Github user squito commented on a diff in the pull request:
https://github.com/apache/spark/pull/11105#discussion_r56421658
--- Diff: core/src/main/scala/org/apache/spark/Accumulable.scala ---
@@ -146,6 +212,32 @@ class Accumulable[R, T] private (
def merge(term: R) { value_ = param.addInPlace(value_, term)}
/**
+ * Merge in pending updates for ac consistent accumulators or merge
accumulated values for
+ * regular accumulators. This is only called on the driver when merging
task results together.
+ */
+ private[spark] def internalMerge(term: Any) {
+ if (!consistent) {
+ merge(term.asInstanceOf[R])
+ } else {
+ mergePending(term.asInstanceOf[mutable.HashMap[(Int, Int, Int), R]])
+ }
+ }
+
+ /**
+ * Merge another Accumulable's pending updates, checks to make sure that
each pending update has
+ * not already been processed before updating.
+ */
+ private[spark] def mergePending(term: mutable.HashMap[(Int, Int, Int),
R]) = {
+ term.foreach{case ((rddId, shuffleId, splitId), v) =>
+ val splits = processed.getOrElseUpdate((rddId, shuffleId), new
mutable.BitSet())
+ if (!splits.contains(splitId)) {
+ splits += splitId
+ value_ = param.addInPlace(value_, v)
+ }
--- End diff --
I don't think you need both `completed` and `processed` -- they seem to be
doing more or less the same thing. You could change this to:
```scala
term.foreach{case (partitionKey, v) =>
if (!completed.contains(partitionKey)) {
completed += partitionKey
value_ = param.addInPlace(value_, v)
}
}
```
If I understand correctly, there is a bit of logical distinction between
the two -- `processed` was being used on the driver, that is across multiple
tasks, to track what had been accumulated and what hadn't. `completed`, OTOH,
was being used on the executors, to track the updates coming from *one task*.
Typically it would only contain a single entry, for the one `(rdd, shuffle,
partition)` tuple that was being used, since tasks *usually* only compute one
partition, but that isn't true when there is a coalesce involved.
If that explanation sounds correct, its probably best to put it into a
comment somewhere, and I'd still say that you can eliminate `processed` and
just explain how `completed` will get used in two different ways on the
executors and on the driver.
---
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]
