DiskBlockManager creates and maintains the logical mapping between logical blocks and physical on-disk locations for a BlockManager.

DiskBlockManager BlockManager
Figure 1. DiskBlockManager and BlockManager

By default, one block is mapped to one file with a name given by its BlockId. It is however possible to have a block map to only a segment of a file.

Block files are hashed among the local directories.

DiskBlockManager is used to create a DiskStore.

Creating Instance

DiskBlockManager takes the following to be created:

When created, DiskBlockManager creates one or many local directories to store block data 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.

Local Directories for Blocks

localDirs: Array[File]

While being created, DiskBlockManager creates local directories for block data. 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:

File Locks for Local Block Store Directories

subDirs: Array[Array[File]]

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 (default: 64).

subDirs(dirId)(subDirId) is used to access subDirId subdirectory in dirId local directory.

subDirs is used when DiskBlockManager is requested for a block file or all block files.

Creating Local Directories for Block Data

  conf: SparkConf): Array[File]

createLocalDirs creates blockmgr-[random UUID] directory under local directories to store block data.

Internally, createLocalDirs finds the configured local directories where Spark can write files and creates a subdirectory blockmgr-[UUID] under every configured parent directory.

For every local directory, createLocalDirs 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 skips the directory.

Failed to create local dir in [rootDir]. Ignoring this directory.

createLocalDirs is used when the localDirs internal registry is initialized.

Finding Block File (and Creating Parent Directories)

  blockId: BlockId): File (1)
  filename: String): File
1 Uses the name of the given BlockId

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:

createTempShuffleBlock Method

createTempShuffleBlock(): (TempShuffleBlockId, File)

createTempShuffleBlock creates a temporary TempShuffleBlockId block.


All Block Files

getAllFiles(): Seq[File]


getAllFiles is used exclusively when DiskBlockManager is requested to getAllBlocks.

Registering Shutdown Hook — addShutdownHook Internal Method

addShutdownHook(): AnyRef

addShutdownHook registers a shutdown hook to execute doStop at shutdown.

When executed, you should see the following DEBUG message in the logs:

DEBUG DiskBlockManager: Adding shutdown hook

addShutdownHook adds the shutdown hook so it prints the following INFO message and executes doStop.

INFO DiskBlockManager: Shutdown hook called

Stopping DiskBlockManager (Removing Local Directories for Blocks) — doStop Internal Method

doStop(): Unit

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 DiskBlockManager is requested to shut down or stop.

Getting Local Directories for Spark to Write Files — Utils.getConfiguredLocalDirs Internal Method

getConfiguredLocalDirs(conf: SparkConf): Array[String]

getConfiguredLocalDirs returns the local directories where Spark can write files.

Internally, getConfiguredLocalDirs uses conf SparkConf to know if External Shuffle Service is enabled (based on spark.shuffle.service.enabled configuration property).

getConfiguredLocalDirs checks if Spark runs on YARN and if so, returns LOCAL_DIRS-controlled local directories.

In non-YARN mode (or for the driver in yarn-client mode), getConfiguredLocalDirs checks the following environment variables (in the order) and returns the value of the first met:

  1. SPARK_EXECUTOR_DIRS environment variable

  2. SPARK_LOCAL_DIRS environment variable

  3. MESOS_DIRECTORY environment variable (only when External Shuffle Service is not used)

In the end, when no earlier environment variables were found, getConfiguredLocalDirs uses spark.local.dir Spark property or falls back on java.io.tmpdir System property.

getConfiguredLocalDirs is used when:

Getting Writable Directories in YARN — getYarnLocalDirs Internal Method

getYarnLocalDirs(conf: SparkConf): String

getYarnLocalDirs uses 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 Exception with the message Yarn Local dirs can’t be empty if LOCAL_DIRS environment variable was not set.

Checking If Spark Runs on YARN — isRunningInYarnContainer Internal Method

isRunningInYarnContainer(conf: SparkConf): Boolean

isRunningInYarnContainer uses conf SparkConf to read Hadoop YARN’s CONTAINER_ID environment variable to find out if Spark runs in a YARN container.

CONTAINER_ID environment variable is exported by YARN NodeManager.

Getting All Blocks (From Files Stored On Disk)

getAllBlocks(): Seq[BlockId]

getAllBlocks gets all the blocks stored on disk.

Internally, getAllBlocks takes the block files and returns their names (as BlockId).

getAllBlocks is used when BlockManager is requested to find IDs of existing blocks for a given filter.

stop Internal Method

stop(): Unit


stop is used exclusively when BlockManager is requested to stop.


Enable ALL logging level for org.apache.spark.storage.DiskBlockManager logger to see what happens inside.

Add the following line to conf/log4j.properties:


Refer to Logging.