IndexShuffleBlockResolver-spark技术分享

IndexShuffleBlockResolver

IndexShuffleBlockResolver is the one and only ShuffleBlockResolver in Spark.

IndexShuffleBlockResolver manages shuffle block data and uses shuffle index files for faster shuffle data access. IndexShuffleBlockResolver can write a shuffle block index and data file, find and remove shuffle index and data files per shuffle and
map.

Note	Shuffle block data files are more often referred as map outputs files.

IndexShuffleBlockResolver is managed exclusively by SortShuffleManager (so BlockManager can access shuffle block data).

spark IndexShuffleBlockResolver SortShuffleManager.png

Figure 1. SortShuffleManager creates IndexShuffleBlockResolver

IndexShuffleBlockResolver is later passed in when SortShuffleManager creates a ShuffleWriter for ShuffleHandle.

Table 1. IndexShuffleBlockResolver’s Internal Properties
Name	Initial Value	Description
`transportConf`	TransportConf for `shuffle` module	Used when `IndexShuffleBlockResolver` creates a `ManagedBuffer` for a `ShuffleBlockId`.

Creating IndexShuffleBlockResolver Instance

IndexShuffleBlockResolver takes the following when created:

SparkConf,
BlockManager (default: unspecified and SparkEnv is used to access one)

IndexShuffleBlockResolver initializes the internal properties.

Note	`IndexShuffleBlockResolver` is created exclusively when `SortShuffleManager` is created.

Writing Shuffle Index and Data Files — `writeIndexFileAndCommit` Method



writeIndexFileAndCommit(
  shuffleId: Int,
  mapId: Int,
  lengths: Array[Long],
  dataTmp: File): Unit

writeIndexFileAndCommit(

shuffleId: Int,

mapId: Int,

lengths: Array[Long],

dataTmp: File): Unit

Internally, writeIndexFileAndCommit first finds the index file for the input shuffleId and mapId.

writeIndexFileAndCommit creates a temporary file for the index file (in the same directory) and writes offsets (as the moving sum of the input lengths starting from 0 to the final offset at the end for the end of the output file).

Note	The offsets are the sizes in the input `lengths` exactly.

spark IndexShuffleBlockResolver writeIndexFileAndCommit.png

Figure 2. writeIndexFileAndCommit and offsets in a shuffle index file

writeIndexFileAndCommit requests a shuffle block data file for the input shuffleId and mapId.

writeIndexFileAndCommit checks if the given index and data files match each other (aka consistency check).

If the consistency check fails, it means that another attempt for the same task has already written the map outputs successfully and so the input dataTmp and temporary index files are deleted (as no longer correct).

If the consistency check succeeds, the existing index and data files are deleted (if they exist) and the temporary index and data files become “official”, i.e. renamed to their final names.

In case of any IO-related exception, writeIndexFileAndCommit throws a IOException with the messages:



fail to rename file [indexTmp] to [indexFile]

fail to rename file [indexTmp] to [indexFile]



fail to rename file [dataTmp] to [dataFile]

fail to rename file [dataTmp] to [dataFile]

Note	`writeIndexFileAndCommit` is used when ShuffleWriter is requested to write records to shuffle system, i.e. SortShuffleWriter, BypassMergeSortShuffleWriter, and UnsafeShuffleWriter.

Creating ManagedBuffer to Read Shuffle Block Data File — `getBlockData` Method



getBlockData(blockId: ShuffleBlockId): ManagedBuffer

getBlockData(blockId: ShuffleBlockId): ManagedBuffer

Note	`getBlockData` is part of ShuffleBlockResolver contract.

Internally, getBlockData finds the index file for the input shuffle blockId.

Note	ShuffleBlockId knows `shuffleId` and `mapId`.

getBlockData discards blockId.reduceId bytes of data from the index file.

Note	`getBlockData` uses Guava’s com.google.common.io.ByteStreams to skip the bytes.

getBlockData reads the start and end offsets from the index file and then creates a FileSegmentManagedBuffer to read the data file for the offsets (using transportConf internal property).

Note	The start and end offsets are the offset and the length of the file segment for the block data.

In the end, getBlockData closes the index file.

Checking Consistency of Shuffle Index and Data Files and Returning Block Lengths — `checkIndexAndDataFile` Internal Method



checkIndexAndDataFile(index: File, data: File, blocks: Int): Array[Long]

checkIndexAndDataFile(index: File, data: File, blocks: Int): Array[Long]

checkIndexAndDataFile first checks if the size of the input index file is exactly the input blocks multiplied by 8.

checkIndexAndDataFile returns null when the numbers, and hence the shuffle index and data files, don’t match.

checkIndexAndDataFile reads the shuffle index file and converts the offsets into lengths of each block.

checkIndexAndDataFile makes sure that the size of the input shuffle data file is exactly the sum of the block lengths.

checkIndexAndDataFile returns the block lengths if the numbers match, and null otherwise.

Note	`checkIndexAndDataFile` is used exclusively when `IndexShuffleBlockResolver` writes shuffle index and data files.

Requesting Shuffle Block Index File (from DiskBlockManager) — `getIndexFile` Internal Method



getIndexFile(shuffleId: Int, mapId: Int): File

getIndexFile(shuffleId: Int, mapId: Int): File

getIndexFile requests BlockManager for the current DiskBlockManager.

Note	`getIndexFile` uses `SparkEnv` to access the current `BlockManager` unless specified when `IndexShuffleBlockResolver` is created.

getIndexFile then requests DiskBlockManager for the shuffle index file given the input shuffleId and mapId (as ShuffleIndexBlockId)

Note	`getIndexFile` is used when `IndexShuffleBlockResolver` writes shuffle index and data files, creates a `ManagedBuffer` to read a shuffle block data file, and ultimately removes the shuffle index and data files.

Requesting Shuffle Block Data File — `getDataFile` Method



getDataFile(shuffleId: Int, mapId: Int): File

getDataFile(shuffleId: Int, mapId: Int): File

getDataFile requests BlockManager for the current DiskBlockManager.

Note	`getDataFile` uses `SparkEnv` to access the current `BlockManager` unless specified when `IndexShuffleBlockResolver` is created.

getDataFile then requests DiskBlockManager for the shuffle block data file given the input shuffleId, mapId, and the special reduce id 0 (as ShuffleDataBlockId).

Note

getDataFile is used when:

IndexShuffleBlockResolver writes an index file, creates a ManagedBuffer for ShuffleBlockId, and removes the data and index files that contain the output data from one map
ShuffleWriter is requested to write records to shuffle system, i.e. SortShuffleWriter, BypassMergeSortShuffleWriter, and UnsafeShuffleWriter.

Removing Shuffle Index and Data Files (For Single Map) — `removeDataByMap` Method



removeDataByMap(shuffleId: Int, mapId: Int): Unit

removeDataByMap(shuffleId: Int, mapId: Int): Unit

removeDataByMap finds and deletes the shuffle data for the input shuffleId and mapId first followed by finding and deleting the shuffle data index file.

When removeDataByMap fails deleting the files, you should see a WARN message in the logs.



WARN Error deleting data [path]

WARN Error deleting data [path]



WARN Error deleting index [path]

WARN Error deleting index [path]

Note	`removeDataByMap` is used exclusively when `SortShuffleManager` unregisters a shuffle, i.e. removes a shuffle from a shuffle system.

Stopping IndexShuffleBlockResolver — `stop` Method



stop(): Unit

stop(): Unit

Note	`stop` is part of ShuffleBlockResolver contract.

stop is a noop operation, i.e. does nothing when called.

IndexShuffleBlockResolver