On May 5, 2011, at 1:56 PM, Jakob Homan wrote: > +1 > > Downloaded, verified, tested on single node cluster to my > satisfaction. We've also brought this release up on a sizable cluster > and checked its basic sanity.
All of you people doing single node tests are missing stuff. For example, the regression in how the secondary namenode addr stuff works vs. 0.20. By far, the biggest problem we've found is that the capacity scheduler documentation doesn't actually match what the code does. I have a hunch that the unit tests were written/change to match the outcome, rather than test what is supposed to happen. For us, this breakage makes it unusable out of the box and we'll likely either go back to our (relatively stable) backport of 0.21's cap sched, try to fix the 0.20.203 code, or maybe even switch to a completely different scheduler.