Github user mccheah commented on a diff in the pull request:

    https://github.com/apache/spark/pull/3638#discussion_r22695144
  
    --- Diff: core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala 
---
    @@ -865,26 +865,6 @@ class DAGScheduler(
         }
     
         if (tasks.size > 0) {
    -      // Preemptively serialize a task to make sure it can be serialized. 
We are catching this
    -      // exception here because it would be fairly hard to catch the 
non-serializable exception
    -      // down the road, where we have several different implementations 
for local scheduler and
    -      // cluster schedulers.
    -      //
    -      // We've already serialized RDDs and closures in taskBinary, but 
here we check for all other
    -      // objects such as Partition.
    -      try {
    -        closureSerializer.serialize(tasks.head)
    -      } catch {
    -        case e: NotSerializableException =>
    -          abortStage(stage, "Task not serializable: " + e.toString)
    -          runningStages -= stage
    -          return
    -        case NonFatal(e) => // Other exceptions, such as 
IllegalArgumentException from Kryo.
    -          abortStage(stage, s"Task serialization failed: 
$e\n${e.getStackTraceString}")
    -          runningStages -= stage
    -          return
    -      }
    -
    --- End diff --
    
    This is the main addition in the patch - to make it so that task 
serialization error handling is only done when the serialization actually 
occurs.
    
    It turns out there are many scenarios where this selective sampling does 
not actually work. For example, when you create an RDD from an in-memory 
collection, perhaps some of the items are serializable but others are not. E.g. 
consider a list of containers, where the first item in the list is an empty 
container, and the second item in the list is a non-empty container with 
non-serializable things.


---
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

Reply via email to