Support columnar processing for FlatMapCoGroupInPandas[databricks] (#…

…6751) * Support columnar processing for FlatMapCoGroupInPandas Signed-off-by: firestarman <[email protected]>
NVIDIA · Oct 14, 2022 · 03b1164 · 03b1164
1 parent 4777935
commit 03b1164
Show file tree

Hide file tree

Showing 24 changed files with 760 additions and 590 deletions.
diff --git a/docs/configs.md b/docs/configs.md
@@ -415,7 +415,7 @@ Name | Description | Default Value | Notes
 <a name="sql.exec.SortMergeJoinExec"></a>spark.rapids.sql.exec.SortMergeJoinExec|Sort merge join, replacing with shuffled hash join|true|None|
 <a name="sql.exec.AggregateInPandasExec"></a>spark.rapids.sql.exec.AggregateInPandasExec|The backend for an Aggregation Pandas UDF, this accelerates the data transfer between the Java process and the Python process. It also supports scheduling GPU resources for the Python process when enabled.|true|None|
 <a name="sql.exec.ArrowEvalPythonExec"></a>spark.rapids.sql.exec.ArrowEvalPythonExec|The backend of the Scalar Pandas UDFs. Accelerates the data transfer between the Java process and the Python process. It also supports scheduling GPU resources for the Python process when enabled|true|None|
-<a name="sql.exec.FlatMapCoGroupsInPandasExec"></a>spark.rapids.sql.exec.FlatMapCoGroupsInPandasExec|The backend for CoGrouped Aggregation Pandas UDF, it runs on CPU itself now but supports scheduling GPU resources for the Python process when enabled|false|This is disabled by default because Performance is not ideal now|
+<a name="sql.exec.FlatMapCoGroupsInPandasExec"></a>spark.rapids.sql.exec.FlatMapCoGroupsInPandasExec|The backend for CoGrouped Aggregation Pandas UDF. Accelerates the data transfer between the Java process and the Python process. It also supports scheduling GPU resources for the Python process when enabled.|false|This is disabled by default because Performance is not ideal with many small groups|
 <a name="sql.exec.FlatMapGroupsInPandasExec"></a>spark.rapids.sql.exec.FlatMapGroupsInPandasExec|The backend for Flat Map Groups Pandas UDF, Accelerates the data transfer between the Java process and the Python process. It also supports scheduling GPU resources for the Python process when enabled.|true|None|
 <a name="sql.exec.MapInPandasExec"></a>spark.rapids.sql.exec.MapInPandasExec|The backend for Map Pandas Iterator UDF. Accelerates the data transfer between the Java process and the Python process. It also supports scheduling GPU resources for the Python process when enabled.|true|None|
 <a name="sql.exec.WindowInPandasExec"></a>spark.rapids.sql.exec.WindowInPandasExec|The backend for Window Aggregation Pandas UDF, Accelerates the data transfer between the Java process and the Python process. It also supports scheduling GPU resources for the Python process when enabled. For now it only supports row based window frame.|false|This is disabled by default because it only supports row based frame for now|

diff --git a/docs/supported_ops.md b/docs/supported_ops.md
@@ -1132,6 +1132,30 @@ Accelerator supports are described below.
 <td><b>NS</b></td>
 </tr>
 <tr>
+<td rowspan="1">FlatMapCoGroupsInPandasExec</td>
+<td rowspan="1">The backend for CoGrouped Aggregation Pandas UDF. Accelerates the data transfer between the Java process and the Python process. It also supports scheduling GPU resources for the Python process when enabled.</td>
+<td rowspan="1">This is disabled by default because Performance is not ideal with many small groups</td>
+<td>Input/Output</td>
+<td>S</td>
+<td>S</td>
+<td>S</td>
+<td>S</td>
+<td>S</td>
+<td>S</td>
+<td>S</td>
+<td>S</td>
+<td><em>PS<br/>UTC is only supported TZ for TIMESTAMP</em></td>
+<td>S</td>
+<td><b>NS</b></td>
+<td><b>NS</b></td>
+<td><b>NS</b></td>
+<td><b>NS</b></td>
+<td><b>NS</b></td>
+<td><b>NS</b></td>
+<td><b>NS</b></td>
+<td><b>NS</b></td>
+</tr>
+<tr>
 <td rowspan="1">FlatMapGroupsInPandasExec</td>
 <td rowspan="1">The backend for Flat Map Groups Pandas UDF, Accelerates the data transfer between the Java process and the Python process. It also supports scheduling GPU resources for the Python process when enabled.</td>
 <td rowspan="1">None</td>
@@ -1204,6 +1228,30 @@ Accelerator supports are described below.
 <td><b>NS</b></td>
 </tr>
 <tr>
+<th>Executor</th>
+<th>Description</th>
+<th>Notes</th>
+<th>Param(s)</th>
+<th>BOOLEAN</th>
+<th>BYTE</th>
+<th>SHORT</th>
+<th>INT</th>
+<th>LONG</th>
+<th>FLOAT</th>
+<th>DOUBLE</th>
+<th>DATE</th>
+<th>TIMESTAMP</th>
+<th>STRING</th>
+<th>DECIMAL</th>
+<th>NULL</th>
+<th>BINARY</th>
+<th>CALENDAR</th>
+<th>ARRAY</th>
+<th>MAP</th>
+<th>STRUCT</th>
+<th>UDT</th>
+</tr>
+<tr>
 <td rowspan="2">WindowExec</td>
 <td rowspan="2">Window-operator backend</td>
 <td rowspan="2">None</td>

diff --git a/integration_tests/src/main/python/udf_cudf_test.py b/integration_tests/src/main/python/udf_cudf_test.py
@@ -1,4 +1,4 @@
-# Copyright (c) 2020-2021, NVIDIA CORPORATION.
+# Copyright (c) 2020-2022, NVIDIA CORPORATION.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -37,7 +37,7 @@
 from pyspark.sql import Window
 from pyspark.sql.functions import pandas_udf, PandasUDFType
 from spark_session import with_cpu_session, with_gpu_session
-from marks import allow_non_gpu, cudf_udf
+from marks import cudf_udf
 
 
 _conf = {
@@ -300,7 +300,6 @@ def gpu_run(spark):
 
 
 # ======= Test CoGroup Map In Pandas =======
-@allow_non_gpu('GpuFlatMapCoGroupsInPandasExec','PythonUDF')
 @cudf_udf
 def test_cogroup(enable_cudf_udf):
     def cpu_run(spark):

diff --git a/integration_tests/src/main/python/udf_test.py b/integration_tests/src/main/python/udf_test.py
@@ -32,7 +32,7 @@
         raise AssertionError("incorrect pyarrow version during required testing " + str(e))
     pytestmark = pytest.mark.skip(reason=str(e))
 
-from asserts import assert_gpu_and_cpu_are_equal_collect
+from asserts import assert_gpu_and_cpu_are_equal_collect, assert_gpu_fallback_collect
 from data_gen import *
 from marks import incompat, approximate_float, allow_non_gpu, ignore_order
 from pyspark.sql import Window
@@ -44,7 +44,7 @@
 arrow_udf_conf = {
     'spark.sql.execution.arrow.pyspark.enabled': 'true',
     'spark.rapids.sql.exec.WindowInPandasExec': 'true',
-    'spark.rapids.sql.exec.AggregateInPandasExec': 'true'
+    'spark.rapids.sql.exec.FlatMapCoGroupsInPandasExec': 'true'
 }
 
 data_gens_nested_for_udf = arrow_array_gens + arrow_struct_gens
@@ -323,7 +323,6 @@ def create_df(spark, data_gen, left_length, right_length):
 
 
 @ignore_order
-@allow_non_gpu('FlatMapCoGroupsInPandasExec', 'PythonUDF', 'Alias')
 @pytest.mark.parametrize('data_gen', [ShortGen(nullable=False)], ids=idfn)
 def test_cogroup_apply_udf(data_gen):
     def asof_join(l, r):
@@ -335,3 +334,18 @@ def do_it(spark):
                 right.groupby('a')).applyInPandas(
                         asof_join, schema="a int, b int")
     assert_gpu_and_cpu_are_equal_collect(do_it, conf=arrow_udf_conf)
+
+
+@ignore_order
+@allow_non_gpu('FlatMapCoGroupsInPandasExec')
+def test_cogroup_apply_fallback():
+    def asof_join(l, r):
+        return r
+
+    def do_it(spark):
+        left = two_col_df(spark, int_gen, int_gen, length=100)
+        right = two_col_df(spark, short_gen, int_gen, length=100)
+        return left.groupby('a').cogroup(
+                right.groupby('a')).applyInPandas(
+                        asof_join, schema="a int, b int")
+    assert_gpu_fallback_collect(do_it, 'FlatMapCoGroupsInPandasExec', conf=arrow_udf_conf)
diff --git a/sql-plugin/src/main/311+-db/scala/com/nvidia/spark/rapids/shims/GpuWindowInPandasExec.scala b/sql-plugin/src/main/311+-db/scala/com/nvidia/spark/rapids/shims/GpuWindowInPandasExec.scala
@@ -19,15 +19,14 @@ package com.nvidia.spark.rapids.shims
 import scala.collection.mutable.ArrayBuffer
 
 import com.nvidia.spark.rapids._
-import com.nvidia.spark.rapids.GpuMetric._
 import com.nvidia.spark.rapids.python.PythonWorkerSemaphore
 
 import org.apache.spark.TaskContext
 import org.apache.spark.api.python.PythonEvalType
 import org.apache.spark.rdd.RDD
 import org.apache.spark.sql.catalyst.expressions._
 import org.apache.spark.sql.execution.SparkPlan
-import org.apache.spark.sql.rapids.execution.python.{BatchQueue, GpuArrowPythonRunner, GpuPythonHelper, GpuPythonUDF, GpuWindowInPandasExecBase, GroupingIterator}
+import org.apache.spark.sql.rapids.execution.python.{BatchQueue, CombiningIterator, GpuArrowPythonRunner, GpuPythonHelper, GpuPythonUDF, GpuWindowInPandasExecBase, GroupingIterator}
 import org.apache.spark.sql.types.{IntegerType, StructField, StructType}
 import org.apache.spark.sql.util.ArrowUtils
 import org.apache.spark.sql.vectorized.ColumnarBatch
@@ -86,11 +85,8 @@ case class GpuWindowInPandasExec(
 
   // Override doExecuteColumnar so we use the correct GpuArrowPythonRunner
   override protected def doExecuteColumnar(): RDD[ColumnarBatch] = {
-    val numInputRows = gpuLongMetric(NUM_INPUT_ROWS)
-    val numInputBatches = gpuLongMetric(NUM_INPUT_BATCHES)
-    val numOutputRows = gpuLongMetric(NUM_OUTPUT_ROWS)
-    val numOutputBatches = gpuLongMetric(NUM_OUTPUT_BATCHES)
-    val spillCallback = GpuMetric.makeSpillCallback(allMetrics)
+    val (numInputRows, numInputBatches, numOutputRows, numOutputBatches,
+         spillCallback) = commonGpuMetrics()
     val sessionLocalTimeZone = conf.sessionLocalTimeZone
 
     // 1) Unwrap the expressions and build some info data:
@@ -228,45 +224,12 @@ case class GpuWindowInPandasExec(
           /* The whole group data should be written in a single call, so here is unlimited */
           Int.MaxValue,
           spillCallback.semaphoreWaitTime,
-          () => queue.finish(),
-          pythonOutputSchema)
+          pythonOutputSchema,
+          () => queue.finish())
 
-        val outputBatchIterator = pyRunner.compute(pyInputIterator, context.partitionId(), context)
-        new Iterator[ColumnarBatch] {
-          // for hasNext we are waiting on the queue to have something inserted into it
-          // instead of waiting for a result to be ready from python. The reason for this
-          // is to let us know the target number of rows in the batch that we want when reading.
-          // It is a bit hacked up but it works. In the future when we support spilling we should
-          // store the number of rows separate from the batch. That way we can get the target batch
-          // size out without needing to grab the GpuSemaphore which we cannot do if we might block
-          // on a read operation.
-          // Besides, when the queue is empty, need to call the `hasNext` of the out iterator to
-          // trigger reading and handling the control data followed with the stream data.
-          override def hasNext: Boolean = queue.hasNext || outputBatchIterator.hasNext
-
-          private [this] def combine(
-                                      origBatch: ColumnarBatch,
-                                      retBatch: ColumnarBatch): ColumnarBatch = {
-            val lColumns = GpuColumnVector.extractColumns(origBatch)
-            val rColumns = GpuColumnVector.extractColumns(retBatch)
-            new ColumnarBatch(lColumns.map(_.incRefCount()) ++ rColumns.map(_.incRefCount()),
-              origBatch.numRows())
-          }
-
-          override def next(): ColumnarBatch = {
-            val numRows = queue.peekBatchSize
-            // Update the expected batch size for next read
-            pyRunner.minReadTargetBatchSize = numRows
-            withResource(outputBatchIterator.next()) { cbFromPython =>
-              assert(cbFromPython.numRows() == numRows)
-              withResource(queue.remove()) { origBatch =>
-                numOutputBatches += 1
-                numOutputRows += numRows
-                projectResult(combine(origBatch, cbFromPython))
-              }
-            }
-          }
-        } // End of new Iterator
+        val outputIterator = pyRunner.compute(pyInputIterator, context.partitionId(), context)
+        new CombiningIterator(queue, outputIterator, pyRunner, numOutputRows,
+          numOutputBatches).map(projectResult(_))
       } else {
         // Empty partition, return the input iterator directly
         inputIter

diff --git a/...ala/org/apache/spark/sql/rapids/execution/python/shims/GpuGroupUDFArrowPythonRunner.scala b/...ala/org/apache/spark/sql/rapids/execution/python/shims/GpuGroupUDFArrowPythonRunner.scala
@@ -28,7 +28,7 @@ import com.nvidia.spark.rapids._
 import org.apache.spark.{SparkEnv, TaskContext}
 import org.apache.spark.api.python._
 import org.apache.spark.sql.execution.python.PythonUDFRunner
-import org.apache.spark.sql.rapids.execution.python.{BufferToStreamWriter, GpuArrowPythonRunner}
+import org.apache.spark.sql.rapids.execution.python.{BufferToStreamWriter, GpuArrowPythonRunner, GpuPythonArrowOutput, GpuPythonRunnerBase}
 import org.apache.spark.sql.types._
 import org.apache.spark.sql.vectorized.ColumnarBatch
 import org.apache.spark.util.Utils
@@ -54,12 +54,10 @@ class GpuGroupUDFArrowPythonRunner(
     timeZoneId: String,
     conf: Map[String, String],
     batchSize: Long,
-    semWait: GpuMetric,
-    onDataWriteFinished: () => Unit,
-    override val pythonOutSchema: StructType,
-    minReadTargetBatchSize: Int)
-    extends GpuArrowPythonRunner(funcs, evalType, argOffsets, pythonInSchema, timeZoneId, conf,
-      batchSize, semWait, onDataWriteFinished, pythonOutSchema, minReadTargetBatchSize) {
+    val semWait: GpuMetric,
+    val pythonOutSchema: StructType)
+  extends GpuPythonRunnerBase[ColumnarBatch](funcs, evalType, argOffsets)
+    with GpuPythonArrowOutput {
 
   protected override def newWriterThread(
       env: SparkEnv,
@@ -91,7 +89,7 @@ class GpuGroupUDFArrowPythonRunner(
             GpuSemaphore.releaseIfNecessary(TaskContext.get())
           })
           // Flatten the names of nested struct columns, required by cudf Arrow IPC writer.
-          flattenNames(pythonInSchema).foreach { case (name, nullable) =>
+          GpuArrowPythonRunner.flattenNames(pythonInSchema).foreach { case (name, nullable) =>
               if (nullable) {
                 builder.withColumnNames(name)
               } else {
@@ -121,19 +119,8 @@ class GpuGroupUDFArrowPythonRunner(
           // tell serializer we are done
           dataOut.writeInt(0)
           dataOut.flush()
-          if (onDataWriteFinished != null) onDataWriteFinished()
         }
       }
-
-      private def flattenNames(d: DataType, nullable: Boolean=true): Seq[(String, Boolean)] =
-        d match {
-          case s: StructType =>
-            s.flatMap(sf => Seq((sf.name, sf.nullable)) ++ flattenNames(sf.dataType, sf.nullable))
-          case m: MapType =>
-            flattenNames(m.keyType, nullable) ++ flattenNames(m.valueType, nullable)
-          case a: ArrayType => flattenNames(a.elementType, nullable)
-          case _ => Nil
-      }
     }
   }
 }
diff --git a/...ala/org/apache/spark/sql/rapids/execution/python/shims/GpuFlatMapGroupsInPandasExec.scala b/...ala/org/apache/spark/sql/rapids/execution/python/shims/GpuFlatMapGroupsInPandasExec.scala
@@ -149,13 +149,7 @@ case class GpuFlatMapGroupsInPandasExec(
           // The whole group data should be written in a single call, so here is unlimited
           Int.MaxValue,
           spillCallback.semaphoreWaitTime,
-          onDataWriteFinished = null,
-          pythonOutputSchema,
-          // We can not assert the result batch from Python has the same row number with the
-          // input batch. Because Grouped Map UDF allows the output of arbitrary length.
-          // So try to read as many as possible by specifying `minReadTargetBatchSize` as
-          // `Int.MaxValue` here.
-          Int.MaxValue)
+          pythonOutputSchema)
 
         executePython(pyInputIter, localOutput, pyRunner, mNumOutputRows, mNumOutputBatches)
       } else {

diff --git a/...1until320-nondb/scala/org/apache/spark/rapids/shims/api/python/ShimBasePythonRunner.scala b/...1until320-nondb/scala/org/apache/spark/rapids/shims/api/python/ShimBasePythonRunner.scala
@@ -24,13 +24,12 @@ import com.nvidia.spark.rapids.Arm
 
 import org.apache.spark.{SparkEnv, TaskContext}
 import org.apache.spark.api.python.BasePythonRunner
-import org.apache.spark.sql.vectorized.ColumnarBatch
 
 // pid is not a constructor argument in 30x and 31x
 abstract class ShimBasePythonRunner[IN, OUT](
     funcs : scala.Seq[org.apache.spark.api.python.ChainedPythonFunctions],
     evalType : scala.Int, argOffsets : scala.Array[scala.Array[scala.Int]]
-) extends BasePythonRunner[ColumnarBatch, ColumnarBatch](funcs, evalType, argOffsets)
+) extends BasePythonRunner[IN, OUT](funcs, evalType, argOffsets)
     with Arm {
   protected abstract class ShimReaderIterator(
     stream: DataInputStream,

diff --git a/.../src/main/31xdb/scala/org/apache/spark/rapids/shims/api/python/ShimBasePythonRunner.scala b/.../src/main/31xdb/scala/org/apache/spark/rapids/shims/api/python/ShimBasePythonRunner.scala
@@ -24,13 +24,12 @@ import com.nvidia.spark.rapids.Arm
 
 import org.apache.spark.{SparkEnv, TaskContext}
 import org.apache.spark.api.python.BasePythonRunner
-import org.apache.spark.sql.vectorized.ColumnarBatch
 
 // pid is not a constructor argument in 30x and 31x
 abstract class ShimBasePythonRunner[IN, OUT](
     funcs : scala.Seq[org.apache.spark.api.python.ChainedPythonFunctions],
     evalType : scala.Int, argOffsets : scala.Array[scala.Array[scala.Int]]
-) extends BasePythonRunner[ColumnarBatch, ColumnarBatch](funcs, evalType, argOffsets)
+) extends BasePythonRunner[IN, OUT](funcs, evalType, argOffsets)
     with Arm {
   protected abstract class ShimReaderIterator(
     stream: DataInputStream,

diff --git a/...ala/org/apache/spark/sql/rapids/execution/python/shims/GpuFlatMapGroupsInPandasExec.scala b/...ala/org/apache/spark/sql/rapids/execution/python/shims/GpuFlatMapGroupsInPandasExec.scala
@@ -154,13 +154,7 @@ case class GpuFlatMapGroupsInPandasExec(
             // The whole group data should be written in a single call, so here is unlimited
             Int.MaxValue,
             spillCallback.semaphoreWaitTime,
-            onDataWriteFinished = null,
-            pythonOutputSchema,
-            // We can not assert the result batch from Python has the same row number with the
-            // input batch. Because Grouped Map UDF allows the output of arbitrary length.
-            // So try to read as many as possible by specifying `minReadTargetBatchSize` as
-            // `Int.MaxValue` here.
-            Int.MaxValue)
+            pythonOutputSchema)
         } else {
           new GpuArrowPythonRunner(
             chainedFunc,
@@ -171,9 +165,7 @@ case class GpuFlatMapGroupsInPandasExec(
             pythonRunnerConf,
             Int.MaxValue,
             spillCallback.semaphoreWaitTime,
-            onDataWriteFinished = null,
-            pythonOutputSchema,
-            Int.MaxValue)
+            pythonOutputSchema)
         }
 
         executePython(pyInputIter, localOutput, pyRunner, mNumOutputRows, mNumOutputBatches)

diff --git a/...n/src/main/320+/scala/org/apache/spark/rapids/shims/api/python/ShimBasePythonRunner.scala b/...n/src/main/320+/scala/org/apache/spark/rapids/shims/api/python/ShimBasePythonRunner.scala
@@ -24,12 +24,11 @@ import com.nvidia.spark.rapids.Arm
 
 import org.apache.spark.{SparkEnv, TaskContext}
 import org.apache.spark.api.python.BasePythonRunner
-import org.apache.spark.sql.vectorized.ColumnarBatch
 
 abstract class ShimBasePythonRunner[IN, OUT](
     funcs : scala.Seq[org.apache.spark.api.python.ChainedPythonFunctions],
     evalType : scala.Int, argOffsets : scala.Array[scala.Array[scala.Int]]
-) extends BasePythonRunner[ColumnarBatch, ColumnarBatch](funcs, evalType, argOffsets)
+) extends BasePythonRunner[IN, OUT](funcs, evalType, argOffsets)
     with Arm {
   protected abstract class ShimReaderIterator(
     stream: DataInputStream,

diff --git a/...ala/org/apache/spark/sql/rapids/execution/python/shims/GpuFlatMapGroupsInPandasExec.scala b/...ala/org/apache/spark/sql/rapids/execution/python/shims/GpuFlatMapGroupsInPandasExec.scala
@@ -155,13 +155,7 @@ case class GpuFlatMapGroupsInPandasExec(
             // The whole group data should be written in a single call, so here is unlimited
             Int.MaxValue,
             spillCallback.semaphoreWaitTime,
-            onDataWriteFinished = null,
-            pythonOutputSchema,
-            // We can not assert the result batch from Python has the same row number with the
-            // input batch. Because Grouped Map UDF allows the output of arbitrary length.
-            // So try to read as many as possible by specifying `minReadTargetBatchSize` as
-            // `Int.MaxValue` here.
-            Int.MaxValue)
+            pythonOutputSchema)
         } else {
           new GpuArrowPythonRunner(
             chainedFunc,
@@ -172,9 +166,7 @@ case class GpuFlatMapGroupsInPandasExec(
             pythonRunnerConf,
             Int.MaxValue,
             spillCallback.semaphoreWaitTime,
-            onDataWriteFinished = null,
-            pythonOutputSchema,
-            Int.MaxValue)
+            pythonOutputSchema)
         }
 
         executePython(pyInputIter, localOutput, pyRunner, mNumOutputRows, mNumOutputBatches)