ShuffleBlockFetcherIterator-spark技术分享

ShuffleBlockFetcherIterator

ShuffleBlockFetcherIterator is a Scala Iterator that fetches shuffle blocks (aka shuffle map outputs) from block managers.

ShuffleBlockFetcherIterator is created exclusively when BlockStoreShuffleReader is requested to read combined key-value records for a reduce task.

ShuffleBlockFetcherIterator allows for iterating over a sequence of blocks as (BlockId, InputStream) pairs so a caller can handle shuffle blocks in a pipelined fashion as they are received.

ShuffleBlockFetcherIterator is exhausted (i.e. can provide no elements) when the number of blocks already processed is at least the total number of blocks to fetch.

ShuffleBlockFetcherIterator throttles the remote fetches to avoid consuming too much memory.

Name Description

numBlocksProcessed

The number of blocks fetched and consumed.

numBlocksToFetch

Total number of blocks to fetch and consume.

ShuffleBlockFetcherIterator can produce up to numBlocksToFetch elements.

numBlocksToFetch is increased every time ShuffleBlockFetcherIterator is requested to splitLocalRemoteBlocks that prints it out as the INFO message to the logs:



Getting [numBlocksToFetch] non-empty blocks out of [totalBlocks] blocks

Getting [numBlocksToFetch] non-empty blocks out of [totalBlocks] blocks

results

Internal FIFO blocking queue (using Java’s java.util.concurrent.LinkedBlockingQueue) to hold FetchResult remote and local fetch results.

Used in:

1. next to take one FetchResult off the queue,

2. sendRequest to put SuccessFetchResult or FailureFetchResult remote fetch results (as part of BlockFetchingListener callback),

3. fetchLocalBlocks (similarly to sendRequest) to put local fetch results,

4. cleanup to release managed buffers for SuccessFetchResult results.

maxBytesInFlight

The maximum size (in bytes) of all the remote shuffle blocks to fetch.

Set when ShuffleBlockFetcherIterator is created.

maxReqsInFlight

The maximum number of remote requests to fetch shuffle blocks.

Set when ShuffleBlockFetcherIterator is created.

bytesInFlight

The bytes of fetched remote shuffle blocks in flight

Starts at 0 when ShuffleBlockFetcherIterator is created.

Incremented every sendRequest and decremented every next.

ShuffleBlockFetcherIterator makes sure that the invariant of bytesInFlight below maxBytesInFlight holds every remote shuffle block fetch.

reqsInFlight

The number of remote shuffle block fetch requests in flight.

Starts at 0 when ShuffleBlockFetcherIterator is created.

Incremented every sendRequest and decremented every next.

ShuffleBlockFetcherIterator makes sure that the invariant of reqsInFlight below maxReqsInFlight holds every remote shuffle block fetch.

isZombie

Flag whether ShuffleBlockFetcherIterator is still active. It is disabled, i.e. false, when ShuffleBlockFetcherIterator is created.

When enabled (when the task using ShuffleBlockFetcherIterator finishes), the block fetch successful callback (registered in sendRequest) will no longer add fetched remote shuffle blocks into results internal queue.

currentResult

The currently-processed SuccessFetchResult

Set when ShuffleBlockFetcherIterator returns the next (BlockId, InputStream) tuple and released (on cleanup).

Tip

Enable ERROR, WARN, INFO, DEBUG or TRACE logging levels for org.apache.spark.storage.ShuffleBlockFetcherIterator logger to see what happens in ShuffleBlockFetcherIterator.

Add the following line to conf/log4j.properties:



log4j.logger.org.apache.spark.storage.ShuffleBlockFetcherIterator=TRACE

log4j.logger.org.apache.spark.storage.ShuffleBlockFetcherIterator=TRACE

Refer to Logging.

`fetchUpToMaxBytes` Method

Caution

FIXME

Creating ShuffleBlockFetcherIterator Instance

When created, ShuffleBlockFetcherIterator takes the following:

TaskContext
ShuffleClient
BlockManager
Blocks to fetch per BlockManager (as Seq[(BlockManagerId, Seq[(BlockId, Long)])])
Function to wrap the returned input stream (as (BlockId, InputStream) ⇒ InputStream)
maxBytesInFlight — the maximum size (in bytes) of map outputs to fetch simultaneously from each reduce task (controlled by spark.reducer.maxSizeInFlight Spark property)
maxReqsInFlight — the maximum number of remote requests to fetch blocks at any given point (controlled by spark.reducer.maxReqsInFlight Spark property)
maxBlocksInFlightPerAddress
maxReqSizeShuffleToMem
detectCorrupt flag to detect any corruption in fetched blocks (controlled by spark.shuffle.detectCorrupt Spark property)

Initializing ShuffleBlockFetcherIterator — `initialize` Internal Method



initialize(): Unit

initialize(): Unit

initialize registers a task cleanup and fetches shuffle blocks from remote and local BlockManagers.

Internally, initialize registers a TaskCompletionListener (that will clean up right after the task finishes).

initialize splitLocalRemoteBlocks.

initialize registers the new remote fetch requests (with fetchRequests internal registry).

As ShuffleBlockFetcherIterator is in initialization phase, initialize makes sure that reqsInFlight and bytesInFlight internal counters are both 0. Otherwise, initialize throws an exception.

initialize fetches shuffle blocks (from remote BlockManagers).

You should see the following INFO message in the logs:



INFO ShuffleBlockFetcherIterator: Started [numFetches] remote fetches in [time] ms

INFO ShuffleBlockFetcherIterator: Started [numFetches] remote fetches in [time] ms

initialize fetches local shuffle blocks.

You should see the following DEBUG message in the logs:



DEBUG ShuffleBlockFetcherIterator: Got local blocks in  [time] ms

DEBUG ShuffleBlockFetcherIterator: Got local blocks in [time] ms

Note	`initialize` is used exclusively when `ShuffleBlockFetcherIterator` is created.

Sending Remote Shuffle Block Fetch Request — `sendRequest` Internal Method



sendRequest(req: FetchRequest): Unit

sendRequest(req: FetchRequest): Unit

Internally, when sendRequest runs, you should see the following DEBUG message in the logs:



DEBUG ShuffleBlockFetcherIterator: Sending request for [blocks.size] blocks ([size] B) from [hostPort]

DEBUG ShuffleBlockFetcherIterator: Sending request for [blocks.size] blocks ([size] B) from [hostPort]

sendRequest increments bytesInFlight and reqsInFlight internal counters.

Note	The input `FetchRequest` contains the remote BlockManagerId address and the shuffle blocks to fetch (as a sequence of BlockId and their sizes).

sendRequest requests ShuffleClient to fetch shuffle blocks (from the host, the port, and the executor as defined in the input FetchRequest).

Note	`ShuffleClient` was defined when `ShuffleBlockFetcherIterator` was created.

sendRequest registers a BlockFetchingListener with ShuffleClient that:

For every successfully fetched shuffle block adds it as SuccessFetchResult to results internal queue.
For every shuffle block fetch failure adds it as FailureFetchResult to results internal queue.

Note	`sendRequest` is used exclusively when `ShuffleBlockFetcherIterator` is requested to fetch remote shuffle blocks.

onBlockFetchSuccess Callback



onBlockFetchSuccess(blockId: String, buf: ManagedBuffer): Unit

onBlockFetchSuccess(blockId: String, buf: ManagedBuffer): Unit

Internally, onBlockFetchSuccess checks if the iterator is not zombie and does the further processing if it is not.

onBlockFetchSuccess marks the input blockId as received (i.e. removes it from all the blocks to fetch as requested in sendRequest).

onBlockFetchSuccess adds the managed buf (as SuccessFetchResult) to results internal queue.

You should see the following DEBUG message in the logs:



DEBUG ShuffleBlockFetcherIterator: remainingBlocks: [blocks]

DEBUG ShuffleBlockFetcherIterator: remainingBlocks: [blocks]

Regardless of zombie state of ShuffleBlockFetcherIterator, you should see the following TRACE message in the logs:



TRACE ShuffleBlockFetcherIterator: Got remote block [blockId] after [time] ms

TRACE ShuffleBlockFetcherIterator: Got remote block [blockId] after [time] ms

onBlockFetchFailure Callback



onBlockFetchFailure(blockId: String, e: Throwable): Unit

onBlockFetchFailure(blockId: String, e: Throwable): Unit

When onBlockFetchFailure is called, you should see the following ERROR message in the logs:



ERROR ShuffleBlockFetcherIterator: Failed to get block(s) from [hostPort]

ERROR ShuffleBlockFetcherIterator: Failed to get block(s) from [hostPort]

onBlockFetchFailure adds the block (as FailureFetchResult) to results internal queue.

Throwing FetchFailedException (for ShuffleBlockId) — `throwFetchFailedException` Internal Method



throwFetchFailedException(
  blockId: BlockId,
  address: BlockManagerId,
  e: Throwable): Nothing

throwFetchFailedException(

blockId: BlockId,

address: BlockManagerId,

e: Throwable): Nothing

throwFetchFailedException throws a FetchFailedException when the input blockId is a ShuffleBlockId.

Note	`throwFetchFailedException` creates a `FetchFailedException` passing on the root cause of a failure, i.e. the input `e`.

Otherwise, throwFetchFailedException throws a SparkException:



Failed to get block [blockId], which is not a shuffle block

Failed to get block [blockId], which is not a shuffle block

Note	`throwFetchFailedException` is used when `ShuffleBlockFetcherIterator` is requested for the next element.

Releasing Resources — `cleanup` Internal Method



cleanup(): Unit

cleanup(): Unit

Internally, cleanup marks ShuffleBlockFetcherIterator a zombie.

cleanup releases the current result buffer.

cleanup iterates over results internal queue and for every SuccessFetchResult, increments remote bytes read and blocks fetched shuffle task metrics, and eventually releases the managed buffer.

Note	`cleanup` is used when `ShuffleBlockFetcherIterator` initializes itself.

Decrementing Reference Count Of and Releasing Result Buffer (for SuccessFetchResult) — `releaseCurrentResultBuffer` Internal Method



releaseCurrentResultBuffer(): Unit

releaseCurrentResultBuffer(): Unit

releaseCurrentResultBuffer decrements the currently-processed SuccessFetchResult reference‘s buffer reference count if there is any.

releaseCurrentResultBuffer releases currentResult.

Note	`releaseCurrentResultBuffer` is used when `ShuffleBlockFetcherIterator` releases resources and `BufferReleasingInputStream` closes.

`fetchLocalBlocks` Internal Method



fetchLocalBlocks(): Unit

fetchLocalBlocks(): Unit

fetchLocalBlocks…FIXME

Note	`fetchLocalBlocks` is used when…FIXME

`hasNext` Method



hasNext: Boolean

hasNext: Boolean

Note	`hasNext` is part of Scala’s Iterator Contract to test whether this iterator can provide another element.

hasNext is positive (true) when numBlocksProcessed is less than numBlocksToFetch.

Otherwise, hasNext is negative (false).

`splitLocalRemoteBlocks` Internal Method



splitLocalRemoteBlocks(): ArrayBuffer[FetchRequest]

splitLocalRemoteBlocks(): ArrayBuffer[FetchRequest]

splitLocalRemoteBlocks…FIXME

Note	`splitLocalRemoteBlocks` is used exclusively when `ShuffleBlockFetcherIterator` is requested to initialize.

`next` Method



next(): (BlockId, InputStream)

next(): (BlockId, InputStream)

Note	`next` is part of Scala’s Iterator Contract to produce the next element of this iterator.

next…FIXME

ShuffleBlockFetcherIterator