On Fri, May 17, 2013 at 9:33 AM, Carsten Ziegeler <cziege...@apache.org> wrote: > ...If you see any issues with releasing this stuff, please speak up now...
I had another (quick) look at the tests, and unless I missed something it's still mostly or only happy cases that are tested. I personally wouldn't trust this stuff until torture tests that break it are implemented...multi-instance discovery and consensus will usually break when someone pulls power and network plugs randomly, and the question is then how robust the recovery algorithms are. Such simulations can be hard to implement, but maybe just running lots of test threads and killing some of them at random would help discover how the discovery implementation handles unexpected failures and timeouts. We can let such tests run on Jenkins even if they are slow. This doesn't prevent us from releasing the API, of course, but for the implementation it might be good to include a note about what level of testing exists so far, to avoid unrealistic expectations. -Bertrand
