Details
-
Improvement
-
Status: Open
-
Major
-
Resolution: Unresolved
-
None
-
None
-
None
Description
For many input types, it is possible to restore the state of a RecordReader by writing a new split for the remaining data (e.g., storing the inflater state with a file offset for gzip text). Similarly, an InputFormat may be capable of further subdividing an InputSplit into smaller tasks to be completed in parallel.
Since some splits are semantically significant, whether a MapTask can be preempted in this way is not necessarily a property of the InputFormat, only. Minimally, we should allow the user to disable this capability.
Attachments
Issue Links
- is part of
-
MAPREDUCE-4584 Umbrella: Preemption and restart of MapReduce tasks
- Open