> You've never run with an older version of TRUNK? Only a recent one, one that > had 1615 in it?
Nope. > I might have seen this in a recent test run. Let me retry. At least there > is better debug since hbase-1603. I think the problem is indeed related to a split as it occurs less and less as I rerun the MR job (and thus the region count is already high and less splits are needed). If I store records in HBase from the mappers, the problem is even worse when starting with an empty table as records are not sorted and more splits occur more rapidly. Mathias.
