[SPARK-11328][SQL] Improve error message when hitting this issue

nongli · yhuai · commit f1122dd2bdc4 · 2015-12-01T15:30:30.000-08:00
The issue is that the output commiter is not idempotent and retry attempts will fail because the output file already exists. It is not safe to clean up the file as this output committer is by design not retryable. Currently, the job fails with a confusing file exists error. This patch is a stop gap to tell the user to look at the top of the error log for the proper message. This is difficult to test locally as Spark is hardcoded not to retry. Manually verified by upping the retry attempts. Author: Nong Li <nong@databricks.com> Author: Nong Li <nongli@gmail.com> Closes apache#10080 from nongli/spark-11328. (cherry picked from commit 47a0abc) Signed-off-by: Yin Huai <yhuai@databricks.com>
diff --git a/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/WriterContainer.scala b/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/WriterContainer.scala
@@ -124,6 +124,24 @@ private[sql] abstract class BaseWriterContainer(
     }
   }
 
+  protected def newOutputWriter(path: String): OutputWriter = {
+    try {
+      outputWriterFactory.newInstance(path, dataSchema, taskAttemptContext)
+    } catch {
+      case e: org.apache.hadoop.fs.FileAlreadyExistsException =>
+        if (outputCommitter.isInstanceOf[parquet.DirectParquetOutputCommitter]) {
+          // Spark-11382: DirectParquetOutputCommitter is not idempotent, meaning on retry
+          // attempts, the task will fail because the output file is created from a prior attempt.
+          // This often means the most visible error to the user is misleading. Augment the error
+          // to tell the user to look for the actual error.
+          throw new SparkException("The output file already exists but this could be due to a " +
+            "failure from an earlier attempt. Look through the earlier logs or stage page for " +
+            "the first error.\n  File exists error: " + e)
+        }
+        throw e
+    }
+  }
+
   private def newOutputCommitter(context: TaskAttemptContext): OutputCommitter = {
     val defaultOutputCommitter = outputFormatClass.newInstance().getOutputCommitter(context)
 
@@ -234,7 +252,7 @@ private[sql] class DefaultWriterContainer(
     executorSideSetup(taskContext)
     val configuration = SparkHadoopUtil.get.getConfigurationFromJobContext(taskAttemptContext)
     configuration.set("spark.sql.sources.output.path", outputPath)
-    val writer = outputWriterFactory.newInstance(getWorkPath, dataSchema, taskAttemptContext)
+    val writer = newOutputWriter(getWorkPath)
     writer.initConverter(dataSchema)
 
     var writerClosed = false
@@ -403,7 +421,7 @@ private[sql] class DynamicPartitionWriterContainer(
       val configuration = SparkHadoopUtil.get.getConfigurationFromJobContext(taskAttemptContext)
       configuration.set(
         "spark.sql.sources.output.path", new Path(outputPath, partitionPath).toString)
-      val newWriter = outputWriterFactory.newInstance(path.toString, dataSchema, taskAttemptContext)
+      val newWriter = super.newOutputWriter(path.toString)
       newWriter.initConverter(dataSchema)
       newWriter
     }
diff --git a/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/parquet/DirectParquetOutputCommitter.scala b/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/parquet/DirectParquetOutputCommitter.scala
@@ -41,7 +41,8 @@ import org.apache.parquet.hadoop.{ParquetFileReader, ParquetFileWriter, ParquetO
  *   no safe way undo a failed appending job (that's why both `abortTask()` and `abortJob()` are
  *   left empty).
  */
-private[parquet] class DirectParquetOutputCommitter(outputPath: Path, context: TaskAttemptContext)
+private[datasources] class DirectParquetOutputCommitter(
+    outputPath: Path, context: TaskAttemptContext)
   extends ParquetOutputCommitter(outputPath, context) {
   val LOG = Log.getLog(classOf[ParquetOutputCommitter])
 
