[MAPREDUCE-7029] FileOutputCommitter is slow on filesystems lacking recursive delete

• Type: Improvement
• Status: Resolved
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 2.8.2
• Fix Version/s: 3.1.0, 2.10.0

Environment

• Google Cloud Storage (with the GCS connector: https://github.com/GoogleCloudPlatform/bigdata-interop/tree/master/gcs) for HCFS compatibility.

• FileOutputCommitter algorithm v2.

• Running on Google Compute Engine with Java 8, Debian 8, Hadoop 2.8.2, Spark 2.2.0.

• Hadoop Flags: Reviewed

Release Note

MapReduce jobs that output to filesystems without direct support for recursive delete can set mapreduce.fileoutputcommitter.task.cleanup.enabled=true to have each task delete their intermediate work directory rather than waiting for the ApplicationMaster to clean up at the end of the job. This can significantly speed up the cleanup phase for large jobs on such filesystems.

• Assignee: Karthik Palaniappan
• Reporter: Karthik Palaniappan
• Created: 28/Dec/17 00:02
• Updated: 23/Jan/18 16:59
• Resolved: 17/Jan/18 16:32

Description

I ran a Spark job that outputs thousands of parquet files (aka there are thousands of reducers), and it hung for several minutes in the driver after all tasks were complete. Here is a very simple repro of the job (to be run in a spark-shell):

spark.range(1L << 20).repartition(1 << 14).write.save("gs://some/path")

Spark actually calls into Mapreduce's FileOuputCommitter. Job committing (specifically cleanupJob()) recursively deletes the job temporary directory, which is something like "gs://some/path/_temporary". If I understand correctly, on HDFS, this would be O(1), but on GCS (and every HCFS I know), this requires a full file tree walk. Deleting tens of thousands of objects in GCS takes several minutes.

I propose that commitTask() recursively deletes its the task attempt temp directory (something like "gs://some/path/_temporary/attempt1/task1"). On HDFS, this is O(1) per task, so this is very little overhead per task. On GCS (and other HCFSs), this adds parallelism for deleting the job temp directory.

With the attached patch, the repro above went from taking ~10 minutes to taking ~5 minutes, and task time did not significantly change.

Side note: I found this issue with Spark, but I assume it applies to a Mapreduce job with thousands of reducers as well.