guiyanakuang commented on a change in pull request #34743:
URL: https://github.com/apache/spark/pull/34743#discussion_r762520646



##########
File path: core/src/main/scala/org/apache/spark/scheduler/TaskSetManager.scala
##########
@@ -259,21 +259,29 @@ private[spark] class TaskSetManager(
       loc match {
         case e: ExecutorCacheTaskLocation =>
           pendingTaskSetToAddTo.forExecutor.getOrElseUpdate(e.executorId, new 
ArrayBuffer) += index
+          pendingTaskSetToAddTo.forHost.getOrElseUpdate(loc.host, new 
ArrayBuffer) += index
         case e: HDFSCacheTaskLocation =>
           val exe = sched.getExecutorsAliveOnHost(loc.host)
           exe match {
             case Some(set) =>
               for (e <- set) {
                 pendingTaskSetToAddTo.forExecutor.getOrElseUpdate(e, new 
ArrayBuffer) += index
               }
+              pendingTaskSetToAddTo.forHost.getOrElseUpdate(loc.host, new 
ArrayBuffer) += index
               logInfo(s"Pending task $index has a cached location at ${e.host} 
" +
                 ", where there are executors " + set.mkString(","))
             case None => logDebug(s"Pending task $index has a cached location 
at ${e.host} " +
               ", but there are no executors alive there.")
           }
-        case _ =>
+        case _: HostTaskLocation =>
+          val exe = sched.getExecutorsAliveOnHost(loc.host)
+          exe match {
+            case Some(_) =>
+              pendingTaskSetToAddTo.forHost.getOrElseUpdate(loc.host, new 
ArrayBuffer) += index
+            case _ => logDebug(s"Pending task $index has a location at 
${loc.host} " +
+              ", but there are no executors alive there.")
+          }
       }
-      pendingTaskSetToAddTo.forHost.getOrElseUpdate(loc.host, new ArrayBuffer) 
+= index

Review comment:
       Yes, I totally agree with you and after discussing with you I realized 
that this pr just solves my special case and breaks other cases.
   
   The reason I haven't closed this pr yet is because I still want to find a 
way to fix this situation with minimal modifications, while not affecting other 
scenarios that are already working well.
   
   @mridulm, thank you for spending a lot of time discussing this and giving me 
new insight into spark scheduling as well.
   
   
   
   




-- 
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.

To unsubscribe, e-mail: [email protected]

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



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

Reply via email to