jerrypeng commented on code in PR #56008: URL: https://github.com/apache/spark/pull/56008#discussion_r3308002189
########## core/src/main/scala/org/apache/spark/StreamingShuffleOutputTracker.scala: ########## @@ -0,0 +1,400 @@ +/* + * 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 +import java.util.concurrent.{ConcurrentHashMap, LinkedBlockingQueue, ThreadPoolExecutor} + +import scala.jdk.CollectionConverters._ +import scala.reflect.ClassTag +import scala.util.control.NonFatal + +import org.apache.spark.internal.{Logging, LogKeys} +import org.apache.spark.internal.config.SHUFFLE_MAPOUTPUT_DISPATCHER_NUM_THREADS +import org.apache.spark.rpc.{RpcCallContext, RpcEndpoint, RpcEndpointRef, RpcEnv} +import org.apache.spark.util.ThreadUtils + +private[spark] sealed trait StreamingShuffleTaskLocationTrackerMessage + +private[spark] case class UpdateStreamingShuffleTaskLocation( + shuffleId: Int, + mapId: Long, + location: StreamingShuffleTaskLocation) + extends StreamingShuffleTaskLocationTrackerMessage + +private[spark] case class GetAllStreamingShuffleTaskLocations(shuffleId: Int) + extends StreamingShuffleTaskLocationTrackerMessage + +private[spark] case class GetAvailableStreamingShuffleTaskLocations(shuffleId: Int) + extends StreamingShuffleTaskLocationTrackerMessage + +private[spark] case object StopStreamingShuffleOutputTracker + extends StreamingShuffleTaskLocationTrackerMessage + +private[spark] sealed trait StreamingShuffleTaskLocationTrackerMasterMessage + +private[spark] case class UpdateStreamingShuffleTaskLocationMasterMessage( + shuffleId: Int, + mapId: Long, + location: StreamingShuffleTaskLocation, + context: RpcCallContext) + extends StreamingShuffleTaskLocationTrackerMasterMessage + +private[spark] case class GetAllStreamingShuffleTaskLocationsMasterMessage( + shuffleId: Int, + context: RpcCallContext) + extends StreamingShuffleTaskLocationTrackerMasterMessage + +private[spark] case class GetAvailableStreamingShuffleTaskLocationsMasterMessage( + shuffleId: Int, + context: RpcCallContext) + extends StreamingShuffleTaskLocationTrackerMasterMessage + +private[spark] case class StreamingShuffleTaskLocation( + executorId: String, + host: String, + port: Int) + +private[spark] class StreamingShuffleOutputTrackerMasterEndpoint( + override val rpcEnv: RpcEnv, + tracker: StreamingShuffleOutputTrackerMaster, + conf: SparkConf) + extends RpcEndpoint + with Logging { + + logDebug("init") // force eager creation of logger + + override def receiveAndReply(context: RpcCallContext): PartialFunction[Any, Unit] = { + case GetAllStreamingShuffleTaskLocations(shuffleId) => + tracker.post(GetAllStreamingShuffleTaskLocationsMasterMessage(shuffleId, context)) + + case GetAvailableStreamingShuffleTaskLocations(shuffleId) => + tracker.post(GetAvailableStreamingShuffleTaskLocationsMasterMessage(shuffleId, context)) + + case UpdateStreamingShuffleTaskLocation( + shuffleId: Int, + mapId: Long, + location: StreamingShuffleTaskLocation) => + tracker.post( + UpdateStreamingShuffleTaskLocationMasterMessage(shuffleId, mapId, location, context)) + + case StopStreamingShuffleOutputTracker => + logInfo(log"StreamingShuffleOutputTrackerMasterEndpoint stopped!") + // Unregister the endpoint from the Dispatcher BEFORE completing the reply. + // This avoids a race where the caller (e.g., initializeShuffleManager) is + // unblocked by context.reply() and tries to register a new endpoint with the + // same name before this handler has finished unregistering the old one. + // stop() synchronously calls Dispatcher.stop(self) which removes the endpoint + // name from the Dispatcher's endpoints map. + // context.reply() just completes a local Promise (p.success(true)) and has no + // dependency on the endpoint's registration state, so it is safe to call after + // stop(). + stop() + context.reply(true) + } +} + +case class ShuffleLocationResponse( + shuffleTaskLocations: Map[Long, StreamingShuffleTaskLocation], + numShuffleWriterTasks: Int) + +/** + * The purpose of this class is to support get task location information for the streaming + * shuffle. For the streaming shuffle to work, tasks need to figure out what where upstream tasks + * are located so it can connect to them to get data. The general usage of the APIs are the + * following + * 1. The DAGScheduler will call registerShuffle for the tracker running on the driver to + * register the shuffle. This is when information like number of mappers will be passed into + * the driver. + * 2. Mapper / Shuffle writer tasks will call registerShuffleWriterTask to + * register themselves with the tracker running on the driver and provide the information + * about where it is running + * 3. Reduce / Shuffle reader tasks will call + * getAllShuffleWriterTaskLocations to get the location information of all the writers. The + * API "getAllShuffleWriterTaskLocations" returns all the writer task locations or nothing at + * all. Partial results will not be returned. Shuffle readers should keep polling this API + * until it returns a non-empty response. + */ +private[spark] abstract class StreamingShuffleOutputTracker(conf: SparkConf) extends Logging { + + /** Reference to the master endpoint living on the driver. */ + var trackerEndpoint: RpcEndpointRef = _ + + /** + * Send a message to the trackerEndpoint and get its result within a default timeout, or + * throw a SparkException if this fails. + */ + protected def askTracker[T: ClassTag](message: Any): T = { + try { + trackerEndpoint.askSync[T](message) + } catch { + case e: Exception => + logError(log"Error communicating with StreamingShuffleOutputTracker", e) + throw new SparkException("Error communicating with StreamingShuffleOutputTracker", e) + } + } + + /** Send a one-way message to the trackerEndpoint, to which we expect it to reply with true. */ + protected def sendTracker(message: Any): Unit = { + val response = askTracker[Boolean](message) + if (response != true) { + throw new SparkException( + "Error reply received from StreamingShuffleOutputTracker. Expecting true, got " + + response.toString) + } + } + + def unregisterShuffle(shuffleId: Int): Unit = {} + + def stop(): Unit = {} + + /** + * Register a shuffle writer task (typically mapper task) for a shuffle and provide information + * about where it is running. Will return false if the register is not completed successfully + */ + def registerShuffleWriterTask( + shuffleId: Int, + mapId: Long, + location: StreamingShuffleTaskLocation): Boolean + + /** + * Get all shuffle writer tasks that are part of the shuffle. This API will return None until + * all shuffle writer task location info is available. + * @param shuffleId The id of the shuffle to get the lacation information about writer tasks + * @return A map in which the key is the shuffle writer task map id and the value + * is the location information of the task. + */ + def getAllShuffleWriterTaskLocations( + shuffleId: Int): Option[Map[Long, StreamingShuffleTaskLocation]] + + /** + * Get the shuffle writer task locations that are currently available for a shuffle. Please note + * that this function may not return all the shuffle writer locations if not all of them are + * available. This contrast with the method "getAllShuffleWriterTaskLocations" in which all + * locations will be returned or none with be returned. + * @param shuffleId The id of the shuffle to get the lacation information about writer tasks + * @return ShuffleLocationResponse which contains a map shuffle writer task locations and the + * total number of shuffle writers to expect. + */ + def getAvailableShuffleWriterTaskLocations(shuffleId: Int): Option[ShuffleLocationResponse] +} + +private[spark] case class StreamingShuffleInfo(numMaps: Int, numReduces: Int, jobId: Int) + +class StreamingShuffleOutputTrackerMaster(conf: SparkConf) + extends StreamingShuffleOutputTracker(conf) { + + // map that stores task location information organized in the following fashion + // shuffle id -> {mapId -> location} + private val taskLocations = + new ConcurrentHashMap[Int, ConcurrentHashMap[Long, StreamingShuffleTaskLocation]]() + + private val shuffleInfos = new ConcurrentHashMap[Int, StreamingShuffleInfo]() + + private val trackerMasterMessages = + new LinkedBlockingQueue[StreamingShuffleTaskLocationTrackerMasterMessage] + + // Thread pool used for handling map output status requests. This is a separate thread pool + // to ensure we don't block the normal dispatcher threads. + private val threadpool: ThreadPoolExecutor = { + val numThreads = conf.get(SHUFFLE_MAPOUTPUT_DISPATCHER_NUM_THREADS) + val pool = + ThreadUtils.newDaemonFixedThreadPool(numThreads, "streaming-shuffle-location-dispatcher") + for (i <- 0 until numThreads) { + pool.execute(new MessageLoop) + } + pool + } + + def registerShuffle(shuffleId: Int, numMaps: Int, numReduces: Int, jobId: Int): Unit = { + logInfo(log"Registering shuffleId ${MDC(LogKeys.SHUFFLE_ID, shuffleId)} with ${ + MDC(LogKeys.NUM_MAPPERS, numMaps)} mappers and ${ + MDC(LogKeys.NUM_REDUCERS, numReduces)} reducers") + shuffleInfos.put(shuffleId, StreamingShuffleInfo(numMaps, numReduces, jobId)) Review Comment: Also added a test -- 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. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
