Steve Loughran created MAPREDUCE-7341:
-----------------------------------------
Summary: Add a task-manifest output committer for Azure and GCS
Key: MAPREDUCE-7341
URL: https://issues.apache.org/jira/browse/MAPREDUCE-7341
Project: Hadoop Map/Reduce
Issue Type: New Feature
Components: client
Affects Versions: 3.3.1
Reporter: Steve Loughran
Assignee: Steve Loughran
Add a task-manifest output committer for Azure and GCS
The S3A committers are very popular in Spark on S3, as they are both correct
and fast.
The classic FileOutputCommitter v1 and v2 algorithms are all that is available
for Azure ABFS and Google GCS,
and they have limitations.
The v2 algorithm isn't safe in the presence of failed task attempt commits, so
we
recommend the v1 algorithm for Azure. But that is slow because it sequentially
lists
then renames files and directories, one-by-one. The latencies of list
and rename make things slow.
Google GCS lacks the atomic directory rename required for v1 correctness;
v2 can be used (which doesn't have the job commit performance limitations),
but it's not safe.
Proposed
* Add a new FileOutputFormat committer which uses an intermediate manifest to
pass the list of files created by a TA to the job committer.
* Job committer to parallelise reading these task manifests and submit all the
rename operations into a pool of worker threads. (also: mkdir, directory
deletions on cleanup)
* Use the committer plugin mechanism added for s3a to make this the default
committer for ABFS
(i.e. no need to make any changes to FileOutputCommitter)
* Add lots of IOStatistics instrumentation + logging of operations in the
JobCommit
for visibility of where delays are occurring.
* Reuse the S3A committer _SUCCESS JSON structure to publish IOStats & other
data
for testing/support.
This committer will be faster than the V1 algorithm because of the
parallelisation, and
because a manifest written by create-and-rename will be exclusive to a single
task
attempt, delivers the isolation which the v2 committer lacks.
This is not an attempt to do an iceberg/hudi/delta-lake style manifest-only
format
for describing the contents of a table; the final output is still a directory
tree
which must be scanned during query planning.
As such the format is still suboptimal for cloud storage -but at least we will
have
faster job execution during the commit phases.
Note: this will also work on HDFS, where again, it should be faster than
the v1 committer. However the target is very much ABFS and GCS.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
