARNING: Do not use this method outside of this context. This method can only be
@@ -65,11 +66,28 @@ object GpuExecutorBroadcastHelper {
     // Use the GPU Shuffle Coalesce iterator to concatenate and load batches onto the 
     // host as needed. Since we don't have GpuShuffleCoalesceExec in the plan for the 
     // executor broadcast scenario, we have to use that logic here to efficiently 
-    // grab and release the semaphore while doing I/O
+    // grab and release the semaphore while doing I/O. We wrap this with GpuCoalesceIterator
+    // to ensure this always a single batch for the following step.
+    val shuffleMetrics = Map(
+      CONCAT_TIME -> metricsMap(CONCAT_TIME),
+      OP_TIME -> metricsMap(OP_TIME)
+    ).withDefaultValue(NoopMetric)
+
     val iter = shuffleDataIterator(shuffleData)
-    new GpuShuffleCoalesceIterator(
-      new HostShuffleCoalesceIterator(iter, targetSize, metricsMap),
-      dataTypes, metricsMap).asInstanceOf[Iterator[ColumnarBatch]]
+    new GpuCoalesceIterator(
+      new GpuShuffleCoalesceIterator(
+        new HostShuffleCoalesceIterator(iter, targetSize, shuffleMetrics),
+        dataTypes, shuffleMetrics).asInstanceOf[Iterator[ColumnarBatch]],
+      dataTypes,
+      RequireSingleBatch,
+      NoopMetric, // numInputRows
+      NoopMetric, // numInputBatches
+      NoopMetric, // numOutputRows
+      NoopMetric, // numOutputBatches
+      NoopMetric, // collectTime
+      metricsMap(CONCAT_TIME), // concatTime
+      metricsMap(OP_TIME), // opTime
+      "GpuBroadcastHashJoinExec").asInstanceOf[Iterator[ColumnarBatch]]
   }
 
   /**