Github user pwendell commented on a diff in the pull request:
https://github.com/apache/spark/pull/126#discussion_r10553029
--- Diff: core/src/main/scala/org/apache/spark/MapOutputTracker.scala ---
@@ -181,15 +178,50 @@ private[spark] class MapOutputTracker(conf:
SparkConf) extends Logging {
}
}
+/**
+ * MapOutputTracker for the workers. This uses BoundedHashMap to keep
track of
+ * a limited number of most recently used map output information.
+ */
+private[spark] class MapOutputTrackerWorker(conf: SparkConf) extends
MapOutputTracker(conf) {
+
+ /**
+ * Bounded HashMap for storing serialized statuses in the worker. This
allows
+ * the HashMap stay bounded in memory-usage. Things dropped from this
HashMap will be
+ * automatically repopulated by fetching them again from the driver. Its
okay to
+ * keep the cache size small as it unlikely that there will be a very
large number of
+ * stages active simultaneously in the worker.
+ */
+ protected val mapStatuses = new BoundedHashMap[Int, Array[MapStatus]](
--- End diff --
I'm suggesting this - when a shuffle dependency goes out of scope in the
driver we can find all of the associated stages. Once the stages are located
then we tell each Executor to clean up everything corresponding to the stage.
This would include the map status and any other information. Right now this
might have to go through a BlockManager message or something like that.
Basically I'm suggesting we try to do principled garbage collection based on
things going out of scope whenever possible.
---
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.
---
