Test build #24765 has started for PR 3785 at commit 9ca6541.

• This patch merges cleanly.

Test build #24765 has finished for PR 3785 at commit 9ca6541.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/SparkPullRequestBuilder/24765/
Test PASSed.

On the surface, this seems like an okay change. I wonder whether this retry logic could have unexpected consequences. Let me try to reason it out:

• askTracker is only called with GetMapOutputStatuses.
• In the master actor, it calls getSerializedMapOutputStatuses. This method never throws exceptions: if a shuffle is missing, then it just stores an empty array and serializes it.
• It's possible that the serialized map statuses could exceed the Akka frame size (although extremely unlikely and perhaps impossible with the new output status compression techniques). In this case, though, the master would throw an exception and fail to send a reply back to the asker. In this case, with this patch we'd end up performing a bunch of retries for an operation that will ultimately fail, so we'll take longer to detect a failure.

In the common cases, though, this seems fine, even if the map output statuses are missing (since it won't introduce a bunch of futile retries). Therefore, I think we should pull this in; I don't know if this fixes an actual bug, but it seems like it could make things more robust.

Alright, I'm going to merge this into master (1.3.0). Thanks!