dongjoon-hyun commented on a change in pull request #28708: URL: https://github.com/apache/spark/pull/28708#discussion_r439853387
########## File path: core/src/main/scala/org/apache/spark/storage/BlockManagerDecommissioner.scala ########## @@ -0,0 +1,281 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.storage + +import java.util.concurrent.ExecutorService + +import scala.collection.JavaConverters._ +import scala.collection.mutable +import scala.util.control.NonFatal + +import org.apache.spark._ +import org.apache.spark.internal.Logging +import org.apache.spark.internal.config +import org.apache.spark.shuffle.{MigratableResolver, ShuffleBlockInfo} +import org.apache.spark.storage.BlockManagerMessages.ReplicateBlock +import org.apache.spark.util.ThreadUtils + +/** + * Class to handle block manager decommissioning retries. + * It creates a Thread to retry offloading all RDD cache and Shuffle blocks + */ +private[storage] class BlockManagerDecommissioner( + conf: SparkConf, + bm: BlockManager) extends Logging { + + private val maxReplicationFailuresForDecommission = + conf.get(config.STORAGE_DECOMMISSION_MAX_REPLICATION_FAILURE_PER_BLOCK) + + /** + * This runnable consumes any shuffle blocks in the queue for migration. This part of a + * producer/consumer where the main migration loop updates the queue of blocks to be migrated + * periodically. On migration failure, the current thread will reinsert the block for another + * thread to consume. Each thread migrates blocks to a different particular executor to avoid + * distribute the blocks as quickly as possible without overwhelming any particular executor. + * + * There is no preference for which peer a given block is migrated to. + * This is notable different than the RDD cache block migration (further down in this file) + * which uses the existing priority mechanism for determining where to replicate blocks to. + * Generally speaking cache blocks are less impactful as they normally represent narrow + * transformations and we normally have less cache present than shuffle data. + * + * The producer/consumer model is chosen for shuffle block migration to maximize + * the chance of migrating all shuffle blocks before the executor is forced to exit. + */ + private class ShuffleMigrationRunnable(peer: BlockManagerId) extends Runnable { + @volatile var running = true + override def run(): Unit = { + var migrating: Option[ShuffleBlockInfo] = None + logInfo(s"Starting migration thread for ${peer}") + // Once a block fails to transfer to an executor stop trying to transfer more blocks + try { + while (running && !Thread.interrupted()) { + val migrating = Option(shufflesToMigrate.poll()) + migrating match { + case None => + logInfo("Nothing to migrate") + // Nothing to do right now, but maybe a transfer will fail or a new block + // will finish being committed. + val SLEEP_TIME_SECS = 1 + Thread.sleep(SLEEP_TIME_SECS * 1000L) + case Some(shuffleBlockInfo) => + logInfo(s"Trying to migrate shuffle ${shuffleBlockInfo} to ${peer}") + val blocks = + bm.migratableResolver.getMigrationBlocks(shuffleBlockInfo) + logInfo(s"Got migration sub-blocks ${blocks}") + blocks.foreach { case (blockId, buffer) => + logInfo(s"Migrating sub-block ${blockId}") + bm.blockTransferService.uploadBlockSync( + peer.host, + peer.port, + peer.executorId, + blockId, + buffer, + StorageLevel.DISK_ONLY, + null)// class tag, we don't need for shuffle + logInfo(s"Migrated sub block ${blockId}") + } + logInfo(s"Migrated ${shuffleBlockInfo} to ${peer}") + } + } + // This catch is intentionally outside of the while running block. + // if we encounter errors migrating to an executor we want to stop. + } catch { + case e: Exception => + migrating match { + case Some(shuffleMap) => + logError(s"Error ${e} during migration, adding ${shuffleMap} back to migration queue") + shufflesToMigrate.add(shuffleMap) + case None => + logError(s"Error ${e} while waiting for block to migrate") Review comment: Could you describe a little more about which root cause do you have in mind? ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org
