We have been giving this a bunch of thought lately. We attempted to replace PARTITION_ASSIGNMENT_STRATEGY_CONFIG with our implementation that hooks into our deployment service. The idea is simple, the new deployment gets *Standby tasks assigned to them until they are caught up*. Once they are caught up, our deployment service takes the older deployment down and the new deployment takes over all the active tasks. We think it is possible to implement but there are large amount of cohesion between consumers and the stream setup that we are wading through.
We ended up writing: 1. AssignmentInfo 2. SubscriptionInfo 3. StickyStandbyTaskAssignor 4. PartitionAssigner All of which were largely copy and paste. Hoping we get to pick it back up soon and able to find a way to make this cleaner. Everything is coupled together in a pretty tight ball of goo today. On Mon, Jun 10, 2019 at 11:55 AM Parthasarathy, Mohan <mpart...@hpe.com> wrote: > Matt, > > I read your email again and this one that you point out: > > > What you also need to take into account is, how often topics are > > compacted, and how large the segment size is, because the active > segment > > is not subject to compaction. > > Are you saying that compaction affects the rebuilding time ? Sorry, I am > not sure I understand what you meant by this in the current context. > > -mohan > > > On 6/10/19, 10:22 AM, "Parthasarathy, Mohan" <mpart...@hpe.com> wrote: > > Thanks. That helps me understand why recreating state might take time. > > -mohan > > > On 6/9/19, 11:50 PM, "Matthias J. Sax" <matth...@confluent.io> wrote: > > By default, Kafka Streams does not "close" windows. > > To handle out-of-order data, windows are maintained until their > retention time passed, and are updated each time an out-of-order > record > arrives (even if window-end time passed). > > Cf > > https://urldefense.proofpoint.com/v2/url?u=https-3A__stackoverflow.com_questions_38935904_how-2Dto-2Dsend-2Dfinal-2Dkafka-2Dstreams-2Daggregation-2Dresult-2Dof-2Da-2Dtime-2Dwindowed-2Dktable&d=DwIGaQ&c=x_Y1Lz9GyeGp2OvBCa_eow&r=ChXZJWKniTslJvQGptpIW7qAh4kkrpgYSer_wfh4G5w&m=NPSMv1cPOmhxttSznyQJ_gbzxa1AHCJGXyS53vGbCFo&s=YUXTr0Z0Neki2SOQESJ6oLUelyCtlfoHyw-CMvWJKkI&e= > > for details; note, that using `suppress()` and setting a > grace-period > change the default behavior. > > Hence, you usually want to keep windows around until you are sure > that > no more out-of-order records for a window may arrive (what is > usually > some time after the window end time). If 1 day is too large, you > can of > course reduce the retention time accordingly. > > If you use Interactive Queries, you would need to set retention > time > large enough to allow your application to query the state. For this > case, retention time might be higher to keep state around for > serving > queries, even if you know it won't be updated any longer. > > For Kafka Streams, using 1 day as default provides a good > out-of-the-box > experience. For production deployments, change the retention time > base > on the need of the application make sense of course. > > There is also one more config you might want to consider: > `windowstore.changelog.additional.retention.ms` -- it's 1 day by > default, too, and you might want to reduce it to reduce the amount > of > data that is restored. > > > In general, there in nothing wrong with using stateful sets > though, and > for large state, it's recommended to avoid long recovery times. > > > > -Matthias > > > On 6/9/19 2:59 PM, Parthasarathy, Mohan wrote: > > Matt, > > > > Thanks for your response. I agree with you that there is no easy > way to answer this. I was trying to see what others experience is which > could simply be "Don't bother, in practice stateful set is better". > > > > Could you explain as to why there has to be more state than the > window size ? In a running application, as the data is being processed from > a topic, there is state being created depending on the stateful primitives. > As the window is closed, this state is not needed and I can see why grace > period has to be taken into account ? So, when would you need it for the > whole store retention time ? Could you clarify ? > > > > Thanks > > Mohan > > > > On 6/8/19, 11:18 PM, "Matthias J. Sax" <matth...@confluent.io> > wrote: > > > > If depends how much state you need to restore and how much > restore-time > > you can accept in your application. > > > > The amount of data that needs to be restored, does not > depend on the > > window-size, but the store retention time (default 1 day, > configurable > > via `Materialized#withRetention()`). The window size (and > grace period, > > if case you use one) is a lower bound for the configurable > retention > > time though, ie, retention time >= grace-period >= > window-size. > > > > What you also need to take into account is, how often topics > are > > compacted, and how large the segment size is, because the > active segment > > is not subject to compaction. > > > > It's always hard to answer a question like this. I would > recommend to do > > some testing and benchmark fail over, by manually killing > some instances > > to simulate a crash. This should give the best insight -- > tuning the > > above parameters, you can see, what works for your > application. > > > > > > > > -Matthias > > > > On 6/8/19 4:28 PM, Pavel Sapozhnikov wrote: > > > I suggest take a look at Strimzi project > https://urldefense.proofpoint.com/v2/url?u=https-3A__strimzi.io_&d=DwIGaQ&c=x_Y1Lz9GyeGp2OvBCa_eow&r=ChXZJWKniTslJvQGptpIW7qAh4kkrpgYSer_wfh4G5w&m=NPSMv1cPOmhxttSznyQJ_gbzxa1AHCJGXyS53vGbCFo&s=qu-Z9wpqbvWtwzM1ED2N1U5Rvj2-N_y6iQgGMFGcAT4&e= > > > > > > > Kafka operator deployed in Kubernetes environment. > > > > > > On Sat, Jun 8, 2019, 6:09 PM Parthasarathy, Mohan < > mpart...@hpe.com> wrote: > > > > > >> Hi, > > >> > > >> I have read several articles about this topic. We are > soon going to deploy > > >> our streaming apps inside k8s. My understanding from > reading these articles > > >> is that stateful set in k8s is not mandatory as the > application can rebuild > > >> its state if the state store is not present. Can people > share their > > >> experience or recommendation when it comes to deploying > the streaming apps > > >> on k8s ? > > >> > > >> Also, let us say the application is using a tumbling > window of 5 mts. When > > >> an application restarts, is it correct to say that it has > to re-build the > > >> state only for that 5 minute window for the partitions > that it was handling > > >> before. I had an instance of such a restart where it was > running a long > > >> time in REBALANCE which makes me think that my > understanding is incorrect. > > >> In this case, the state store was available during the > restart. Can someone > > >> clarify ? > > >> > > >> Thanks > > >> Mohan > > >> > > >> > > > > > > > > > > > > > > >
