By default, the mapreduce manifest committer is used for jobs working with abfs and gcs.
Hadoop mapreduce will pick this up automatically; for Spark it is a bit complicated: read the docs
to see the steps required.
This modifies the manifest committer so that the list of files
to rename is passed between stages as a file of
writeable entries on the local filesystem.
The map of directories to create is still passed in memory;
this map is built across all tasks, so even if many tasks
created files, if they all write into the same set of directories
the memory needed is O(directories) with the
task count not a factor.
The _SUCCESS file reports on heap size through gauges.
This should give a warning if there are problems.
Contributed by Steve Loughran
This:
1. Adds optLong, optDouble, mustLong and mustDouble
methods to the FSBuilder interface to let callers explicitly
passin long and double arguments.
2. The opt() and must() builder calls which take float/double values
now only set long values instead, so as to avoid problems
related to overloaded methods resulting in a ".0" being appended
to a long value.
3. All of the relevant opt/must calls in the hadoop codebase move to
the new methods
4. And the s3a code is resilient to parse errors in is numeric options
-it will downgrade to the default.
This is nominally incompatible, but the floating-point builder methods
were never used: nothing currently expects floating point numbers.
For anyone who wants to safely set numeric builder options across all compatible
releases, convert the number to a string and then use the opt(String, String)
and must(String, String) methods.
Contributed by Steve Loughran
Declares its compatibility with Spark's dynamic
output partitioning by having the stream capability
"mapreduce.job.committer.dynamic.partitioning"
Requires a Spark release with SPARK-40034, which
does the probing before deciding whether to
accept/rejecting instantiation with
dynamic partition overwrite set
This feature can be declared as supported by
any other PathOutputCommitter implementations
whose algorithm and destination filesystem
are compatible.
None of the S3A committers are compatible.
The classic FileOutputCommitter is, but it
does not declare itself as such out of our fear
of changing that code. The Spark-side code
will automatically infer compatibility if
the created committer is of that class or
a subclass.
Contributed by Steve Loughran.
* HADOOP-18321.Fix when to read an additional record from a BZip2 text file split
Co-authored-by: Ashutosh Gupta <ashugpt@amazon.com> and Reviewed by Akira Ajisaka.
Speed up the magic committer with key changes being
* Writes under __magic always retain directory markers
* File creation under __magic skips all overwrite checks,
including the LIST call intended to stop files being
created over dirs.
* mkdirs under __magic probes the path for existence
but does not look any further.
Extra parallelism in task and job commit directory scanning
Use of createFile and openFile with parameters which all for
HEAD checks to be skipped.
The committer can write the summary _SUCCESS file to the path
`fs.s3a.committer.summary.report.directory`, which can be in a
different file system/bucket if desired, using the job id as
the filename.
Also: HADOOP-15460. S3A FS to add `fs.s3a.create.performance`
Application code can set the createFile() option
fs.s3a.create.performance to true to disable the same
safety checks when writing under magic directories.
Use with care.
The createFile option prefix `fs.s3a.create.header.`
can be used to add custom headers to S3 objects when
created.
Contributed by Steve Loughran.
These changes ensure that sequential files are opened with the
right read policy, and split start/end is passed in.
As well as offering opportunities for filesystem clients to
choose fetch/cache/seek policies, the settings ensure that
processing text files on an s3 bucket where the default policy
is "random" will still be processed efficiently.
This commit depends on the associated hadoop-common patch,
which must be committed first.
Contributed by Steve Loughran.
Change-Id: Ic6713fd752441cf42ebe8739d05c2293a5db9f94
This is a mapreduce/spark output committer optimized for
performance and correctness on Azure ADLS Gen 2 storage
(via the abfs connector) and Google Cloud Storage
(via the external gcs connector library).
* It is safe to use with HDFS, however it has not been optimized
for that use.
* It is *not* safe for use with S3, and will fail if an attempt
is made to do so.
Contributed by Steve Loughran
Change-Id: I6f3502e79c578b9fd1a8c1485f826784b5421fca
