On Aug 26, 2009, at 8:15 AM, Steve Loughran wrote:
I havent' seen those. As owen says, swallowing interruptions is naughty, I don't know what the hadoop-wide policy is on processing them; something to look at.
It is currently broken. Currently the biggest breakage is the RPC layer that swallows them in a really evil way. We need to fix that, but it is a big job. In particular, all of the methods in all of the protocols need to have InterruptedExceptions added to their signatures. That is pretty painful.
The requirement is that in all cases, threads that are interrupted should exit quickly. We've found many of the problems by livelocks in the mini-mr and mini-hdfs based test cases, but really someone should take a sweep through and fix them all.
-- Owen
