Agreed. This is a regression and is not easy to reason about. This is a side effect of reading the partitions as a set from zookeeper. Please can you file a JIRA to get this fixed? Feel free to upload a patch as well.
Thanks, Neha On Thu, Sep 19, 2013 at 8:17 AM, Sam Meder <sam.me...@jivesoftware.com>wrote: > The latest consumer changes to read data from Zookeeper during rebalance > have made the consumer rebalance code incompatible with older versions > (making rolling upgrades without downtime hard). The problem relates to how > partitions are ordered. The old code seems to have returned the partitions > sorted: > > ... rebalancing the following partitions: ArrayBuffer(0, 1, 2, 3, 4, 5, 6, > 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19) for topic > produce-indexable-views with consumers: ... > > the new code instead uses: > > ... rebalancing the following partitions: List(0, 5, 10, 14, 1, 6, 9, 13, > 2, 17, 12, 7, 3, 18, 16, 11, 8, 19, 4, 15) for topic > produce-indexable-views with consumers: ... > > This causes new consumers and old consumers to claim the same partitions. > I realize that this may not be a big deal (although painful for us since it > disagrees with our deployment automation) since the code wasn't officially > released, but it seems simple enough to sort the partitions if you'd take > such a patch. > > /Sam > > > >
