Thanks a lot for the detailed explanation here Konstantine! I strongly agree that a rolling start of Kafka broker is not the optimal solution when we have an alternative to just upgrade the client. Also I fully understood your explanation on task shuffle minimum impact in the workers scenario, because the local storage usage is very limited.
Focusing on the current KIP, a few more suggestions are: 1. I copy-pasted partial scenario on the Leader bounces section d' is the remaining delay W1, which is the leader, leaves Rebalance is triggered W2 joins with assignment: [] W3 joins with assignment: [BT1] W3 becomes leader. There's an active delay in progress. W3 computes and sends assignments: W2(delay: d'', assigned: [], revoked: []) W3(delay: d'', assigned: [BT1, AT1], revoked: []) after we start d' round of delayed rebalance. Why does W3 send assignment [BT1, AT1] instead of just [BT1] here? I guess we won't do the actual rebalance until the original scheduled delay d is reached right? 2. we are basically relying on the leader subscription to persist the group assignment across the generation and leader rejoin to trigger necessary rebalance. This assumption could potentially be broken with future upgrades of broker as we are discussing https://issues.apache.org/jira/browse/KAFKA-7728. This JIRA will be converted to a ready KIP by Mayuresh pretty soon, and our goal here is to avoid unnecessary rebalance due to leader bounces by specifying a field called JoinReason for broker to interpret. With that change in mind, I think it's worth mentioning this potential dependency within KIP-415 so that we don't forget to have corresponding change to adapt to 7728 broker upgrade in case JoinReason change happens before KIP-415. Am I clear on the situation explanation? 3. cooperative cmeans -> means that only Incremental Cooperative Connect protocol is enabled (version 1 or higher). 4. For the compatibility change, I'm wondering whether we could just use 2 connect protocols instead of 3. Because the user knows when all the workers all upgraded to version 1, we could just use `compatible` for the first rolling bounce and 'cooperative' for the second bounce. Could you explain a bit why we need to start from `eager` stage? cc Mayuresh on this thread. Thanks, Boyang ________________________________ From: Konstantine Karantasis <konstant...@confluent.io> Sent: Friday, January 18, 2019 8:32 AM To: dev@kafka.apache.org Subject: Re: [DISCUSS] KIP-415: Incremental Cooperative Rebalancing in Kafka Connect Hi Stanislav and Boyang. Thanks for your comments. You are both asking how KIP-345 affects this KIP, so, first, I'll address this point. KIP-345 has recently introduced an option that will allow Kafka consumer applications to avoid rebalances due to the departure and subsequently return of a member in the group. That way KIP-345 offers a remedy for the cases of rolling bounces and replacement of nodes due to failures that can happen quickly. Without ruling out that policies of Incremental Cooperative Rebalancing may use static membership eventually in order to better address such use cases, next I'll mention a few reasons why I believe KIP-415, which is the proposal of Incremental Cooperative Rebalancing in Connect, should proceed independently at first: * KIP-345 requires an upgrade to both Kafka brokers and Connect workers. This requirement is very strict for a big group of Connect users that anticipate a solution to the stop-the-world effect in Connect in order to grow their Connect clusters, but can not afford to also have to upgrade their Kafka brokers in order to enjoy the suggested improvements. * Connect clusters are traditionally locally stateless and lightweight, in the sense that they don't store state outside Kafka and that this state is easy to process during startup. Overall, based on their overall configuration and deployment requirements, Connect Workers very suitable to run in containers. With respect to the resources that Connect Workers are rebalancing, connectors and tasks are (and honestly should be) easy to spin and redistribute in a Connect cluster. This is true because connectors depend on Kafka and the systems they connect in order to save their progress. They don't use the Workers' local instances. Given this reality, the configuration of a unique id, per KIP-345's requirement, doesn't seem necessary for Connect to introduce yet. The upgrade path is made even easier without having to define unique ids and in practice the heuristics of Incremental Cooperative Rebalancing have the potential to cover (immediately or eventually) most of the scenarios that make rebalancing and stop-the-world problematic in Connect today. * Static membership has not been merged yet. Given that KIP-415 addresses also scale-up and scale-down scenarios and the important side-effect that the submission of a new connector has to other connectors in the worker's group, it seems to me that introducing an interdependency between the two proposals is not necessary. Again, this doesn't prevent reconsidering integration in the future. * Finally, it's not immediately obvious to me that integration between the two proposals also means significantly simpler implementation in Connect. That's because Connect Workers will have to handle a delay one way or the other. Plus, the group management and resource assignment code is mostly separate between Connect and the Consumer. With respect to your other comments, Stanislav, glad you found the examples easy to read. I'll change the KIP to show who's leader at the beginning as well. Boyang, I'll add a paragraph to highlight why local state is not the most pressing issue in Connect. Thank you both for your initial comments. Konstantine On Mon, Jan 14, 2019 at 9:24 AM Boyang Chen <bche...@outlook.com> wrote: > Hey Konstantine, > > great work for making this happen! Incremental rebalancing is super > important for avoiding unnecessary resource shuffle and improving the > overall Connect framework stability. > > After the first pass, two questions across my mind are: > > 1. For my understanding, the general rebalancing case could be covered > by configuring the workers as static members, so that we don't need to > worry about worker temporarily leaving group case. Basically KIP-345 could > help with avoiding unexpected rebalances during cluster rolling bounce > which I feel the same way as Stanislav that parts of 415 logic could be > simplified. It would be great if we could look at these two initiatives > holistically to help reduce the common workload. > 2. Since I never used Connect before, I do hope you could enlighten me > on the potential effort involved in task transfer between workers. The > reason I ask is to estimate how much burden will we introduce by starting a > task on the brand new worker? Is there any local state to be replayed? It > would be good to also provide this background in the KIP motivation so that > people could understand better of the symptom and build constructive > feedbacks. > > Thanks a lot! > > Boyang > ________________________________ > From: Stanislav Kozlovski <stanis...@confluent.io> > Sent: Monday, January 14, 2019 3:15 PM > To: dev@kafka.apache.org > Subject: Re: [DISCUSS] KIP-415: Incremental Cooperative Rebalancing in > Kafka Connect > > Hey Konstantine, > > This is a very exciting and fundamental-improving KIP, thanks a lot for > working on it! > > Have you seen KIP-345 > <https://cwiki.apache.org/confluence/display/KAFKA/KIP-345>? I was > wondering whether Connect would support the static group membership - > potentially limiting the need to handle "node bounce" cases through a > rebalance (even though there wouldn't be downtime). I find it is somewhat > related to the `scheduled.rebalance.max.delay.ms` config described in > KIP-415. The main difference I think is that rebalance delay in KIP-345 is > configurable through `session.timeout.ms` which is tied to the liveness > heartbeat, whereas here we have a separate config. > > The original design document suggested > > Assignment response includes usual assignment information. Start > processing any new partitions. (Since we expect sticky assignment, we could > also optimize this and omit the assignment when it is just repeating a > previous assignment) > Have we decided on whether we would make use of the optimization as to not > send the assignment that the worker already knows about? > > I enjoyed reading the rebalancing examples. As a small readability > improvement, could I suggest we clarify which Worker (W1,W2,W3) is the > leader in the "Initial group and assignment" part? For example, in the > `Leader bounces` I was left thinking whether the leaving W2 was the initial > leader or not. > > Thanks, > Stanislav > > On Sat, Jan 12, 2019 at 1:44 AM Konstantine Karantasis < > konstant...@confluent.io> wrote: > > > Hi all, > > > > I just published KIP-415: Incremental Cooperative Rebalancing in Kafka > > Connect > > on the wiki here: > > > > > https://cwiki.apache.org/confluence/display/KAFKA/Kafka+Improvement+Proposals > > > > This is the first KIP to suggest an implementation of incremental and > > cooperative rebalancing in the context of Kafka Connect. It aims to > provide > > an adequate solution to the stop-the-world effect that occurs in a > Connect > > cluster whenever a new connector configuration is submitted or a Connect > > Worker is added or removed from the cluster. > > > > Looking forward to your insightful feedback! > > > > Regards, > > Konstantine > > > > > -- > Best, > Stanislav >
