SaurabhChawla100 commented on a change in pull request #27636: 
[SPARK-30873][CORE][YARN]Handling Node Decommissioning for Yarn cluster manger 
in Spark
URL: https://github.com/apache/spark/pull/27636#discussion_r393149928
 
 

 ##########
 File path: 
core/src/main/scala/org/apache/spark/scheduler/DecommissionTracker.scala
 ##########
 @@ -0,0 +1,405 @@
+/*
+ * 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 java.text.SimpleDateFormat
+import java.util.concurrent.TimeUnit
+import java.util.concurrent.atomic.{AtomicBoolean, AtomicInteger}
+
+import scala.collection.mutable.HashMap
+
+import org.apache.spark.{ExecutorAllocationClient, SparkConf, SparkContext}
+import org.apache.spark.annotation.DeveloperApi
+import org.apache.spark.internal.Logging
+import org.apache.spark.internal.config
+import org.apache.spark.util.{Clock, SystemClock, ThreadUtils, Utils}
+
+/**
+ * DecommissionTracker tracks the list of decommissioned nodes.
+ * This decommission trackers maintains the decommissioned nodes state.
+ * Decommission tracker schedules the executor decommission and shuffle
+ * decommission for that node.
+ */
+private[scheduler] class DecommissionTracker (
+  conf: SparkConf,
+  executorAllocClient: Option[ExecutorAllocationClient],
+  dagScheduler: Option[DAGScheduler],
+  clock: Clock = new SystemClock()) extends Logging {
+
+  def this(sc: SparkContext,
+           client: Option[ExecutorAllocationClient],
+           dagScheduler: Option[DAGScheduler]) = {
+    this(sc.conf, client, dagScheduler)
+  }
+
+  // Decommission thread of node decommissioning!!
+  private val decommissionThread =
+    
ThreadUtils.newDaemonThreadPoolScheduledExecutor("node-decommissioning-thread", 
20)
 
 Review comment:
   So the above code creates the thread pool of 20. A new thread will be 
created if none of the thread is free in the thread pool after scheduled time 
for the executors and shuffle decommission(executorsDecomissionTimeMs and 
ShuffleDecosmmissionTimeMs). These active threads completes the task in less 
than second (for executorDecommission it will kill all the executors on the 
host and for shuffleDecommission remove all the entries of shuffle data for 
that node from the MapOutputTracker).
   
   **You are also calling executedecomission separate from the shuffle 
decommission and then have delays to have them happen not at the same time. 
Couldn't they be done in sequence and then you know they aren't done at same 
time?** - So the reason why it is not done in sequence . If inside  
executordecommission after decommission of executors the same thread wait for 
some time and than do the shuffleDecommission . In this case we are holding 
thread for long time which can be free in 1 sec. Now consider a scenario where 
aws spotloss happened after 120 sec of receiving the decommission event, As per 
the logic written executor decommission will takes place after 50% time i.e. 
after 60 sec  and shuffleDecommission will happen after 90% of the time 108 
sec. If we do this sequentially than we have to hold current thread for another 
(108-60=48) 48 secs. So there are more chances of creating more thread from the 
pool if we do it sequentially.
   
   Now as per the current code the executor decommission will takes place after 
50% time i.e. after 60 sec. ExecutorDecommissionThread can take one of the 
thread completes its task in max 1 sec and than that thread is free and same is 
for shuffle decommission which will happened after 90% time 108sec completes 
the task in 1 sec and than the thread is free again. This free thread can be 
used by another Executordecommision/Shuffle decommission if decommission is 
received for  multiple nodes.
   
   Its very rare scenario when all the 20 threads would be active where the 
decommission is very frequent. I think we can have the config for the thread 
pool and believe 8 would be good number and if someone facing the issue very 
frequent  decommission than they can change it 

----------------------------------------------------------------
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.
 
For queries about this service, please contact Infrastructure at:
[email protected]


With regards,
Apache Git Services

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to