From 1c32557f0cb086a2d54ca7054cf370fd1bfae969 Mon Sep 17 00:00:00 2001 From: Takuya UESHIN Date: Wed, 23 Dec 2020 14:48:01 -0800 Subject: [PATCH] [SPARK-33277][PYSPARK][SQL] Use ContextAwareIterator to stop consuming after the task ends ### What changes were proposed in this pull request? This is a retry of #30177. This is not a complete fix, but it would take long time to complete (#30242). As discussed offline, at least using `ContextAwareIterator` should be helpful enough for many cases. As the Python evaluation consumes the parent iterator in a separate thread, it could consume more data from the parent even after the task ends and the parent is closed. Thus, we should use `ContextAwareIterator` to stop consuming after the task ends. ### Why are the changes needed? Python/Pandas UDF right after off-heap vectorized reader could cause executor crash. E.g.,: ```py spark.range(0, 100000, 1, 1).write.parquet(path) spark.conf.set("spark.sql.columnVector.offheap.enabled", True) def f(x): return 0 fUdf = udf(f, LongType()) spark.read.parquet(path).select(fUdf('id')).head() ``` This is because, the Python evaluation consumes the parent iterator in a separate thread and it consumes more data from the parent even after the task ends and the parent is closed. If an off-heap column vector exists in the parent iterator, it could cause segmentation fault which crashes the executor. ### Does this PR introduce _any_ user-facing change? No. ### How was this patch tested? Added tests, and manually. Closes #30899 from ueshin/issues/SPARK-33277/context_aware_iterator. Authored-by: Takuya UESHIN Signed-off-by: Dongjoon Hyun (cherry picked from commit 5c9b421c3711ba373b4d5cbbd83a8ece91291ed0) Signed-off-by: Dongjoon Hyun --- .../apache/spark/ContextAwareIterator.scala | 40 +++++++++++++++++++ .../sql/execution/python/EvalPythonExec.scala | 5 ++- .../execution/python/MapInPandasExec.scala | 9 +++-- 3 files changed, 48 insertions(+), 6 deletions(-) create mode 100644 core/src/main/scala/org/apache/spark/ContextAwareIterator.scala diff --git a/core/src/main/scala/org/apache/spark/ContextAwareIterator.scala b/core/src/main/scala/org/apache/spark/ContextAwareIterator.scala new file mode 100644 index 0000000000000..c4d0dd8aceab0 --- /dev/null +++ b/core/src/main/scala/org/apache/spark/ContextAwareIterator.scala @@ -0,0 +1,40 @@ +/* + * 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 + +import org.apache.spark.annotation.DeveloperApi + +/** + * :: DeveloperApi :: + * A TaskContext aware iterator. + * + * As the Python evaluation consumes the parent iterator in a separate thread, + * it could consume more data from the parent even after the task ends and the parent is closed. + * If an off-heap access exists in the parent iterator, it could cause segmentation fault + * which crashes the executor. + * Thus, we should use [[ContextAwareIterator]] to stop consuming after the task ends. + */ +@DeveloperApi +class ContextAwareIterator[+T](val context: TaskContext, val delegate: Iterator[T]) + extends Iterator[T] { + + override def hasNext: Boolean = + !context.isCompleted() && !context.isInterrupted() && delegate.hasNext + + override def next(): T = delegate.next() +} diff --git a/sql/core/src/main/scala/org/apache/spark/sql/execution/python/EvalPythonExec.scala b/sql/core/src/main/scala/org/apache/spark/sql/execution/python/EvalPythonExec.scala index 7c476ab03c002..fca43e454bff5 100644 --- a/sql/core/src/main/scala/org/apache/spark/sql/execution/python/EvalPythonExec.scala +++ b/sql/core/src/main/scala/org/apache/spark/sql/execution/python/EvalPythonExec.scala @@ -21,7 +21,7 @@ import java.io.File import scala.collection.mutable.ArrayBuffer -import org.apache.spark.{SparkEnv, TaskContext} +import org.apache.spark.{ContextAwareIterator, SparkEnv, TaskContext} import org.apache.spark.api.python.ChainedPythonFunctions import org.apache.spark.rdd.RDD import org.apache.spark.sql.catalyst.InternalRow @@ -89,6 +89,7 @@ trait EvalPythonExec extends UnaryExecNode { inputRDD.mapPartitions { iter => val context = TaskContext.get() + val contextAwareIterator = new ContextAwareIterator(context, iter) // The queue used to buffer input rows so we can drain it to // combine input with output from Python. @@ -120,7 +121,7 @@ trait EvalPythonExec extends UnaryExecNode { }.toSeq) // Add rows to queue to join later with the result. - val projectedRowIter = iter.map { inputRow => + val projectedRowIter = contextAwareIterator.map { inputRow => queue.add(inputRow.asInstanceOf[UnsafeRow]) projection(inputRow) } diff --git a/sql/core/src/main/scala/org/apache/spark/sql/execution/python/MapInPandasExec.scala b/sql/core/src/main/scala/org/apache/spark/sql/execution/python/MapInPandasExec.scala index 2bb808119c0ae..71f51f1abc6f5 100644 --- a/sql/core/src/main/scala/org/apache/spark/sql/execution/python/MapInPandasExec.scala +++ b/sql/core/src/main/scala/org/apache/spark/sql/execution/python/MapInPandasExec.scala @@ -19,7 +19,7 @@ package org.apache.spark.sql.execution.python import scala.collection.JavaConverters._ -import org.apache.spark.TaskContext +import org.apache.spark.{ContextAwareIterator, TaskContext} import org.apache.spark.api.python.{ChainedPythonFunctions, PythonEvalType} import org.apache.spark.rdd.RDD import org.apache.spark.sql.catalyst.InternalRow @@ -61,16 +61,17 @@ case class MapInPandasExec( val pythonRunnerConf = ArrowUtils.getPythonRunnerConfMap(conf) val outputTypes = child.schema + val context = TaskContext.get() + val contextAwareIterator = new ContextAwareIterator(context, inputIter) + // Here we wrap it via another row so that Python sides understand it // as a DataFrame. - val wrappedIter = inputIter.map(InternalRow(_)) + val wrappedIter = contextAwareIterator.map(InternalRow(_)) // DO NOT use iter.grouped(). See BatchIterator. val batchIter = if (batchSize > 0) new BatchIterator(wrappedIter, batchSize) else Iterator(wrappedIter) - val context = TaskContext.get() - val columnarBatchIter = new ArrowPythonRunner( chainedFunc, PythonEvalType.SQL_MAP_PANDAS_ITER_UDF,