[SPARK-29649][SQL] Stop task set if FileAlreadyExistsException was thrown when writing to output file

### What changes were proposed in this pull request?

We already know task attempts that do not clean up output files in staging directory can cause job failure (SPARK-27194). There was proposals trying to fix it by changing output filename, or deleting existing output files. These proposals are not reliable completely.

The difficulty is, as previous failed task attempt wrote the output file, at next task attempt the output file is still under same staging directory, even the output file name is different.

If the job will go to fail eventually, there is no point to re-run the task until max attempts are reached. For the jobs running a lot of time, re-running the task can waste a lot of time.

This patch proposes to let Spark detect such file already exist exception and stop the task set early.

### Why are the changes needed?

For now, if FileAlreadyExistsException is thrown during data writing job in SQL, the job will continue re-running task attempts until max failure number is reached. It is no point for re-running tasks as task attempts will also fail because they can not write to the existing file too. We should stop the task set early.

### Does this PR introduce any user-facing change?

Yes. If FileAlreadyExistsException is thrown during data writing job in SQL, no more task attempts are re-tried and the task set will be stoped early.

### How was this patch tested?

Unit test.

Closes apache#26312 from viirya/stop-taskset-if-outputfile-exists.

Authored-by: Liang-Chi Hsieh <[email protected]>
Signed-off-by: Dongjoon Hyun <[email protected]>

Author: viirya

core/src/main/scala/org/apache/spark/TaskOutputFileAlreadyExistException.scala

@@ -0,0 +1,23 @@
+/*
+ * 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
+
+/**
+ * Exception thrown when a task cannot write to output file due to the file already exists.
+ */
+private[spark] class TaskOutputFileAlreadyExistException(error: Throwable) extends Exception(error)

core/src/main/scala/org/apache/spark/scheduler/TaskSetManager.scala

@@ -799,6 +799,15 @@ private[spark] class TaskSetManager(
             info.id, taskSet.id, tid, ef.description))
           return
         }
+        if (ef.className == classOf[TaskOutputFileAlreadyExistException].getName) {
+          // If we can not write to output file in the task, there's no point in trying to
+          // re-execute it.
+          logError("Task %s in stage %s (TID %d) can not write to output file: %s; not retrying"
+            .format(info.id, taskSet.id, tid, ef.description))
+          abort("Task %s in stage %s (TID %d) can not write to output file: %s".format(
+            info.id, taskSet.id, tid, ef.description))
+          return
+        }
         val key = ef.description
         val now = clock.getTimeMillis()
         val (printFull, dupCount) = {

core/src/test/scala/org/apache/spark/scheduler/TaskSetManagerSuite.scala

@@ -22,6 +22,7 @@ import java.util.{Properties, Random}
 import scala.collection.mutable
 import scala.collection.mutable.ArrayBuffer
 
+import org.apache.hadoop.fs.FileAlreadyExistsException
 import org.mockito.ArgumentMatchers.{any, anyBoolean, anyInt, anyString}
 import org.mockito.Mockito._
 import org.mockito.invocation.InvocationOnMock
@@ -1775,4 +1776,23 @@ class TaskSetManagerSuite extends SparkFunSuite with LocalSparkContext with Logg
     assert(!manager.checkSpeculatableTasks(0))
     assert(manager.resourceOffer("exec1", "host1", ANY).isEmpty)
   }
+
+  test("TaskOutputFileAlreadyExistException lead to task set abortion") {
+    sc = new SparkContext("local", "test")
+    sched = new FakeTaskScheduler(sc, ("exec1", "host1"))
+    val taskSet = FakeTask.createTaskSet(1)
+    val manager = new TaskSetManager(sched, taskSet, MAX_TASK_FAILURES)
+    assert(sched.taskSetsFailed.isEmpty)
+
+    val offerResult = manager.resourceOffer("exec1", "host1", ANY)
+    assert(offerResult.isDefined,
+      "Expect resource offer on iteration 0 to return a task")
+    assert(offerResult.get.index === 0)
+    val reason = new ExceptionFailure(
+      new TaskOutputFileAlreadyExistException(
+        new FileAlreadyExistsException("file already exists")),
+      Seq.empty[AccumulableInfo])
+    manager.handleFailedTask(offerResult.get.taskId, TaskState.FAILED, reason)
+    assert(sched.taskSetsFailed.contains(taskSet.id))
+  }
 }

sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/FileFormatWriter.scala

@@ -20,7 +20,7 @@ package org.apache.spark.sql.execution.datasources
 import java.util.{Date, UUID}
 
 import org.apache.hadoop.conf.Configuration
-import org.apache.hadoop.fs.Path
+import org.apache.hadoop.fs.{FileAlreadyExistsException, Path}
 import org.apache.hadoop.mapreduce._
 import org.apache.hadoop.mapreduce.lib.output.FileOutputFormat
 import org.apache.hadoop.mapreduce.task.TaskAttemptContextImpl
@@ -281,6 +281,10 @@ object FileFormatWriter extends Logging {
     } catch {
       case e: FetchFailedException =>
         throw e
+      case f: FileAlreadyExistsException =>
+        // If any output file to write already exists, it does not make sense to re-run this task.
+        // We throw the exception and let Executor throw ExceptionFailure to abort the job.
+        throw new TaskOutputFileAlreadyExistException(f)
       case t: Throwable =>
         throw new SparkException("Task failed while writing rows.", t)
     }

sql/core/src/test/scala/org/apache/spark/sql/sources/InsertSuite.scala

@@ -20,6 +20,8 @@ package org.apache.spark.sql.sources
 import java.io.File
 import java.sql.Date
 
+import org.apache.hadoop.fs.{FileAlreadyExistsException, FSDataOutputStream, Path, RawLocalFileSystem}
+
 import org.apache.spark.SparkException
 import org.apache.spark.sql._
 import org.apache.spark.sql.catalyst.TableIdentifier
@@ -735,4 +737,35 @@ class InsertSuite extends DataSourceTest with SharedSparkSession {
       assert(msg.contains("Cannot write nullable values to non-null column 's'"))
     }
   }
+
+  test("Stop task set if FileAlreadyExistsException was thrown") {
+    withSQLConf("fs.file.impl" -> classOf[FileExistingTestFileSystem].getName,
+        "fs.file.impl.disable.cache" -> "true") {
+      withTable("t") {
+        sql(
+          """
+            |CREATE TABLE t(i INT, part1 INT) USING PARQUET
+            |PARTITIONED BY (part1)
+          """.stripMargin)
+
+        val df = Seq((1, 1)).toDF("i", "part1")
+        val err = intercept[SparkException] {
+          df.write.mode("overwrite").format("parquet").insertInto("t")
+        }
+        assert(err.getCause.getMessage.contains("can not write to output file: " +
+          "org.apache.hadoop.fs.FileAlreadyExistsException"))
+      }
+    }
+  }
+}
+
+class FileExistingTestFileSystem extends RawLocalFileSystem {
+  override def create(
+      f: Path,
+      overwrite: Boolean,
+      bufferSize: Int,
+      replication: Short,
+      blockSize: Long): FSDataOutputStream = {
+    throw new FileAlreadyExistsException(s"${f.toString} already exists")
+  }
 }