BlockManagerMaster

BlockManagerMaster runs on the driver.

BlockManagerMaster uses BlockManagerMasterEndpoint — BlockManagerMaster RPC Endpoint registered under BlockManagerMaster RPC endpoint name on the driver (with the endpoint references on executors) to allow executors for sending block status updates to it and hence keep track of block statuses.

Creating Instance

BlockManagerMaster takes the following to be created:

BlockManagerMaster is created when SparkEnv utility is used to create a SparkEnv (for the driver and executors) (to create a BlockManager).

removeExecutorAsync Method

FIXME

contains Method

FIXME

Removing Executor — removeExecutor Method

removeExecutor(execId: String): Unit

removeExecutor posts RemoveExecutor to BlockManagerMaster RPC endpoint and waits for a response.

If false in response comes in, a SparkException is thrown with the following message:

BlockManagerMasterEndpoint returned false, expected true.

If all goes fine, you should see the following INFO message in the logs:

INFO BlockManagerMaster: Removed executor [execId]
removeExecutor is executed when DAGScheduler processes ExecutorLost event.

Removing Block — removeBlock Method

removeBlock(blockId: BlockId): Unit

removeBlock simply posts a RemoveBlock blocking message to BlockManagerMasterEndpoint — BlockManagerMaster RPC Endpoint (and ultimately disregards the reponse).

Removing RDD Blocks — removeRdd Method

removeRdd(rddId: Int, blocking: Boolean)

removeRdd removes all the blocks of rddId RDD, possibly in blocking fashion.

Internally, removeRdd posts a RemoveRdd(rddId) message to BlockManagerMasterEndpoint — BlockManagerMaster RPC Endpoint on a separate thread.

If there is an issue, you should see the following WARN message in the logs and the entire exception:

WARN Failed to remove RDD [rddId] - [exception]

If it is a blocking operation, it waits for a result for spark.rpc.askTimeout, spark.network.timeout or 120 secs.

Removing Shuffle Blocks — removeShuffle Method

removeShuffle(shuffleId: Int, blocking: Boolean)

removeShuffle removes all the blocks of shuffleId shuffle, possibly in a blocking fashion.

It posts a RemoveShuffle(shuffleId) message to BlockManagerMasterEndpoint — BlockManagerMaster RPC Endpoint on a separate thread.

If there is an issue, you should see the following WARN message in the logs and the entire exception:

WARN Failed to remove shuffle [shuffleId] - [exception]

If it is a blocking operation, it waits for the result for spark.rpc.askTimeout, spark.network.timeout or 120 secs.

removeShuffle is used exclusively when ContextCleaner removes a shuffle.

Removing Broadcast Blocks — removeBroadcast Method

removeBroadcast(broadcastId: Long, removeFromMaster: Boolean, blocking: Boolean)

removeBroadcast removes all the blocks of broadcastId broadcast, possibly in a blocking fashion.

It posts a RemoveBroadcast(broadcastId, removeFromMaster) message to BlockManagerMasterEndpoint — BlockManagerMaster RPC Endpoint on a separate thread.

If there is an issue, you should see the following WARN message in the logs and the entire exception:

WARN Failed to remove broadcast [broadcastId] with removeFromMaster = [removeFromMaster] - [exception]

If it is a blocking operation, it waits for the result for spark.rpc.askTimeout, spark.network.timeout or 120 secs.

Stopping BlockManagerMaster — stop Method

stop(): Unit

stop sends a StopBlockManagerMaster message to BlockManagerMasterEndpoint — BlockManagerMaster RPC Endpoint and waits for a response.

It is only executed for the driver.

If all goes fine, you should see the following INFO message in the logs:

INFO BlockManagerMaster: BlockManagerMaster stopped

Otherwise, a SparkException is thrown.

BlockManagerMasterEndpoint returned false, expected true.

Registering BlockManager with Driver

registerBlockManager(
  blockManagerId: BlockManagerId,
  maxMemSize: Long,
  slaveEndpoint: RpcEndpointRef): BlockManagerId

registerBlockManager prints the following INFO message to the logs:

Registering BlockManager [blockManagerId]
BlockManagerMaster RegisterBlockManager
Figure 1. Registering BlockManager with the Driver

registerBlockManager then notifies the driver that the BlockManagerId is registering itself. registerBlockManager posts a blocking RegisterBlockManager message to BlockManagerMaster RPC endpoint.

registerBlockManager waits until a confirmation comes (as BlockManagerId).

In the end, registerBlockManager prints the following INFO message to the logs and returns the BlockManagerId received.

Registered BlockManager [updatedId]

registerBlockManager is used when BlockManager is requested to initialize and re-register itself with the driver.

Relaying Block Status Update From BlockManager to Driver

updateBlockInfo(
  blockManagerId: BlockManagerId,
  blockId: BlockId,
  storageLevel: StorageLevel,
  memSize: Long,
  diskSize: Long): Boolean

updateBlockInfo sends a blocking UpdateBlockInfo event to BlockManagerMaster RPC endpoint (and waits for a response).

updateBlockInfo prints out the following DEBUG message to the logs:

DEBUG BlockManagerMaster: Updated info of block [blockId]

updateBlockInfo returns the response from the BlockManagerMaster RPC endpoint.

updateBlockInfo is used exclusively when BlockManager is requested to report a block status update to the driver.

Get Block Locations of One Block — getLocations Method

getLocations(blockId: BlockId): Seq[BlockManagerId]

Get Block Locations for Multiple Blocks — getLocations Method

getLocations(blockIds: Array[BlockId]): IndexedSeq[Seq[BlockManagerId]]

Finding Peers of BlockManager — getPeers Internal Method

getPeers(blockManagerId: BlockManagerId): Seq[BlockManagerId]
Peers of a BlockManager are the other BlockManagers in a cluster (except the driver’s BlockManager). Peers are used to know the available executors in a Spark application.
getPeers is used when BlockManager finds the peers of a BlockManager, Structured Streaming’s KafkaSource and Spark Streaming’s KafkaRDD.

getExecutorEndpointRef Method

getExecutorEndpointRef(executorId: String): Option[RpcEndpointRef]

getExecutorEndpointRef posts GetExecutorEndpointRef(executorId) message to BlockManagerMasterEndpoint — BlockManagerMaster RPC Endpoint and waits for a response which becomes the return value.

getMemoryStatus Method

getMemoryStatus: Map[BlockManagerId, (Long, Long)]

getMemoryStatus posts a GetMemoryStatus message BlockManagerMasterEndpoint — BlockManagerMaster RPC Endpoint and waits for a response which becomes the return value.

Storage Status (Posting GetStorageStatus to BlockManagerMaster RPC endpoint) — getStorageStatus Method

getStorageStatus: Array[StorageStatus]

getStorageStatus posts a GetStorageStatus message to BlockManagerMasterEndpoint — BlockManagerMaster RPC Endpoint and waits for a response which becomes the return value.

getBlockStatus Method

getBlockStatus(
  blockId: BlockId,
  askSlaves: Boolean = true): Map[BlockManagerId, BlockStatus]

getBlockStatus posts a GetBlockStatus(blockId, askSlaves) message to BlockManagerMasterEndpoint — BlockManagerMaster RPC Endpoint and waits for a response (of type Map[BlockManagerId, Future[Option[BlockStatus]]]).

It then builds a sequence of future results that are BlockStatus statuses and waits for a result for spark.rpc.askTimeout, spark.network.timeout or 120 secs.

No result leads to a SparkException with the following message:

BlockManager returned null for BlockStatus query: [blockId]

getMatchingBlockIds Method

getMatchingBlockIds(
  filter: BlockId => Boolean,
  askSlaves: Boolean): Seq[BlockId]

getMatchingBlockIds posts a GetMatchingBlockIds(filter, askSlaves) message to BlockManagerMasterEndpoint — BlockManagerMaster RPC Endpoint and waits for a response which becomes the result for spark.rpc.askTimeout, spark.network.timeout or 120 secs.

hasCachedBlocks Method

hasCachedBlocks(executorId: String): Boolean

hasCachedBlocks posts a HasCachedBlocks(executorId) message to BlockManagerMasterEndpoint — BlockManagerMaster RPC Endpoint and waits for a response which becomes the result.

Logging

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

Add the following line to conf/log4j.properties:

log4j.logger.org.apache.spark.storage.BlockManagerMaster=ALL

Refer to Logging.