Hey Viktor, Thanks for thinking about Tiered Storage. I'm not so familiar there, so if you could add some of your expectations about how the two features will interact, I would appreciate that.
It appears to me that follower-fetch-from-remote is a significant optimization within TS, and so similar optimizations to support cross-cluster-replicate-from-remote and out-of-band remote replication could also be desirable. I think we can explore the idea further, and make sure that CCR is extensible to tiered topics if it doesn't make it into the initial implementation. Thanks! Greg On Wed, Oct 4, 2023 at 6:13 AM Viktor Somogyi-Vass <viktor.somo...@cloudera.com.invalid> wrote: > > Hi Greg, > > Thanks for the answers. I think they all make sense. > > Another point I realized last evening is that now that tiered storage (TS) > is available, it might complicate things with CCR. What I'm thinking of is > that if you have multiple clusters in multiple regions, enabling the object > storage's replication between zones could be much more cost efficient than > replicating local+remote offsets through Kafka. You'd only need to copy > local segments over and remote partition replication would be done by the > remote layer. Or the user could simply choose to not replicate remote > segments between regions but instead just reference them (so that the > backup cluster's remote offsets point to the original region). These > options however likely require bigger coordination between clusters than in > pre-TS Kafka. Do you think we should take this into consideration in the > design and in the UX? > > Thanks, > Viktor > > On Tue, Oct 3, 2023 at 6:30 PM Greg Harris <greg.har...@aiven.io.invalid> > wrote: > > > Hi Viktor, > > > > Thanks for your questions! I agree, replication is very fundamental in > > Kafka, so it's been implemented in many different ways by different > > people. I hope that this is the last implementation we'll need, but > > every software engineer says that :) > > > > GT-1: I think as this KIP is very focused on the UX of the feature, > > that user stories are appropriate to include. I think it isn't > > necessary to explain how the different applications are accomplished > > with MM2 or other solutions, but describing what they will look like > > after this KIP would be a wonderful addition. +1 > > > > MM2-1: I think that replacing the consumer is insufficient, as we need > > a more expressive producer as well. This is not possible within the > > design constraints of MM2 as a Connector, as MM2 uses the > > connect-managed producer. This could be implemented in MM3 as a new > > process that can use more expressive "internal clients", but then > > we've thrown away the Connect runtime that made MM2 easier to run for > > some users. > > MM2-2: This is technically possible, but sounds operationally hazardous to > > me. > > MM2-3: From the user perspective, I believe that CCR can be made more > > simple to use and operate than MM2, while providing better guarantees. > > From the implementation standpoint, I think that CCR will be > > significantly more complex, as the architecture of MM2 leverages a lot > > of the Connect infrastructure. > > > > LaK-1: Yes, I think you understand what I was going for. > > LaK-2: I don't think that this is a user experience that we could add > > to CCR without changing the Kafka clients to be aware of both clusters > > concurrently. In order to redirect clients away from a failed cluster > > with a metadata refresh, the cluster that they're currently connected > > to must give them that data. But because the cluster failed, that > > refresh will not be reliable. With a proxy between the client and > > Kafka, that proxy can be available while the original Kafka cluster is > > not. Failovers would happen between distinct sets of clients that are > > part of the same logical application. > > > > Thanks for taking a look at the rejected alternatives! > > Greg > > > > On Tue, Oct 3, 2023 at 3:24 AM Viktor Somogyi-Vass > > <viktor.somo...@cloudera.com.invalid> wrote: > > > > > > 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 > > > > > > > > > > > > > > > > > >
