Hi Greg, Seems like finding the perfect replication solution is a never ending story for Kafka :).
Some general thoughts: GT-1. While as you say it would be good to have some kind of built-in replication in Kafka, we definitely need to understand the problem better to provide a better solution. Replication has lots of user stories as you iterated over a few and I think it's very well worth the time to detail each one in the KIP. This may help understanding the problem on a deeper level to others who may want to contribute, somewhat sets the scope and describes the problem in a way that a good solution can be deduced from it. I also have a few questions regarding some of the rejected solutions: MM2: I think your points about MM2 are fair (offset transparency and operational complexity), however I think it needs more reasoning about why are we moving in a different direction? A few points I can think about what we could improve in MM2 that'd transform it into more like a solution that you aim for: MM2-1. What if we consider replacing the client based mechanism with a follower fetch protocol? MM2-2. Operating an MM2 cluster might be familiar to those who operate Connect anyway. For those who don't, can we provide a "built-in" version that runs in the same process as Kafka, like an embedded dedicated MM2 cluster? MM2-3. Will we actually be able to achieve less complexity with a built-in solution? Layer above Kafka: LaK-1. Would you please add more details about this? What I can currently think of is that this "layer above Kafka" would be some kind of a proxy which would proactively send an incoming request to multiple clusters like "broadcast" it. Is that a correct assumption? LaK-2. In case of a cluster failover a client needs to change bootstrap servers to a different cluster. A layer above Kafka or a proxy can solve this by abstracting away the cluster itself. It could force out a metadata refresh and from that point on clients can fetch from the other cluster. Is this problem within the scope of this KIP or not? Thanks, Viktor On Tue, Oct 3, 2023 at 2:55 AM Greg Harris <greg.har...@aiven.io.invalid> wrote: > Hey Tom, > > Thanks for the high-level questions, as I am certainly approaching > this KIP differently than I've seen before. > > I think that ideally this KIP will expand to include lots of > requirements and possible implementations, and that through discussion > we can narrow the scope and form a roadmap for implementation across > multiple KIPs. I don't plan to be the decision-maker for this project, > as I'm more interested in building consensus among the co-authors. I > can certainly poll that consensus and update the KIP to keep the > project moving, and any other co-author can do the same. And to set an > example, I'll clarify your questions and for anything that I agree > with, I'll ask that you make the update to the KIP, so that the KIP > captures your understanding of the problem and your requirements. If > you don't get the chance to make the changes yourself, I'll make sure > they get included eventually, as they're very good ideas :) > > For your remaining questions: > > M1: I was trying to draw analogies to databases, but your suggested > properties are much more compelling and informative. I'd love it if > you added some formalism here, so that we have a better grasp on what > we're trying to accomplish. +1 > M2: I think the "asynchronous" problem corresponds to the goal of > "exactly once semantics" but the two are not obviously opposites. I > think the MM2 deficiencies could focus less on the architecture > (asynchronicity) and more on the user-facing effect (semantics). +1 > M3: I had a "non-goals" section that ended up becoming the "rejected > alternatives" section instead. If you have some non-goals in mind, > please add them. > M4+M5: I think it's too early to nail down the assumptions directly, > but if you believe that "separate operators of source and target" is a > requirement, that would be good to write down. +1 > M6: That is a concerning edge case, and I don't know how to handle it. > I was imagining that there would be a many:many relationship of > clusters and links, but I understand that the book-keeping of that > decision may be significant. > M7: I think this may be appropriate to cover in a "user story" or > "example usages". I naturally thought that the feature would describe > some minimal way of linking two topics, and the applications > (combining multiple links, performing failovers, or running > active-active, etc) would be left to users to define. I included the > regex configurations because I imagine that creating 100s or 1000s of > links would be unnecessarily tedious. The feature may also encode > those use-cases directly as first-class citizens as well. > > U1: These are states that can happen in reality, and I meant for that > section to imply that we should expect these states and model them for > operations and observability. > > D1: I think I may have introduced this confusion by trying to be > terse. I imagined that there will be two different topics on the > source and target, which would be synced to have the same > configuration contents, similar to MM2's implementation. This would > allow for the replication link to be permanently disconnected and the > target topic to become just a regular topic, Later, a new replication > link and new target topic (with another separate topic-id) can be > created to rebuild the replication. I also thought that it was > possible that two clusters had already chosen the same topic-id, and > that attempting to interpret one topic-id in two different clusters > was error-prone. As far as replicating __cluster_metadata: I hadn't > considered that, but that might be required depending on the semantics > we choose. > D2: Thanks, that's a good clarification. Uptime and bandwidth should > be assumed to be lower, and latency should be assumed to be higher. +1 > D3: I included this restriction because it would not be transparent to > source consumers. They would need special support for connecting to > brokers from multiple clusters, with potentially distinct metadata. > > Thanks so much! > Greg > > On Mon, Oct 2, 2023 at 4:24 PM Tom Bentley <tbent...@redhat.com> wrote: > > > > Hi Greg, > > > > Thanks for this KIP! It is obviously very ambitious, but it's great to > have > > a conversation about it. > > > > I'll start with some general points: > > > > Do you have a plan in mind for how to proceed with elaborating this KIP? > > While I like how you're involving the community in elaborating the KIP, I > > think there is a danger, which is more likely with this inclusive > approach, > > in trying to attempt too much at once. > > > > In my opinion someone needs to take the difficult decisions necessary to > > limit the initial scope (and, just as importantly, communicate that > > clearly) in order to maximise the chances of actually getting something > > accepted and implemented. Can we assume that you're that person? Defining > > the what and how of the metadata replication, and the log replication > seem > > to me to be the core of what you're trying to achieve here. We should > make > > anything that is not crucial to that (i.e. NAT punching) a non-goal of > this > > KIP. Future KIPs can easily add those features. > > > > I also had a few specific points: > > > > Motivation > > M1. I don't find the "logical replication" vs "physical replication" > > particularly helpful. I think one key property is "offset preserving", > > which is also self-explanatory. Slightly more generally, we could define > > the concept of "consumer transparency", i.e. a consumer could reconnect > to > > either cluster and observe the same sequences of records (same order, > same > > offsets, and same transaction visibility). Consumer transparency requires > > synchronous replication, but offset preserving does not. > > M2. In the motivation you mention that MM offers asynchronous > replication, > > but the Goals subsection doesn't mention support for synchronous > > replication. We should be clear which (or both) we're aiming for. > > M3. A Non-Goals section would be useful, especially for a KIP that's > large > > and ambitious like this one. > > M4. It might also be worth having a list of Assumptions. Here we could > list > > all the things we want to assume in order to make the initial KIP > feasible. > > M5. For example we should be explicit about whether or not it is assumed > > that the same people are operating (and thus have visibility into) both > > clusters. > > M6. One thing worth calling out is whether the clusters themselves are > in a > > leader/follower relationship (e.g. the DR scenario), or whether this is a > > topic-level concern. I guess it's topic level from the topic and consumer > > group regexes. But this has consequences we should explore. For example > > what if a transaction includes records in topics X and Y, where X is > > replicated but Y is not? > > M7. I think you should be clear about whether this leader/follower > > relationship can be reversed, and in what circumstances. In the user > > interface section you talk about "disconnected", but not this kind of > > fail-back. > > > > > > User interface > > U1. "Links can be temporarily or permanently disconnected." Are you > > describing a fact about the network between the two clusters, or is this > > disconnection something actively managed by the system, or by the > operator? > > > > Data semantics > > D1. The KIP says "both cross-cluster topics and intra-cluster replicas: > > Have the same configuration as their source" but you also say > > "cross-cluster replicas: Have a separate topic-id", this seems like a > > contradiction, on the face of it. It seems like there's a whole host of > > devils in the detail behind this. It implies replication of (some of) the > > __cluster_metadata, I think, but not all (since you say ACLs are not > > replicated). If that's right, then what does it imply about referential > > integrity between metadata records? i.e. what if metadata record A (which > > is replicated) references record B (which is not)? Even if this is not > > possible by design initially, how does it constrain the future evolution > of > > metadata record schemas? Is any such metadata replication going to be > > transaction preserving? If the topic ids can differ then what is > > responsible for the mapping and rewriting of metadata records which > include > > topic ids? > > D2. "The network path between Kafka clusters is assumed to be less > reliable > > than the intra-cluster network," we should be explicit about whether or > not > > we're assuming similar network latencies and bandwidth for the > > inter-cluster network links as for the in-cluster ones. > > D3 "Are not eligible for fetch-from-follower on the source cluster" the > > reason for this isn't immediately apparent to me. > > > > Thanks again, > > > > Tom > > > > On Tue, 3 Oct 2023 at 09:37, Greg Harris <greg.har...@aiven.io.invalid> > > wrote: > > > > > Hi all, > > > > > > I've opened an extremely early draft of the Cross-Cluster Replication > > > feature, and I would like to invite any and all co-authors to expand > > > on it. Find the page here: > > > > > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-986%3A+Cross-Cluster+Replication > > > > > > This is not strictly an invitation to "review" the KIP, as the > > > document has much less detail than other KIPs of similar complexity. > > > But if you are knowledgeable in this area, some early sanity checks > > > would be greatly appreciated. > > > > > > I've included a "shopping list" of properties that appear to me to be > > > desirable, but I don't have an implementation in mind that meets these > > > requirements. If you have additional requirements, an alternative UX > > > in mind, or wish to propose some implementation details, please edit > > > the KIP with your contributions. > > > > > > Thanks everyone! > > > > > > Greg Harris > > > Aiven, Inc > > > > > > >
