Github user andrewor14 commented on a diff in the pull request:
https://github.com/apache/spark/pull/12154#discussion_r58762264
--- Diff:
streaming/src/main/scala/org/apache/spark/streaming/scheduler/ExecutorAllocationManager.scala
---
@@ -0,0 +1,233 @@
+/*
+ * 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.streaming.scheduler
+
+import scala.util.Random
+
+import org.apache.spark.{ExecutorAllocationClient, SparkConf}
+import org.apache.spark.internal.Logging
+import org.apache.spark.streaming.util.RecurringTimer
+import org.apache.spark.util.{Clock, Utils}
+
+/**
+ * Class that manages executor allocated to a StreamingContext, and
dynamically request or kill
+ * executors based on the statistics of the streaming computation. This is
different from the core
+ * dynamic allocation policy; the core policy relies on executors being
idle for a while, but the
+ * micro-batch model of streaming prevents any particular executors from
being idle for a long
+ * time. Instead, the measure of "idle-ness" needs to be based on the time
taken to process
+ * each batch.
+ *
+ * At a high level, the policy implemented by this class is as follows:
+ * - Use StreamingListener interface get batch processing times of
completed batches
+ * - Periodically take the average batch completion times and compare with
the batch interval
+ * - If (avg. proc. time / batch interval) >= scaling up ratio, then
request more executors.
+ * The number of executors requested is based on the ratio = (avg. proc.
time / batch interval).
+ * - If (avg. proc. time / batch interval) <= scaling down ratio, then try
to kill a executor that
+ * is not running a receiver.
+ *
+ * This features should ideally be used in conjunction with backpressure,
as backpressure ensures
+ * system stability, while executors are being readjusted.
+ */
+private[streaming] class ExecutorAllocationManager(
+ client: ExecutorAllocationClient,
+ receiverTracker: ReceiverTracker,
+ conf: SparkConf,
+ batchDurationMs: Long,
+ clock: Clock) extends StreamingListener with Logging {
+
+ import ExecutorAllocationManager._
+
+ private val scalingIntervalSecs = conf.getTimeAsSeconds(
+ SCALING_INTERVAL_KEY,
+ s"${SCALING_INTERVAL_DEFAULT_SECS}s")
+ private val scalingUpRatio = conf.getDouble(SCALING_UP_RATIO_KEY,
SCALING_UP_RATIO_DEFAULT)
+ private val scalingDownRatio = conf.getDouble(SCALING_DOWN_RATIO_KEY,
SCALING_DOWN_RATIO_DEFAULT)
+ private val minNumExecutors = conf.getInt(
+ MIN_EXECUTORS_KEY,
+ math.max(1, receiverTracker.numReceivers))
+ private val maxNumExecutors = conf.getInt(MAX_EXECUTORS_KEY,
Integer.MAX_VALUE)
+ private val timer = new RecurringTimer(clock, scalingIntervalSecs * 1000,
+ _ => manageAllocation(), "streaming-executor-allocation-manager")
+
+ @volatile private var batchProcTimeSum = 0L
+ @volatile private var batchProcTimeCount = 0
+
+ validateSettings()
+
+ def start(): Unit = {
+ timer.start()
+ logInfo(s"ExecutorAllocationManager started with " +
+ s"ratios = [$scalingUpRatio, $scalingDownRatio] and interval =
$scalingIntervalSecs sec")
+ }
+
+ def stop(): Unit = {
+ timer.stop(interruptTimer = true)
+ logInfo("ExecutorAllocationManager stopped")
+ }
+
+ /**
+ * Manage executor allocation by requesting or killing executors based
on the collected
+ * batch statistics.
+ */
+ private def manageAllocation(): Unit = synchronized {
+ logInfo(s"Managing executor allocation with ratios = [$scalingUpRatio,
$scalingDownRatio]")
+ if (batchProcTimeCount > 0) {
+ val averageBatchProcTime = batchProcTimeSum / batchProcTimeCount
+ val ratio = averageBatchProcTime.toDouble / batchDurationMs
+ logInfo(s"Average: $averageBatchProcTime, ratio = $ratio" )
+ if (ratio >= scalingUpRatio) {
+ logDebug("Requesting executors")
+ val numNewExecutors = math.max(math.round(ratio).toInt, 1)
+ requestExecutors(numNewExecutors)
+ } else if (ratio <= scalingDownRatio) {
+ logDebug("Killing executors")
+ killExecutor()
+ }
+ }
+ batchProcTimeSum = 0
+ batchProcTimeCount = 0
+ }
+
+ /** Request the specified number of executors over the currently active
one */
+ private def requestExecutors(numNewExecutors: Int): Unit = {
+ require(numNewExecutors >= 1)
+ val allExecIds = client.getExecutorIds()
+ logDebug(s"Executors (${allExecIds.size}) = ${allExecIds}")
+ val targetTotalExecutors =
+ math.max(math.min(maxNumExecutors, allExecIds.size +
numNewExecutors), minNumExecutors)
+ client.requestTotalExecutors(targetTotalExecutors, 0, Map.empty)
+ logInfo(s"Requested total $targetTotalExecutors executors")
+ }
+
+ /** Kill a executors that is not running a receiver */
--- End diff --
grammar
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
