I created HBASE-4545 to clean up TestHLog. Among other things, it didn't even need to spin up a MiniHBaseCluster, only a MiniDFSCluster. Several of the tests were also leaking unclosed HLog instances.
Will post a patch shortly. On Wed, Oct 5, 2011 at 5:52 PM, Gary Helmling <[email protected]> wrote: > On Wed, Oct 5, 2011 at 3:47 PM, Roman Shaposhnik <[email protected]> wrote: > >> On Wed, Oct 5, 2011 at 12:00 PM, Todd Lipcon <[email protected]> wrote: >> > +CC Roman who worked on the patch identified by the bisect. >> >> Thanks for the CC. I've just subscribed myself to the list. >> >> > Roman, does Gary's analysis make sense to you? >> >> Hm. I need to look into it. There *could* be some issues with the >> mini-cluster, >> since I've enabled a whole new codepath for it. >> >> >> > I'm thinking this may just be sloppiness or lack of cleanup in TestHLog > that was just exposed by the HBASE-4209 change. Even if I revert the change > to JVMClusterUtil from 4209 and instead use the patch to call shutdown in > TestHLog.tearDownAfterClass(), I still get hangs every 3-4 runs when > executing TestHLog in a loop. > > Previously, when TestHLog had neither the shutdown hook nor an explicit > shutdown of the mini cluster, I'm guessing that the test would "pass" even > though the mini cluster never fully shut down. So it's good to have the > shutdown hooks in place for tests to help us ferret out these issues. > > >
