Filed KAFKA-1062, including trivial patch. /Sam
On Sep 19, 2013, at 5:52 PM, Neha Narkhede <neha.narkh...@gmail.com> wrote: > 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 >> >> >> >>
