ShuffleMapStage and pendingPartitions vs isAvailable or findMissingPartitions?

classic Classic list List threaded Threaded
2 messages Options
Reply | Threaded
Open this post in threaded view
|

ShuffleMapStage and pendingPartitions vs isAvailable or findMissingPartitions?

Jacek Laskowski
Hi,

I found that ShuffleMapStage has this (apparently superfluous) pendingPartitions registry [1] for DAGScheduler and the description says:

"  /**
   * Partitions that either haven't yet been computed, or that were computed on an executor
   * that has since been lost, so should be re-computed.  This variable is used by the
   * DAGScheduler to determine when a stage has completed. Task successes in both the active
   * attempt for the stage or in earlier attempts for this stage can cause paritition ids to get
   * removed from pendingPartitions. As a result, this variable may be inconsistent with the pending
   * tasks in the TaskSetManager for the active attempt for the stage (the partitions stored here
   * will always be a subset of the partitions that the TaskSetManager thinks are pending).
   */
"

I'm curious why there is a need for this pendingPartitions since isAvailable or findMissingPartitions (using MapOutputTrackerMaster) know it already and I think are even more up-to-date. Why is there this extra registry?

Reply | Threaded
Open this post in threaded view
|

Re: ShuffleMapStage and pendingPartitions vs isAvailable or findMissingPartitions?

ZHANG Wei
AFAICT, not must have `pendingPartitions`, `mapOutputTrackerMaster` is
added by a later change, `pendingPartitions` can be cleaned up.

--
Cheers,
-z

On Sun, 26 Apr 2020 11:53:09 +0200
Jacek Laskowski <[hidden email]> wrote:

> Hi,
>
> I found that ShuffleMapStage has this (apparently superfluous)
> pendingPartitions registry [1] for DAGScheduler and the description says:
>
> "  /**
>    * Partitions that either haven't yet been computed, or that were
> computed on an executor
>    * that has since been lost, so should be re-computed.  This variable is
> used by the
>    * DAGScheduler to determine when a stage has completed. Task successes
> in both the active
>    * attempt for the stage or in earlier attempts for this stage can cause
> paritition ids to get
>    * removed from pendingPartitions. As a result, this variable may be
> inconsistent with the pending
>    * tasks in the TaskSetManager for the active attempt for the stage (the
> partitions stored here
>    * will always be a subset of the partitions that the TaskSetManager
> thinks are pending).
>    */
> "
>
> I'm curious why there is a need for this pendingPartitions
> since isAvailable or findMissingPartitions (using MapOutputTrackerMaster)
> know it already and I think are even more up-to-date. Why is there this
> extra registry?
>
> [1]
> https://github.com/apache/spark/blob/master/core/src/main/scala/org/apache/spark/scheduler/ShuffleMapStage.scala#L60
>
> Pozdrawiam,
> Jacek Laskowski
> ----
> https://about.me/JacekLaskowski
> "The Internals Of" Online Books <https://books.japila.pl/>
> Follow me on https://twitter.com/jaceklaskowski
>
> <https://twitter.com/jaceklaskowski>

---------------------------------------------------------------------
To unsubscribe e-mail: [hidden email]