[SPARK-20732][CORE] Decommission cache blocks to other executors when an executor is decommissioned #27864
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -54,6 +54,7 @@ import org.apache.spark.rpc.RpcEnv | |
| import org.apache.spark.scheduler.ExecutorCacheTaskLocation | ||
| import org.apache.spark.serializer.{SerializerInstance, SerializerManager} | ||
| import org.apache.spark.shuffle.{ShuffleManager, ShuffleWriteMetricsReporter} | ||
| import org.apache.spark.storage.BlockManagerMessages.ReplicateBlock | ||
| import org.apache.spark.storage.memory._ | ||
| import org.apache.spark.unsafe.Platform | ||
| import org.apache.spark.util._ | ||
|
|
@@ -241,6 +242,9 @@ private[spark] class BlockManager( | |
|
|
||
| private var blockReplicationPolicy: BlockReplicationPolicy = _ | ||
|
|
||
| private var blockManagerDecommissioning: Boolean = false | ||
| private var decommissionManager: Option[BlockManagerDecommissionManager] = None | ||
|
|
||
| // A DownloadFileManager used to track all the files of remote blocks which are above the | ||
| // specified memory threshold. Files will be deleted automatically based on weak reference. | ||
| // Exposed for test | ||
|
|
@@ -1551,30 +1555,36 @@ private[spark] class BlockManager( | |
| } | ||
|
|
||
| /** | ||
| * Replicates a block to peer block managers based on existingReplicas and maxReplicas | ||
| * | ||
| * @param blockId blockId being replicate | ||
| * @param existingReplicas existing block managers that have a replica | ||
| * @param maxReplicas maximum replicas needed | ||
| * @param maxReplicationFailures number of replication failures to tolerate before | ||
| * giving up. | ||
| * @return whether block was successfully replicated or not | ||
| */ | ||
| def replicateBlock( | ||
| blockId: BlockId, | ||
| existingReplicas: Set[BlockManagerId], | ||
| maxReplicas: Int, | ||
| maxReplicationFailures: Option[Int] = None): Boolean = { | ||
| logInfo(s"Using $blockManagerId to pro-actively replicate $blockId") | ||
| blockInfoManager.lockForReading(blockId).forall { info => | ||
| val data = doGetLocalBytes(blockId, info) | ||
| val storageLevel = StorageLevel( | ||
| useDisk = info.level.useDisk, | ||
| useMemory = info.level.useMemory, | ||
| useOffHeap = info.level.useOffHeap, | ||
| deserialized = info.level.deserialized, | ||
| replication = maxReplicas) | ||
| // we know we are called as a result of an executor removal or because the current executor | ||
| // is getting decommissioned. so we refresh peer cache before trying replication, we won't | ||
| // try to replicate to a missing executor/another decommissioning executor | ||
| getPeers(forceFetch = true) | ||
| try { | ||
| replicate( | ||
| blockId, data, storageLevel, info.classTag, existingReplicas, maxReplicationFailures) | ||
| } finally { | ||
| logDebug(s"Releasing lock for $blockId") | ||
| releaseLockAndDispose(blockId, data) | ||
|
|
@@ -1591,9 +1601,11 @@ private[spark] class BlockManager( | |
| data: BlockData, | ||
| level: StorageLevel, | ||
| classTag: ClassTag[_], | ||
| existingReplicas: Set[BlockManagerId] = Set.empty, | ||
| maxReplicationFailures: Option[Int] = None): Boolean = { | ||
|
|
||
| val maxReplicationFailureCount = maxReplicationFailures.getOrElse( | ||
| conf.get(config.STORAGE_MAX_REPLICATION_FAILURE)) | ||
| val tLevel = StorageLevel( | ||
| useDisk = level.useDisk, | ||
| useMemory = level.useMemory, | ||
|
|
@@ -1617,7 +1629,7 @@ private[spark] class BlockManager( | |
| blockId, | ||
| numPeersToReplicateTo) | ||
|
|
||
| while(numFailures <= maxReplicationFailureCount && | ||
| !peersForReplication.isEmpty && | ||
| peersReplicatedTo.size < numPeersToReplicateTo) { | ||
| val peer = peersForReplication.head | ||
|
|
@@ -1665,9 +1677,11 @@ private[spark] class BlockManager( | |
| if (peersReplicatedTo.size < numPeersToReplicateTo) { | ||
| logWarning(s"Block $blockId replicated to only " + | ||
| s"${peersReplicatedTo.size} peer(s) instead of $numPeersToReplicateTo peers") | ||
| return false | ||
| } | ||
|
|
||
| logDebug(s"block $blockId replicated to ${peersReplicatedTo.mkString(", ")}") | ||
| return true | ||
| } | ||
|
|
||
| /** | ||
|
|
@@ -1761,6 +1775,58 @@ private[spark] class BlockManager( | |
| blocksToRemove.size | ||
| } | ||
|
|
||
| def decommissionBlockManager(): Unit = { | ||
| if (!blockManagerDecommissioning) { | ||
prakharjain09 marked this conversation as resolved.
Show resolved
Hide resolved
|
||
| logInfo("Starting block manager decommissioning process") | ||
| blockManagerDecommissioning = true | ||
| decommissionManager = Some(new BlockManagerDecommissionManager(conf)) | ||
| decommissionManager.foreach(_.start()) | ||
| } else { | ||
| logDebug("Block manager already in decommissioning state") | ||
| } | ||
| } | ||
|
|
||
| /** | ||
| * Tries to offload all cached RDD blocks from this BlockManager to peer BlockManagers | ||
| * Visible for testing | ||
| */ | ||
| def decommissionRddCacheBlocks(): Unit = { | ||
| val replicateBlocksInfo = master.getReplicateInfoForRDDBlocks(blockManagerId) | ||
|
|
||
| if (replicateBlocksInfo.nonEmpty) { | ||
| logInfo(s"Need to replicate ${replicateBlocksInfo.size} blocks " + | ||
| "for block manager decommissioning") | ||
| } | ||
|
|
||
| // Maximum number of storage replication failure which replicateBlock can handle | ||
| val maxReplicationFailures = conf.get( | ||
| config.STORAGE_DECOMMISSION_MAX_REPLICATION_FAILURE_PER_BLOCK) | ||
|
|
||
| // TODO: We can sort these blocks based on some policy (LRU/blockSize etc) | ||
| // so that we end up prioritize them over each other | ||
| val blocksFailedReplication = ThreadUtils.parmap( | ||
| replicateBlocksInfo, "decommissionRddCacheBlocks", 4) { | ||
| case ReplicateBlock(blockId, existingReplicas, maxReplicas) => | ||
| val replicatedSuccessfully = replicateBlock( | ||
|
There was a problem hiding this comment. I don't know if we need it but replicateBlock is blocking and it seems like maybe async + futures might help us migrate more blocks? Especially if one host is underload we might block on sending a block to that host before we move forward. There was a problem hiding this comment. @holdenk I am not sure how this will behave when multiple executors on same host machine are decommissioning. And each one of them is doing it in parallel - may cause some sort of network congestion? I have updated code to do replication in ThreadPool of size 4. Maybe we should make this configurable? any suggestions? There was a problem hiding this comment. Network congestion is certainly a possibility, I think that for now this strike a good balance between simple code and avoiding hanging all transfers if we have one slow target host. We can revisit this in the future if it turns out we need more control in production environments. Sound good? |
||
| blockId, | ||
| existingReplicas.toSet, | ||
| maxReplicas, | ||
| maxReplicationFailures = Some(maxReplicationFailures)) | ||
| if (replicatedSuccessfully) { | ||
| logInfo(s"Block $blockId offloaded successfully, Removing block now") | ||
| removeBlock(blockId) | ||
| logInfo(s"Block $blockId removed") | ||
| } else { | ||
| logWarning(s"Failed to offload block $blockId") | ||
| } | ||
| (blockId, replicatedSuccessfully) | ||
| }.filterNot(_._2).map(_._1) | ||
| if (blocksFailedReplication.nonEmpty) { | ||
| logWarning("Blocks failed replication in cache decommissioning " + | ||
| s"process: ${blocksFailedReplication.mkString(",")}") | ||
| } | ||
| } | ||
|
|
||
| /** | ||
| * Remove all blocks belonging to the given broadcast. | ||
| */ | ||
|
|
@@ -1829,7 +1895,52 @@ private[spark] class BlockManager( | |
| data.dispose() | ||
| } | ||
|
|
||
| /** | ||
| * Class to handle block manager decommissioning retries | ||
| * It creates a Thread to retry offloading all RDD cache blocks | ||
| */ | ||
| private class BlockManagerDecommissionManager(conf: SparkConf) { | ||
| @volatile private var stopped = false | ||
| private val blockReplicationThread = new Thread { | ||
| override def run(): Unit = { | ||
| while (blockManagerDecommissioning && !stopped) { | ||
| try { | ||
| logDebug("Attempting to replicate all cached RDD blocks") | ||
| decommissionRddCacheBlocks() | ||
| logInfo("Attempt to replicate all cached blocks done") | ||
| val sleepInterval = conf.get( | ||
| config.STORAGE_DECOMMISSION_REPLICATION_REATTEMPT_INTERVAL) | ||
| Thread.sleep(sleepInterval) | ||
| } catch { | ||
| case _: InterruptedException => | ||
| // no-op | ||
| case NonFatal(e) => | ||
| logError("Error occurred while trying to " + | ||
| "replicate cached RDD blocks for block manager decommissioning", e) | ||
| } | ||
| } | ||
| } | ||
| } | ||
| blockReplicationThread.setDaemon(true) | ||
| blockReplicationThread.setName("block-replication-thread") | ||
|
|
||
| def start(): Unit = { | ||
prakharjain09 marked this conversation as resolved.
Show resolved
Hide resolved
|
||
| logInfo("Starting block replication thread") | ||
| blockReplicationThread.start() | ||
| } | ||
|
|
||
| def stop(): Unit = { | ||
| if (!stopped) { | ||
| stopped = true | ||
| logInfo("Stopping block replication thread") | ||
| blockReplicationThread.interrupt() | ||
| blockReplicationThread.join() | ||
| } | ||
| } | ||
| } | ||
|
|
||
| def stop(): Unit = { | ||
| decommissionManager.foreach(_.stop()) | ||
| blockTransferService.close() | ||
| if (blockStoreClient ne blockTransferService) { | ||
| // Closing should be idempotent, but maybe not for the NioBlockTransferService. | ||
|
|
||
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Could you explain about the meaning of
decommissionhere?Let say, we have Spark Thrift Server. In that case, dynamic allocation wants to decommission all workers. Then, what this PR can provide for zero-worker situation?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@dongjoon-hyun Thanks for the detailed review.
I couldn't understand your question with respect to dynamic allocation completely (as per my understanding dynamic allocation doesn't initiates decommissioning). This change works together with SPARK-20628 PR. After SPARK-20628, Spark scheduler starts receiving decommissionExecutor message (because of AWS spot loss, Azure low-priority VMs etc). In such case, we should offload cache data from that executor to other active executors so that we don't eventually loose those cache blocks. This is done in a best effort way i.e. move cache blocks if space is available on active executors. If space is not available on other active executors, then keep the cache blocks.
Next step after this change is to initiate decommissioning from DynamicAllocation. Let say minExecutor=2, maxExecutor=50. A spark application is running at max 50 executors and all of them have little bit of cache data (say 2GB out of total available 5GB capacity). So DynamicAllocation is not able to downscale any of the executors (although they are idle). In future, DynamicAllocation can leverage changes in this PR to do defragmentation of cache data to fewer set of executors so that some of the executors can be freed up.
Hope this answers your question.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Since this is focused on executor to executor migration supporting zero executors isn't going to be a viable option here. However since were using the same methods for putting blocks, assuming that we add support for storing these blocks in external storage in the future, this approach should be able to be generalized to use that same mechanism if configured.