DiskBlockManager manages a logical mapping of logical blocks and their physical on-disk locations for a BlockManager.
By default, one block is mapped to one file with a name given by BlockId. It is however possible to have a block to be mapped to a segment of a file only.
Block files are hashed among the local directories.
DiskBlockManager is used to create a DiskStore.
DiskBlockManager takes the following to be created:
DiskBlockManager creates the local directories for block storage and initializes the internal subDirs collection of locks for every local directory.
In the end,
DiskBlockManager registers a shutdown hook to clean up the local directories for blocks.
DiskBlockManager is created for BlockManager.
createLocalDirsForMergedShuffleBlocks is a noop with isPushBasedShuffleEnabled disabled (YARN mode only).
DiskBlockManager is available using SparkEnv.
Local Directories for Block Storage¶
DiskBlockManager creates blockmgr directory in every local root directory when created.
localDirs internal registry of all the
DiskBlockManager expects at least one local directory or prints out the following ERROR message to the logs and exits the JVM (with exit code 53):
Failed to create any local dir.
localDirs is used when:
DiskBlockManageris created (and creates localDirsString and subDirs), requested to look up a file (among local subdirectories) and doStop
BlockManageris requested to register with an external shuffle server
BasePythonRunner(PySpark) is requested to
localDirsString internal registry of the paths of the local blockmgr directories.
localDirsString is used by
BlockManager when requested for getLocalDiskDirs.
Creating blockmgr Directory in Every Local Root Directory¶
createLocalDirs( conf: SparkConf): Array[File]
blockmgr local directories for storing block data.
createLocalDirs creates a
blockmgr-[randomUUID] directory under every root directory for local storage and prints out the following INFO message to the logs:
Created local directory at [localDir]
In case of an exception,
createLocalDirs prints out the following ERROR message to the logs and ignore the directory:
Failed to create local dir in [rootDir]. Ignoring this directory.
File Locks for Local Block Store Directories¶
subDirs is a lookup table for file locks of every local block directory (with the first dimension for local directories and the second for locks).
The number of block subdirectories is controlled by spark.diskStore.subDirectories configuration property.
subDirs(dirId)(subDirId) is used to access
subDirId subdirectory in
dirId local directory.
subDirs is used when:
DiskBlockManageris requested for a block file and all the block files
Finding Block File (and Creating Parent Directories)¶
getFile( blockId: BlockId): File getFile( filename: String): File
getFile computes a hash of the file name of the input BlockId that is used for the name of the parent directory and subdirectory.
getFile creates the subdirectory unless it already exists.
getFile is used when:
DiskBlockManageris requested to containsBlock, createTempLocalBlock, createTempShuffleBlock
DiskStoreis requested to getBytes, remove, contains, and put
IndexShuffleBlockResolveris requested to getDataFile and getIndexFile
createTempShuffleBlock(): (TempShuffleBlockId, File)
createTempShuffleBlock creates a temporary
Registering Shutdown Hook¶
addShutdownHook registers a shutdown hook to execute doStop at shutdown.
When executed, you should see the following DEBUG message in the logs:
Adding shutdown hook
addShutdownHook adds the shutdown hook so it prints the following INFO message and executes doStop:
Shutdown hook called
Getting Writable Directories in YARN¶
getYarnLocalDirs( conf: SparkConf): String
conf SparkConf to read
LOCAL_DIRS environment variable with comma-separated local directories (that have already been created and secured so that only the user has access to them).
getYarnLocalDirs throws an
LOCAL_DIRS environment variable was not set:
Yarn Local dirs can't be empty
Checking Whether Spark Runs on YARN¶
isRunningInYarnContainer( conf: SparkConf): Boolean
conf SparkConf to read Hadoop YARN's CONTAINER_ID environment variable to find out if Spark runs in a YARN container (that is exported by YARN NodeManager).
Getting All Blocks (From Files Stored On Disk)¶
getAllBlocks gets all the blocks stored on disk.
getAllBlocks takes the block files and returns their names (as
getAllBlocks is used when
BlockManager is requested to find IDs of existing blocks for a given filter.
All Block Files¶
getAllFiles uses the subDirs registry to list all the files (in all the directories) that are currently stored on disk by this disk manager.
stop is used when
BlockManager is requested to stop.
Stopping DiskBlockManager (Removing Local Directories for Blocks)¶
doStop deletes the local directories recursively (only when the constructor's
deleteFilesOnStop is enabled and the parent directories are not registered to be removed at shutdown).
doStop is used when:
Demo: DiskBlockManager and Block Data
ALL logging level for
org.apache.spark.storage.DiskBlockManager logger to see what happens inside.
Add the following line to
Refer to Logging.