Github user mccheah commented on a diff in the pull request:
https://github.com/apache/spark/pull/21366#discussion_r193934305
--- Diff:
resource-managers/kubernetes/core/src/main/scala/org/apache/spark/scheduler/cluster/k8s/ExecutorPodsSnapshotsStoreImpl.scala
---
@@ -0,0 +1,95 @@
+/*
+ * 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.scheduler.cluster.k8s
+
+import java.util.concurrent.{ExecutorService, ScheduledExecutorService,
TimeUnit}
+
+import com.google.common.collect.Lists
+import io.fabric8.kubernetes.api.model.Pod
+import io.reactivex.disposables.Disposable
+import io.reactivex.functions.Consumer
+import io.reactivex.schedulers.Schedulers
+import io.reactivex.subjects.PublishSubject
+import javax.annotation.concurrent.GuardedBy
+import scala.collection.JavaConverters._
+import scala.collection.mutable
+
+import org.apache.spark.util.{ThreadUtils, Utils}
+
+private[spark] class ExecutorPodsSnapshotsStoreImpl(
+ bufferSnapshotsExecutor: ScheduledExecutorService,
+ executeSubscriptionsExecutor: ExecutorService)
+ extends ExecutorPodsSnapshotsStore {
+
+ private val SNAPSHOT_LOCK = new Object()
+
+ private val snapshotsObservable =
PublishSubject.create[ExecutorPodsSnapshot]()
+ private val observedDisposables = mutable.Buffer.empty[Disposable]
+
+ @GuardedBy("SNAPSHOT_LOCK")
+ private var currentSnapshot = ExecutorPodsSnapshot()
+
+ override def addSubscriber(
+ processBatchIntervalMillis: Long)
+ (subscriber: ExecutorPodsSnapshot => Unit): Unit = {
+ observedDisposables += snapshotsObservable
+ // Group events in the time window given by the caller. These
buffers are then sent
+ // to the caller's lambda at the given interval, with the pod
updates that occurred
+ // in that given interval.
+ .buffer(
+ processBatchIntervalMillis,
+ TimeUnit.MILLISECONDS,
+ // For testing - specifically use the given scheduled executor
service to trigger
+ // buffer boundaries. Allows us to inject a deterministic
scheduler here.
+ Schedulers.from(bufferSnapshotsExecutor))
+ // Trigger an event cycle immediately. Not strictly required to be
fully correct, but
+ // in particular the pod allocator should try to request executors
immediately instead
+ // of waiting for one pod allocation delay.
+ .startWith(Lists.newArrayList(ExecutorPodsSnapshot()))
+ // Force all triggered events - both the initial event above and the
buffered ones in
+ // the following time windows - to execute asynchronously to this
call's thread.
+ .observeOn(Schedulers.from(executeSubscriptionsExecutor))
+ .subscribe(toReactivexConsumer { snapshots:
java.util.List[ExecutorPodsSnapshot] =>
+ Utils.tryLogNonFatalError {
+ snapshots.asScala.foreach(subscriber)
+ }
+ })
+ }
+
+ override def stop(): Unit = {
+ observedDisposables.foreach(_.dispose())
+ snapshotsObservable.onComplete()
+ ThreadUtils.shutdown(bufferSnapshotsExecutor)
+ ThreadUtils.shutdown(executeSubscriptionsExecutor)
+ }
+
+ override def updatePod(updatedPod: Pod): Unit =
SNAPSHOT_LOCK.synchronized {
+ currentSnapshot = currentSnapshot.withUpdate(updatedPod)
--- End diff --
So the watch only creates new snapshots by applying some "diff" according
to the next event to the previous snapshot. One downside of this data model is
that we end up buffering multiple collections of pods in the observable stream
which can all perhaps only differ by a single pod per update. Thus we end up
temporarily storing redundant information in the snapshots. But the observable
buffers are ephemeral and will be periodically processed by the periodic
iterations of the subscribers. I wouldn't mind thinking about a more optimal
representation here.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
