[GitHub] vvraskin commented on a change in pull request #3240: Add a loadbalancer with local state and horizontal invoker sharding.

[GitBox <gi...@apache.org>]

vvraskin commented on a change in pull request #3240: Add a loadbalancer with local state and horizontal invoker sharding.
URL: https://github.com/apache/incubator-openwhisk/pull/3240#discussion_r166048402
 
 

 ##########
 File path: core/controller/src/main/scala/whisk/core/loadBalancer/ShardingContainerPoolBalancer.scala
 ##########
 @@ -0,0 +1,439 @@
+/*
+ * 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 whisk.core.loadBalancer
+
+import java.nio.charset.StandardCharsets
+import java.util.concurrent.atomic.LongAdder
+import java.util.concurrent.ThreadLocalRandom
+
+import akka.actor.{Actor, ActorSystem, Props}
+import akka.cluster.ClusterEvent._
+import akka.cluster.{Cluster, Member, MemberStatus}
+import akka.event.Logging.InfoLevel
+import akka.stream.ActorMaterializer
+import org.apache.kafka.clients.producer.RecordMetadata
+import pureconfig._
+import whisk.common.{ForcableSemaphore, Logging, LoggingMarkers, TransactionId}
+import whisk.core.WhiskConfig._
+import whisk.core.connector._
+import whisk.core.entity._
+import whisk.core.{ConfigKeys, WhiskConfig}
+import whisk.spi.SpiLoader
+
+import scala.annotation.tailrec
+import scala.collection.concurrent.TrieMap
+import scala.concurrent.duration._
+import scala.concurrent.{ExecutionContext, Future, Promise}
+import scala.util.{Failure, Success}
+
+/**
+ * A loadbalancer that uses "horizontal" sharding to not collide with fellow loadbalancers.
+ *
+ * Horizontal sharding means, that each invoker's capacity is evenly divided between the loadbalancers. If an invoker
+ * has at most 16 slots available, those will be divided to 8 slots for each loadbalancer (if there are 2).
+ */
+class ShardingContainerPoolBalancer(config: WhiskConfig, controllerInstance: InstanceId)(
+  implicit val actorSystem: ActorSystem,
+  logging: Logging,
+  materializer: ActorMaterializer)
+    extends LoadBalancer {
+
+  private implicit val executionContext: ExecutionContext = actorSystem.dispatcher
+
+  /** Build a cluster of all loadbalancers */
+  val seedNodesProvider = new StaticSeedNodesProvider(config.controllerSeedNodes, actorSystem.name)
+  val cluster = Cluster(actorSystem)
+  cluster.joinSeedNodes(seedNodesProvider.getSeedNodes())
+
+  /** Used to manage an action for testing invoker health */
+  private val entityStore = WhiskEntityStore.datastore(config)
+
+  /** State related to invocations and throttling */
+  private val activations = TrieMap[ActivationId, ActivationEntry]()
+  private val activationsPerNamespace = TrieMap[UUID, LongAdder]()
+  private val totalActivations = new LongAdder()
+
+  /** State needed for scheduling. */
+  private val schedulingState = ShardingContainerPoolBalancerState()()
+
+  /**
+   * Monitors invoker supervision and the cluster to update the state sequentially
+   *
+   * All state updates should go through this actor to guarantee, that `updateState` and `updateCluster` are called
+   * mutually exclusive and not concurrently.
+   */
+  private val monitor = actorSystem.actorOf(Props(new Actor {
+    override def preStart(): Unit = {
+      cluster.subscribe(self, classOf[MemberEvent], classOf[ReachabilityEvent])
+    }
+
+    // all members of the cluster that are available
+    var availableMembers = Set.empty[Member]
+
+    override def receive: Receive = {
+      case CurrentInvokerPoolState(newState) =>
+        schedulingState.updateInvokers(newState)
+
+      // State of the cluster as it is right now
+      case CurrentClusterState(members, _, _, _, _) =>
+        availableMembers = members.filter(_.status == MemberStatus.Up)
+        schedulingState.updateCluster(availableMembers.size)
+
+      // General lifecycle events and events concerning the reachability of members. Split-brain is not a huge concern
+      // in this case as only the invoker-threshold is adjusted according to the perceived cluster-size.
+      // Taking the unreachable member out of the cluster from that point-of-view results in a better experience
+      // even under split-brain-conditions, as that (in the worst-case) results in premature overloading of invokers vs.
+      // going into overflow mode prematurely.
+      case event: ClusterDomainEvent =>
+        availableMembers = event match {
+          case MemberUp(member)          => availableMembers + member
+          case ReachableMember(member)   => availableMembers + member
+          case MemberRemoved(member, _)  => availableMembers - member
+          case UnreachableMember(member) => availableMembers - member
+          case _                         => availableMembers
+        }
+
+        schedulingState.updateCluster(availableMembers.size)
+    }
+  }))
+
+  /** Loadbalancer interface methods */
+  override def invokerHealth(): Future[IndexedSeq[InvokerHealth]] = Future.successful(schedulingState.invokers)
+  override def activeActivationsFor(namespace: UUID): Future[Int] =
+    Future.successful(activationsPerNamespace.get(namespace).map(_.intValue()).getOrElse(0))
+  override def totalActiveActivations: Future[Int] = Future.successful(totalActivations.intValue())
+
+  /** 1. Publish a message to the loadbalancer */
+  override def publish(action: ExecutableWhiskActionMetaData, msg: ActivationMessage)(
+    implicit transid: TransactionId): Future[Future[Either[ActivationId, WhiskActivation]]] = {
+
+    val invokersToUse = if (action.exec.pull) schedulingState.blackboxInvokers else schedulingState.managedInvokers
+    val chosen = if (invokersToUse.nonEmpty) {
+      val hash = ShardingContainerPoolBalancer.generateHash(msg.user.namespace, action.fullyQualifiedName(false))
+      val homeInvoker = hash % invokersToUse.size
+      val stepSize = schedulingState.stepSizes(hash % schedulingState.stepSizes.size)
+      ShardingContainerPoolBalancer.schedule(invokersToUse, schedulingState.invokerSlots, homeInvoker, stepSize)
+    } else {
+      None
+    }
+
+    chosen
+      .map { invoker =>
+        val entry = setupActivation(msg, action, invoker)
+        sendActivationToInvoker(messageProducer, msg, invoker).map { _ =>
+          entry.promise.future
+        }
+      }
+      .getOrElse(Future.failed(LoadBalancerException("No invokers available")))
+  }
+
+  /** 2. Update local state with the to be executed activation */
+  private def setupActivation(msg: ActivationMessage,
+                              action: ExecutableWhiskActionMetaData,
+                              instance: InstanceId): ActivationEntry = {
+
+    totalActivations.increment()
+    activationsPerNamespace.getOrElseUpdate(msg.user.uuid, new LongAdder()).increment()
+
+    val timeout = action.limits.timeout.duration.max(TimeLimit.STD_DURATION) + 1.minute
+    // Install a timeout handler for the catastrophic case where an active ack is not received at all
+    // (because say an invoker is down completely, or the connection to the message bus is disrupted) or when
+    // the active ack is significantly delayed (possibly dues to long queues but the subject should not be penalized);
+    // in this case, if the activation handler is still registered, remove it and update the books.
+    activations.getOrElseUpdate(
+      msg.activationId, {
+        val timeoutHandler = actorSystem.scheduler.scheduleOnce(timeout) {
+          processCompletion(Left(msg.activationId), msg.transid, forced = true, invoker = instance)
+        }
+
+        // please note: timeoutHandler.cancel must be called on all non-timeout paths, e.g. Success
+        ActivationEntry(
+          msg.activationId,
+          msg.user.uuid,
+          instance,
+          timeoutHandler,
+          Promise[Either[ActivationId, WhiskActivation]]())
+      })
+  }
+
+  private val messagingProvider = SpiLoader.get[MessagingProvider]
+  private val messageProducer = messagingProvider.getProducer(config, executionContext)
+
+  /** 3. Send the activation to the invoker */
+  private def sendActivationToInvoker(producer: MessageProducer,
+                                      msg: ActivationMessage,
+                                      invoker: InstanceId): Future[RecordMetadata] = {
+    implicit val transid: TransactionId = msg.transid
+
+    val topic = s"invoker${invoker.toInt}"
+    val start = transid.started(
+      this,
+      LoggingMarkers.CONTROLLER_KAFKA,
+      s"posting topic '$topic' with activation id '${msg.activationId}'",
+      logLevel = InfoLevel)
+
+    producer.send(topic, msg).andThen {
+      case Success(status) =>
+        transid.finished(
+          this,
+          start,
+          s"posted to ${status.topic()}[${status.partition()}][${status.offset()}]",
+          logLevel = InfoLevel)
+      case Failure(e) => transid.failed(this, start, s"error on posting to topic $topic")
+    }
+  }
+
+  /**
+   * Subscribes to active acks (completion messages from the invokers), and
+   * registers a handler for received active acks from invokers.
+   */
+  private val maxActiveAcksPerPoll = 128
+  private val activeAckPollDuration = 1.second
+  private val activeAckConsumer =
+    messagingProvider.getConsumer(
+      config,
+      "completions",
+      s"completed${controllerInstance.toInt}",
+      maxPeek = maxActiveAcksPerPoll)
+
+  private val activationFeed = actorSystem.actorOf(Props {
+    new MessageFeed(
+      "activeack",
+      logging,
+      activeAckConsumer,
+      maxActiveAcksPerPoll,
+      activeAckPollDuration,
+      processActiveAck)
+  })
+
+  /** 4. Get the active-ack message and parse it */
+  private def processActiveAck(bytes: Array[Byte]): Future[Unit] = Future {
+    val raw = new String(bytes, StandardCharsets.UTF_8)
+    CompletionMessage.parse(raw) match {
+      case Success(m: CompletionMessage) =>
+        processCompletion(m.response, m.transid, forced = false, invoker = m.invoker)
+        activationFeed ! MessageFeed.Processed
+
+      case Failure(t) =>
+        activationFeed ! MessageFeed.Processed
+        logging.error(this, s"failed processing message: $raw with $t")
+    }
+  }
+
+  /** 5. Process the active-ack and update the state accordingly */
+  private def processCompletion(response: Either[ActivationId, WhiskActivation],
+                                tid: TransactionId,
+                                forced: Boolean,
+                                invoker: InstanceId): Unit = {
+    val aid = response.fold(l => l, r => r.activationId)
+
+    // treat left as success (as it is the result of a message exceeding the bus limit)
+    val isSuccess = response.fold(l => true, r => !r.response.isWhiskError)
+
+    activations.remove(aid) match {
+      case Some(entry) =>
+        totalActivations.decrement()
+        activationsPerNamespace.get(entry.namespaceId).foreach(_.decrement())
+        schedulingState.invokerSlots.lift(invoker.toInt).foreach(_.release())
+
+        logging.info(this, s"${if (!forced) "received" else "forced"} active ack for '$aid'")(tid)
+        // Active acks that are received here are strictly from user actions - health actions are not part of
+        // the load balancer's activation map. Inform the invoker pool supervisor of the user action completion.
+        invokerPool ! InvocationFinishedMessage(invoker, isSuccess)
+        if (!forced) {
+          entry.timeoutHandler.cancel()
+          entry.promise.trySuccess(response)
+        } else {
+          entry.promise.tryFailure(new Throwable("no active ack received"))
+        }
+      case None if !forced =>
+        // the entry has already been removed but we receive an active ack for this activation Id.
+        // This happens for health actions, because they don't have an entry in Loadbalancerdata or
+        // for activations that already timed out.
+        invokerPool ! InvocationFinishedMessage(invoker, isSuccess)
+        logging.debug(this, s"received active ack for '$aid' which has no entry")(tid)
+      case None =>
+        // the entry has already been removed by an active ack. This part of the code is reached by the timeout.
+        // As the active ack is already processed we don't have to do anything here.
+        logging.debug(this, s"forced active ack for '$aid' which has no entry")(tid)
+    }
+  }
+
+  private val invokerPool = {
+    InvokerPool.prepare(controllerInstance, WhiskEntityStore.datastore(config))
+
+    actorSystem.actorOf(
+      InvokerPool.props(
+        (f, i) => f.actorOf(InvokerActor.props(i, controllerInstance)),
+        (m, i) => sendActivationToInvoker(messageProducer, m, i),
+        messagingProvider.getConsumer(config, s"health${controllerInstance.toInt}", "health", maxPeek = 128),
+        Some(monitor)))
+  }
+}
+
+object ShardingContainerPoolBalancer extends LoadBalancerProvider {
+
+  override def loadBalancer(whiskConfig: WhiskConfig, instance: InstanceId)(
+    implicit actorSystem: ActorSystem,
+    logging: Logging,
+    materializer: ActorMaterializer): LoadBalancer = new ShardingContainerPoolBalancer(whiskConfig, instance)
+
+  def requiredProperties: Map[String, String] =
+    kafkaHosts ++
+      Map(controllerLocalBookkeeping -> null, controllerSeedNodes -> null)
+
+  def generateHash(namespace: EntityName, action: FullyQualifiedEntityName): Int = {
+    (namespace.asString.hashCode() ^ action.asString.hashCode()).abs
+  }
+
+  /**
+   * Scans through all invokers and searches for an invoker tries to get a free slot on an invoker. If no slot can be
+   * obtained, randomly picks a healthy invoker.
+   *
+   * @param invokers a list of available invokers to search in, including their state
+   * @param dispatched semaphores for each invoker to give the slots away from
+   * @param index the index to start from (initially should be the "homeInvoker"
+   * @param step stable identifier of the entity to be scheduled
+   * @return an invoker to schedule to or None of no invoker is available
+   */
+  @tailrec
+  def schedule(invokers: IndexedSeq[InvokerHealth],
+               dispatched: IndexedSeq[ForcableSemaphore],
+               index: Int,
+               step: Int,
+               stepsDone: Int = 0): Option[InstanceId] = {
+    val numInvokers = invokers.size
+
+    if (numInvokers > 0) {
+      val invoker = invokers(index)
+      // If the current invoker is healthy and we can get a slot
+      if (invoker.status == Healthy && dispatched(invoker.id.toInt).tryAcquire()) {
+        Some(invoker.id)
+      } else {
+        // If we've gone through all invokers
+        if (stepsDone == numInvokers + 1) {
+          val healthyInvokers = invokers.filter(_.status == Healthy)
+          if (healthyInvokers.nonEmpty) {
+            // Choose a healthy invoker randomly
+            val random = ThreadLocalRandom.current().nextInt(healthyInvokers.size)
+            dispatched(random).forceAcquire()
+            Some(healthyInvokers(random).id)
+          } else {
+            None
+          }
+        } else {
+          val newIndex = (index + step) % numInvokers
+          schedule(invokers, dispatched, newIndex, step, stepsDone + 1)
+        }
+      }
+    } else {
+      None
+    }
+  }
+}
+
+/**
+ * Holds the state necessary for scheduling of actions.
+ *
+ * @param invokers all of the known invokers in the system
+ * @param managedInvokers all invokers for managed runtimes
+ * @param blackboxInvokers all invokers for blackbox runtimes
+ * @param stepSizes the step-sizes possible for the current invoker count
+ * @param invokerSlots state of accessible slots of each invoker
+ */
+case class ShardingContainerPoolBalancerState(
+  var invokers: IndexedSeq[InvokerHealth] = IndexedSeq.empty[InvokerHealth],
+  var managedInvokers: IndexedSeq[InvokerHealth] = IndexedSeq.empty[InvokerHealth],
+  var blackboxInvokers: IndexedSeq[InvokerHealth] = IndexedSeq.empty[InvokerHealth],
+  var stepSizes: Seq[Int] = ContainerPoolBalancer.pairwiseCoprimeNumbersUntil(0),
+  var invokerSlots: IndexedSeq[ForcableSemaphore] = IndexedSeq.empty[ForcableSemaphore])(
+  lbConfig: ShardingContainerPoolBalancerConfig =
+    loadConfigOrThrow[ShardingContainerPoolBalancerConfig](ConfigKeys.loadbalancer))(implicit logging: Logging) {
+
+  private val totalInvokerThreshold = lbConfig.invokerBusyThreshold
+  private var currentInvokerThreshold = totalInvokerThreshold
+
+  private val blackboxFraction: Double = Math.max(0.0, Math.min(1.0, lbConfig.blackboxFraction))
+  logging.info(this, s"blackboxFraction = $blackboxFraction")(TransactionId.loadbalancer)
+
+  /**
+   * Updates the scheduling state with the new invokers.
+   *
+   * This is okay to not happen atomically since dirty reads of the values set are not dangerous. It is important though
+   * to update the "invokers" variables last, since they will determine the range of invokers to choose from.
+   *
+   * Handling a shrinking invokers list is not necessary, because InvokerPool won't shrink its own list but rather
+   * report the invoker as "Offline".
+   *
+   * It is important that this method does not run concurrently to itself and/or to `updateCluster`
+   */
+  def updateInvokers(newInvokers: IndexedSeq[InvokerHealth]): Unit = {
+    val oldSize = invokers.size
+    val newSize = newInvokers.size
+
+    if (oldSize != newSize) {
+      stepSizes = ContainerPoolBalancer.pairwiseCoprimeNumbersUntil(newSize)
+      if (oldSize < newSize) {
+        // Keeps the existing state..
+        invokerSlots = invokerSlots.padTo(newSize, new ForcableSemaphore(currentInvokerThreshold))
+      }
+    }
+
+    val blackboxes = Math.max(1, (newSize.toDouble * blackboxFraction).toInt)
+    val managed = Math.max(1, newSize - blackboxes)
+
+    invokers = newInvokers
+    blackboxInvokers = invokers.takeRight(blackboxes)
+    managedInvokers = invokers.take(managed)
+
+    logging.info(
+      this,
+      s"loadbalancer invoker status updated. managedInvokers = $managed blackboxInvokers = $blackboxes")(
+      TransactionId.loadbalancer)
+  }
+
+  /**
+   * Updates the size of a cluster. Throws away all state for simplicity.
+   *
+   * This is okay to not happen atomically, since a dirty read of the values set are not dangerous. At worst the
+   * scheduler works on outdated invoker-load data which is acceptable.
+   *
+   * It is important that this method does not run concurrently to itself and/or to `updateState`
+   */
+  def updateCluster(newSize: Int): Unit = {
+    val actualSize = newSize max 1 // if a cluster size < 1 is reported, falls back to a size of 1 (alone)
+    val newTreshold = totalInvokerThreshold / actualSize
 
 Review comment:
   Considering a hypothetical case when the number of available controllers get > then the threshold, then we won't be able to schedule. Should we also set the value to `val newTreshold = (totalInvokerThreshold / actualSize) max 1` as above?

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services