Affects Version/s: 0.22.0
Fix Version/s: None
We have been running the new speculative logic in
HADOOP-2141 done by Andy in our production cluster.
One thing that we observed is that there are significantly low number of speculative reducers.
We have seen usually 100 speculative mappers launched per minute but speculative reducers are usually less than 5.
But reducers are usually where we get complains about having speculative issues.
These two types of tasks has different properties and different needs.
It would be nice if we can configure the slow threshold separately to deal with them separately.
We can add the following config keys to allow setting them independently.