spark-SortShuffleWriter.adoc

SortShuffleWriter — Fallback ShuffleWriter

SortShuffleWriter is a ShuffleWriter that is used when SortShuffleManager returns a ShuffleWriter for ShuffleHandle (and the more specialized BypassMergeSortShuffleWriter and UnsafeShuffleWriter could not be used).

Note	SortShuffleWriter is parameterized by types for K keys, V values, and C combiner values.

Table 1. SortShuffleWriter’s Internal Registries and Counters

Name	Description
mapStatus	MapStatus the SortShuffleWriter has recently persisted (as a shuffle partitioned file in disk store). NOTE: Since write does not return a value, mapStatus attribute is used to be returned when SortShuffleWriter is closed.
stopping	Internal flag to mark that SortShuffleWriter is closed.

Tip	Enable ERROR logging level for org.apache.spark.shuffle.sort.SortShuffleWriter logger to see what happens in SortShuffleWriter. Add the following line to conf/log4j.properties: log4j.logger.org.apache.spark.shuffle.sort.SortShuffleWriter=ERROR Refer to Logging.

Creating SortShuffleWriter Instance

SortShuffleWriter takes the following when created:

1. IndexShuffleBlockResolver

2. BaseShuffleHandle

3. mapId — the mapper task id

4. TaskContext

Note	SortShuffleWriter is created when SortShuffleManager returns a ShuffleWriter for the fallback BaseShuffleHandle.

Writing Records Into Shuffle Partitioned File In Disk Store — write Method

write(records: Iterator[Product2[K, V]]): Unit

Note	write is a part of ShuffleWriter contract to write a sequence of records (for a RDD partition).

Internally, write creates a ExternalSorter with the types K, V, C or K, V, V depending on mapSideCombine flag of the ShuffleDependency being enabled or not, respectively.

Note	ShuffleDependency is defined when SortShuffleWriter is created (as the dependency of BaseShuffleHandle).

Note	write makes sure that Aggregator is defined for ShuffleDependency when mapSideCombine flag is enabled.

write inserts all the records to ExternalSorter

write requests IndexShuffleBlockResolver for the shuffle data output file (for the ShuffleDependency and mapId) and creates a temporary file for the shuffle data file in the same directory.

write creates a ShuffleBlockId (for the ShuffleDependency and mapId and the special IndexShuffleBlockResolver.NOOP_REDUCE_ID reduce id).

write requests ExternalSorter to write all the records (previously inserted in) into the temporary partitioned file in the disk store.

write requests IndexShuffleBlockResolver to write an index file (for the temporary partitioned file).

write creates a MapStatus (with the location of the shuffle server that serves the executor’s shuffle files and the sizes of the shuffle partitioned file’s partitions).

Note	The newly-created MapStatus is available as mapStatus internal attribute.

Note	write does not handle exceptions so when they occur, they will break the processing.

In the end, write deletes the temporary partitioned file. You may see the following ERROR message in the logs if write did not manage to do so:

ERROR Error while deleting temp file [path]

Closing SortShuffleWriter (and Calculating MapStatus) — stop Method

stop(success: Boolean): Option[MapStatus]

Note	stop is a part of ShuffleWriter contract to close itself (and return the last written MapStatus).

stop turns stopping flag on and returns the internal mapStatus if the input success is enabled.

Otherwise, when stopping flag is already enabled or the input success is disabled, stop returns no MapStatus (i.e. None).

In the end, stop stops the ExternalSorter and increments the shuffle write time task metrics.