Github user squito commented on a diff in the pull request:
https://github.com/apache/spark/pull/8760#discussion_r45771875
--- Diff:
core/src/main/scala/org/apache/spark/scheduler/BlacklistStrategy.scala ---
@@ -0,0 +1,146 @@
+/*
+ * 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
+
+import scala.collection.mutable
+import org.apache.spark.Logging
+import org.apache.spark.SparkConf
+import org.apache.spark.util.SystemClock
+import org.apache.spark.util.Clock
+
+/**
+ * The interface to determine executor blacklist and node blacklist.
+ */
+private [scheduler] trait BlacklistStrategy {
+ /** Defined a time interval to expire failure information of executors */
+ val expireTimeInMilliseconds: Long
+
+ /** Return executors in blacklist which are related to given taskIndex */
+ def getExecutorBlacklist(
+ executorIdToFailureStatus: mutable.HashMap[String, FailureStatus],
+ atomTask: BlacklistAtomTask, clock: Clock): Set[String]
+
+ /** Return all nodes in blacklist */
+ def getNodeBlacklist(
+ executorIdToFailureStatus: mutable.HashMap[String, FailureStatus]):
Set[String]
+
+ /**
+ * Default implementation to remove failure executors from HashMap based
on given time period.
+ * The return value identity whether or not it updated anything
+ */
+ def expireExecutorsInBlackList(
+ executorIdToFailureStatus: mutable.HashMap[String, FailureStatus],
clock: Clock): Boolean = {
+ val now = clock.getTimeMillis()
+ val expiredKey = executorIdToFailureStatus.filter {
+ case (executorid, failureStatus) => {
+ (now - failureStatus.updatedTime) >= expireTimeInMilliseconds
+ }
+ }.keySet
+
+ if (expiredKey.isEmpty) {
+ false
+ } else {
+ executorIdToFailureStatus --= expiredKey
+ true
+ }
+ }
+}
+
+/**
+ * This strategy is simply based on given threshold and is taskId
unrelated. An executor will be
+ * in blacklist, if it failed more than "maxFailureTaskNumber" times. A
node will be in blacklist,
+ * if there are more than "maxBlackExecutorNumber" executors on it in
executor blacklist.
+ *
+ * In this case, provided taskId will be ignored. The benefit for taskId
unrelated strategy is that
+ * different taskSets can learn experience from other taskSet to avoid
allocating tasks on
+ * problematic executors.
+ */
+private[scheduler] class SimpleStrategy(
+ maxFailureTaskNumber: Int,
+ maxBlackExecutorNumber: Int,
+ val expireTimeInMilliseconds: Long
+ )extends BlacklistStrategy {
+
+ private def getSelectedExecutorMap(
+ executorIdToFailureStatus: mutable.HashMap[String, FailureStatus]) =
{
+ executorIdToFailureStatus.filter{
+ case (id, failureStatus) => failureStatus.totalNumFailures >
maxFailureTaskNumber
+ }
+ }
+
+ // As this is a taskId unrelated strategy, the input taskId will be
ignored
+ def getExecutorBlacklist(
+ executorIdToFailureStatus: mutable.HashMap[String, FailureStatus],
+ atomTask: BlacklistAtomTask, clock: Clock): Set[String] = {
+ getSelectedExecutorMap(executorIdToFailureStatus).keys.toSet
+ }
+
+ def getNodeBlacklist(
+ executorIdToFailureStatus: mutable.HashMap[String, FailureStatus]):
Set[String] = {
+ getSelectedExecutorMap(executorIdToFailureStatus)
+ .groupBy{case (id, failureStatus) => failureStatus.host}
+ .filter {case (host, executorIdToFailureStatus) =>
+ executorIdToFailureStatus.size > maxBlackExecutorNumber}
+ .keys.toSet
+ }
+}
+
+/**
+ * This strategy is applied as default to keep the same semantics as
original. It's an taskId
+ * related strategy. If an executor failed running "task A", then we think
this executor is
+ * blacked for "task A". And we think the executor is still healthy for
other task. node blacklist
+ * is always empty.
+ *
+ * It was the standard behavior before spark 1.6
+ */
+private[scheduler] class DefaultStrategy(
+ val expireTimeInMilliseconds: Long) extends BlacklistStrategy {
+ def getExecutorBlacklist(
+ executorIdToFailureStatus: mutable.HashMap[String, FailureStatus],
+ atomTask: BlacklistAtomTask, clock: Clock): Set[String] = {
+ executorIdToFailureStatus.filter{
+ case (_, failureStatus) =>
failureStatus.numFailuresPerTask.keySet.contains(atomTask) &&
+ clock.getTimeMillis() - failureStatus.updatedTime <
expireTimeInMilliseconds
+ }.keys.toSet
+ }
+
+ def getNodeBlacklist(
+ executorIdToFailureStatus: mutable.HashMap[String, FailureStatus]):
Set[String] =
+ Set.empty[String]
+}
+
+/**
+ * Create BlacklistStrategy instance according to SparkConf
+ */
+private[scheduler] object BlacklistStrategy {
+ def apply(sparkConf: SparkConf): BlacklistStrategy = {
+ val timeout =
sparkConf.getTimeAsMs("spark.scheduler.blacklist.timeout",
+ sparkConf.getLong("spark.scheduler.executorTaskBlacklistTime",
0L).toString() + "ms")
+ sparkConf.get("spark.scheduler.blacklist.strategy", "default") match {
+ case "default" =>
+ new DefaultStrategy(timeout)
+ case "threshold" =>
+ new SimpleStrategy(
+
sparkConf.getInt("spark.scheduler.blacklist.threshold.maxFailureTaskNumber", 3),
+
sparkConf.getInt("spark.scheduler.blacklist.threshold.maxBlackExecutorNumber",
3),
+ timeout)
+ case unsupported =>
+ throw new Exception(s"No matching blacklist strategy for:
$unsupported")
--- End diff --
IllegalArgumentException
---
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 [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
