On Thu, Dec 20, 2012 at 1:31 PM, Stack <[email protected]> wrote: > On Thu, Dec 20, 2012 at 12:01 PM, Jesse Yates <[email protected] > >wrote: > > > Not being intimately familiar with the replication tests, how well > covered > > on that stuff are we by the remaining unit tests? Also, since the > hbase-it > > tests still get run, are we really gaining anything in terms of CI > > reliability? > > > > The hbase-it tests are not run up on jenkins Jesse, not at the moment at > least. > > I'm just a bit worried that we will break things and not have it caught on check-in (the real arbiter of what's a valid patch), meaning we will break the branch without realizing. :-/ We need some way to ensure that the underlying code is covered.
At the minimum it needs to be part of the release checklist that we run the replication IT test on two real clusters (assuming this is a black-box test and not messing with things too much). I don't expect needing more than small functional clusters (say max 5 nodes?) to test this adequately. The Jenkins machines don't seem sufficient for this, so my gut feel is that it will have to be a release item that the RM needs to verify works (either personally or by proxy). This could even apply to all the hbase-it, for releases going forward Ideally, we will also have some unit tests that subsume some of the tested functionality when/if we move them to a more infrequent tests, though its hard to say how possible/useful this would be to manage in practice. Short term, maybe we disable them and file a Jira to get them rock solid (or on -it and tested regularly somehow)? This goes back to the long standing discussion that we should disable flappers until they tell us something useful. Again don't know those tests very well, so these are just general platitudes :) > Agree to moving TestReplication and perhaps some of the mapreduce tests out > of unit test core. > > That means we should set up a regular run of hbase-it tests somewhere > though? > > St.Ack >
