Details
-
Bug
-
Status: Open
-
Minor
-
Resolution: Unresolved
-
3.0.0
-
None
-
None
Description
I recently encountered a problem where jobs non-deterministically failed with
TaskResultLost (result lost from block manager)
exceptions.
It turned out that this was due to some sort of networking issue where the Spark driver was unable to initiate outgoing connections to executors' block managers in order to fetch indirect task results.
In this situation, the error message was slightly misleading: the "result lost from block manager" makes it sound like we received an error / block-not-found response from the remote host, whereas in my case the problem was actually a network connectivity issue where we weren't even able to connect in the first place.
If it's easy to do so, it might be nice to refine the error-handling / logging code so that we distinguish between the receipt of an error response vs. a lower-level networking / connectivity issue.