Hey Andy, Thanks for jumping in. A couple comments:
In addition, I think it is important that during a rebalance consumers do > not first have all partitions revoked, only to have a very similar, (or the > same!), set reassigned. This is less than initiative and complicates client > code unnecessarily. Instead, the `ConsumerPartitionListener` should only be > called for true changes in assignment I.e. any new partitions assigned and > any existing ones revoked, when comparing the new assignment to the > previous one. The problem is that the revocation callback is called before you know what the assignment for the next generation will be. This is necessary for the consumer to be able to commit offsets for its assigned partitions. Once the consumer has a new assignment, it is no longer safe to commit offsets from the previous generation. Unless sticky assignment can give us some guarantee on which partitions will remain after the rebalance, all of them must be included in the revocation callback. > There is one last scenario I'd like to highlight that I think the KIP > should describe: say you have a group consuming from two topics, each topic > with two partitions. As of 0.9.0.1 the maximum number of consumers you can > have is 2, not 4. With 2 consumers each will get one partition from each > topic. A third consumer with not have any partitions assigned. This should > be fixed by the 'fair' part of the strategy, but it would be good to see > this covered explicitly in the KIP. This would be true for range assignment, but with 4 partitions total, round-robin assignment would give one partition to each of the 4 consumers (assuming subscriptions match). Thanks, Jason On Thu, Jun 23, 2016 at 1:42 AM, Andrew Coates <big.andy.coa...@gmail.com> wrote: > Hi all, > > I think sticky assignment is immensely important / useful in many > situations. Apps that use Kafka are many and varied. Any app that stores > any state, either in the form of data from incoming messages, cached > results from previous out-of-process calls or expensive operations, (and > let's face it, that's most!), can see a big negative impact from partition > movement. > > The main issue partition movement brings is that it makes building elastic > services very hard. Consider: you've got an app consuming from Kafka that > locally caches data to improve performance. You want the app to auto scale > as the throughout to the topic(s) increases. Currently, when one or more > new instance are added and the group rebalances, all existing instances > have all partitions revoked, and then a new, potentially quite different, > set assigned. An intuitive pattern is to evict partition state, I.e. the > cached data, when a partition is revoked. So in this case all apps flush > their entire cache causing throughput to drop massively, right when you > want to increase it! > > Even if the app is not flushing partition state when partitions are > revoked, the lack of a 'sticky' strategy means that a proportion of the > cached state is now useless, and instances have partitions assigned for > which they have no cached state, again negatively impacting throughout. > > With a 'sticky' strategy throughput can be maintained and indeed increased, > as intended. > > The same is also true in the presence of failure. An instance failing, > (maybe due to high load), can invalidate the caching of existing instances, > negatively impacting throughout of the remaining instances, (possibly at a > time the system needs throughput the most!) > > My question would be 'why move partitions if you don't have to?'. I will > certainly be setting the 'sticky' assignment strategy as the default once > it's released, and I have a feeling it will become the default in the > communitie's 'best-practice' guides. > > In addition, I think it is important that during a rebalance consumers do > not first have all partitions revoked, only to have a very similar, (or the > same!), set reassigned. This is less than initiative and complicates client > code unnecessarily. Instead, the `ConsumerPartitionListener` should only be > called for true changes in assignment I.e. any new partitions assigned and > any existing ones revoked, when comparing the new assignment to the > previous one. > > I think the change to how the client listener is called should be part of > this work. > > There is one last scenario I'd like to highlight that I think the KIP > should describe: say you have a group consuming from two topics, each topic > with two partitions. As of 0.9.0.1 the maximum number of consumers you can > have is 2, not 4. With 2 consumers each will get one partition from each > topic. A third consumer with not have any partitions assigned. This should > be fixed by the 'fair' part of the strategy, but it would be good to see > this covered explicitly in the KIP. > > Thanks, > > > Andy > > > > > > > > > On Thu, 23 Jun 2016, 00:41 Jason Gustafson, <ja...@confluent.io> wrote: > > > Hey Vahid, > > > > Thanks for the updates. I think the lack of comments on this KIP suggests > > that the motivation might need a little work. Here are the two main > > benefits of this assignor as I see them: > > > > 1. It can give a more balanced assignment when subscriptions do not match > > in a group (this is the same problem solved by KIP-49). > > 2. It potentially allows applications to save the need to cleanup > partition > > state when rebalancing since partitions are more likely to stay assigned > to > > the same consumer. > > > > Does that seem right to you? > > > > I think it's unclear how serious the first problem is. Providing better > > balance when subscriptions differ is nice, but are rolling updates the > only > > scenario where this is encountered? Or are there more general use cases > > where differing subscriptions could persist for a longer duration? I'm > also > > wondering if this assignor addresses the problem found in KAFKA-2019. It > > would be useful to confirm whether this problem still exists with the new > > consumer's round robin strategy and how (whether?) it is addressed by > this > > assignor. > > > > The major selling point seems to be the second point. This is definitely > > nice to have, but would you expect a lot of value in practice since > > consumer groups are usually assumed to be stable? It might help to > describe > > some specific use cases to help motivate the proposal. One of the > downsides > > is that it requires users to restructure their code to get any benefit > from > > it. In particular, they need to move partition cleanup out of the > > onPartitionsRevoked() callback and into onPartitionsAssigned(). This is a > > little awkward and will probably make explaining the consumer more > > difficult. It's probably worth including a discussion of this point in > the > > proposal with an example. > > > > Thanks, > > Jason > > > > > > > > On Tue, Jun 7, 2016 at 4:05 PM, Vahid S Hashemian < > > vahidhashem...@us.ibm.com > > > wrote: > > > > > Hi Jason, > > > > > > I updated the KIP and added some details about the user data, the > > > assignment algorithm, and the alternative strategies to consider. > > > > > > > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-54+-+Sticky+Partition+Assignment+Strategy > > > > > > Please let me know if I missed to add something. Thank you. > > > > > > Regards, > > > --Vahid > > > > > > > > > > > >
