Hi Jun, Thanks for the review comments. Answer them below.
> JR1. "Epoch rewriting: When records are appended to the destination log, the batch epochs are rewritten to match the destination cluster's leader epochs, maintaining consistency within the destination cluster." This has a couple of impacts. We have an updated design to support unclean leader election in this doc, where we will NOT rewrite the leader epoch anymore. https://cwiki.apache.org/confluence/display/KAFKA/Unclean+Leader+Election+in+Cluster+Mirroring > JR2. "Tiered Storage is not initially supported": Ideally, we should support tiered storage. Same as RS9, the destination cluster issues consumer requests, which support tiering. The OffsetMovedToTieredStorageException is used by replication inside AbstractFetcherThread. This suggests that it's probably not a good fit for cluster mirroring to use AbstractFetcherThread. Yes, thanks for pointing this out. I was wrong about this. Before we support tiered storage, the mirroring will mirror all records including data in the remote storage into the destination cluster. > JR3. For each new cluster mirroring command, it would be useful to document the underlying workflow (e.g, which RPC requests are issued, to which node; what records are written to metadata topic, or internal topic, which actions are triggered on the broker, etc). Will do. > JR4. Truncating a log to LMO. Currently, there is no internal API that truncates a partition from the end. Could you describe how this will be implemented to ensure all replicas are consistent after the truncation? The truncation flow is like this: 1. When the MirrorMetadataManager in the node gets notified about the partition leader assignment when onMetadataUpdate (via TopicsDelta), it'll query the mirror coordinator about mirror partition state. 2. When it's the PREPARING state, the MirrorMetadataManager in the leader node will get the last mirrored offset (or epoch) from the source cluster (new API) and then do the log truncate. 3. In (2), we'll also register a callback in Partition instance, and wait until all ISRs complete the truncation by checking the follower replica's LEO. 4. In (3), this check will be invoked every time the leader node update follower fetch state, like how we check if high watermark should be incremented. 5. After all ISRs complete the truncation, we'll invoke the callback and move the mirror partition state to MIRRORING, and then start fetching data from the source cluster. Note: (1) In PREPARING state, the partition is READ-ONLY, so there will no any data written in the leader node (2) During step (1) ~ (4), if any leadership change happens, the new leader will start from step (1) to complete the log truncation process. (3) If unclean leader election is supported (i.e. mirror.support.unclean.leader.election=true), then we'll wait until ALL registered replicas complete the truncation before moving on to MIRRORING state. We'll update the KIP in the following days to address the community feedback. Some questions need more thinking. Please give us some time. :) Thank you, Luke On Fri, Mar 20, 2026 at 9:10 AM Jun Rao via dev <[email protected]> wrote: > Hi, Federico, > > Thanks for the KIP. A few comments. > > JR1. "Epoch rewriting: When records are appended to the destination log, > the batch epochs are rewritten to match the destination cluster's leader > epochs, maintaining consistency within the destination cluster." This has a > couple of impacts. > JR1.1 How do we ensure that the leader epoch in committed offsets is > consistent with the leader epoch in the batch? This consistency is > important when the consumer fails over to a different cluster. It seems the > KIP doesn't translate the leader epoch when mirroring the comitted offsets. > JR1.2 Typically, leader epochs increase monotonically in the log. Do we > ensure this remains the case after failover and failback? > > JR2. "Tiered Storage is not initially supported": Ideally, we should > support tiered storage. Same as RS9, the destination cluster issues > consumer requests, which support tiering. > The OffsetMovedToTieredStorageException is used by replication > inside AbstractFetcherThread. This suggests that it's probably not a good > fit for cluster mirroring to use AbstractFetcherThread. > > JR3. For each new cluster mirroring command, it would be useful to document > the underlying workflow (e.g, which RPC requests are issued, to which node; > what records are written to metadata topic, or internal topic, which > actions are triggered on the broker, etc). > > JR4. Truncating a log to LMO. Currently, there is no internal API that > truncates a partition from the end. Could you describe how this will be > implemented to ensure all replicas are consistent after the truncation? > > Jun > > On Mon, Mar 16, 2026 at 2:44 AM Federico Valeri <[email protected]> > wrote: > > > I mentioned a corner case in the chained mirroring use case. Let me > > clarify what I mean with a simple example: > > > > 1. B is fetching from A, and C is fetching from B (A --> B --> C) > > 2. A producer with PID 5 sends records to A > > 3. Failover happens and B becomes writable (A -x-> B --> C) > > 4. A different producer with PID 5 sends records to B > > 5. Collision on cluster C (two different producers mapped to PID -7 in C) > > > > (arrows represent data flowing, not fetch direction) > > > > > > On Sun, Mar 15, 2026 at 7:14 PM Federico Valeri <[email protected]> > > wrote: > > > > > > Hi Rajini, thanks for your thoughtful review and for catching a few > > > bugs. I'll skip some questions that we will address later. > > > > > > RS2: The metadata records are described in "Mirror Metadata Records" > > > paragraph. Currently there are only two records: "LastMirroredOffsets" > > > record tracks the latest successfully mirrored offset for each > > > partition, while "MirrorPartitionState" record represents the > > > lifecycle states of a mirrored partition. > > > > > > RS3: That's a good point that was also raised by Andrew. It was an > > > easy solution that we used for our prototype, but we need to think > > > about a better solution. > > > > > > RS4: Current design is that mirror fetcher threads behaves like a > > > read_committed consumer fetching up to "source LSO". On failover we > > > truncate destination log to "local LSO". > > > > > > The approach of fetching up to HW that you propose is still safe as we > > > keep truncating to local LSO on failover, but it trades lower > > > steady-state lag (especially when long-running transactions exist on > > > the source) for more data loss on failover (the net data loss relative > > > to the source is the same in both approaches). In other words, with > > > your approach we fetch more data that we may then need to truncate. > > > Also, read_uncommited consumers on the destination cluster would be > > > able to read records that may be truncated on failover. These are just > > > my consideration, but we are open to discussion on which is the best > > > approach here. > > > > > > When a failover is triggered (RemoveTopicsFromMirror), the sequence is: > > > > > > 1. Partitions transition to STOPPING state > > > 2. Fetchers are removed > > > 3. For each partition, truncate to local LSO is called > > > 3.1. Reads LSO from each partition's local log > > > 3.2. Calls log.truncateTo(offset) on the UnifiedLog > > > 3.3. Ensures ISR members complete truncation before the partition > > > becomes writable > > > 4. For each partition, the LSO is recorded as the last mirrored offset > > > (LMO) in __mirror_state > > > 5. Partitions transition to STOPPED and become writable > > > > > > When a failback is triggered (AddTopicsToMirror), the sequence is: > > > > > > 1. Partitions transition to PREPARING state > > > 2. For each partition, truncation to LMO is called > > > 2.1. This sends a LastMirroredOffsetsRequest to the source cluster > > > to fetch the offsets that were recorded during the previous failover > > > 2.2.a. The response offsets are used to truncate local logs > > > 2.2.b. If the source cluster doesn't support the LastMirroredOffsets > > > API or first-time mirror, it truncates to offset 0 > > > 3. Partitions transition to MIRRORING > > > > > > RS7: Can you point me to the section that says configs are stored in > > > __mirror_state? Mirror connection configs (bootstrap servers, > > > credentials, etc.) are stored in KRaft metadata via > > > ConfigResource.Type.MIRROR, not in __mirror_state. The internal topic > > > only stores partition states and last mirrored offsets. Sensitive > > > credentials follow the existing KRaft config handling, which is > > > already protected by controller/broker access controls and sensitive > > > config redaction in DescribeConfigs responses. > > > > > > RS8: Not sure what's the recommended approach here. Adding a new error > > > code does not change the response schema and older clients that don't > > > recognize the new error code will surface it as an > > > UnknownServerException (non-retriable). > > > > > > RS11: Good catch. This is a prototype simplification that we need to > > > address. To properly sync consumer group offsets, the implementation > > > would need to send ListGroups to all source brokers (or use the > > > AdminClient which does this internally), send FindCoordinator to > > > discover the group coordinator for each group, send OffsetFetch to the > > > correct group coordinator. > > > > > > RS12: You are absolutely right, the transformation is not idempotent, > > > so it is not safe for chained mirroring (A -> B -> C). Instead, > > > round-trip mirroring (A -> B, then B -> A) works because, when doing a > > > failback, the log is truncated before mirroring resumes, so previously > > > mirrored records with negative pids are removed and the transformation > > > is only applied to new records produced natively on that cluster > > > (double-transformation never occurs). Non-transactional batches stay > > > at -1 [ -(-1 + 2) = -(1) = -1], which is correct. > > > > > > The chained mirroring would work if we skip the transformation when > > > pid is negative, but there is still an edge case: A -> B -> C with > > > local B producer. If cluster A has local pid 5 and cluster B also has > > > local pid 5, both end up as -7 on cluster C. Collision: two different > > > producers with the same pid on the destination. No pid-only > > > transformation can solve that. We would need to incorporate cluster > > > identity. > > > > > > Possible solution that would handle any topology: The producer IDs are > > > 64-bit signed longs used to identify a producer. The clusterId (UUID) > > > is a globally 128-bit unique identifier for each source cluster. We > > > could use the clusterId hash to partition the entire negative PID > > > space into regions, one per source cluster. Basically we divide the 64 > > > bits into three fields: bit 63 (sign bit), bits 62-31 (region > > > selector), bits 30-0 (producer identity). Once a non-negative PID is > > > mapped to a region, it passes through unchanged no matter how many > > > hops follow (i.e. we apply the transformation only for PIDs >= 0). > > > > > > Example with two clusters: > > > > > > - Bit 63: This is the sign bit that makes the value negative and > > > distinguishes mirrored pids from local ones (which are non-negative). > > > - Bits 62-31 cluster A: clusterId = "abc-123", clusterHash = 42 > > > - Bits 62-31 cluster B: clusterId = "xyz-789", clusterHash = 99 > > > - Bits 30-0: Local producer ID 5 that is the same on both clusters. > > > > > > A's pid 5 --> > > > 1|00000000000000000000000000101010|0000000000000000000000000000101 > > > B's pid 5 --> > > > 1|00000000000000000000000001100011|0000000000000000000000000000101 > > > > > > On Wed, Mar 11, 2026 at 1:23 PM Rajini Sivaram < > [email protected]> > > wrote: > > > > > > > > A few more questions about the KIP for clarification: > > > > > > > > RS8: The KIP says produce requests to mirror topics will throw > > > > ReadOnlyTopicException. For Produce Requests returning a new error to > > > > clients, don’t we need to bump Produce request version? > > > > > > > > RS9: The KIP says we use OffsetMovedToTieredStorageException to > prevent > > > > mirroring of data in tiered storage. But doesn’t the mirror client > look > > > > like a regular consumer to the source cluster and return records > > fetched > > > > from tiered storage? > > > > > > > > RS10: Client-id based quotas for the source cluster look hard to > manage > > > > since there is no hierarchy or grouping possible. Seems better to > rely > > on > > > > secure user-principal based quotas on the source-side. > > > > > > > > RS11: The KIP says `The manager maintains a connection pool with one > > > > blocking sender per source cluster`. If this is the connection used > for > > > > periodic sync of offsets, topic configs etc. the coordinator is > likely > > to > > > > need connections to all source brokers (i.e. all group coordinators). > > > > > > > > RS12: The KIP proposes to transform producer ids for mirror records > to > > > > avoid conflicts. This comes at a cost because CRC checksum needs to > be > > > > recomputed. To justify this cost, we need to ensure that this > > > > transformation works in all cases. What happens if you are mirroring > a > > > > mirror topic? Is that a supported scenario? Or mirroring back > mirrored > > data > > > > during failback because the source was truncated? > > > > > > > > > > > > > > > > Thanks, > > > > > > > > Rajini > > > > > > > > > > > > On Tue, Mar 10, 2026 at 8:19 PM Rajini Sivaram < > > [email protected]> > > > > wrote: > > > > > > > > > Hi team, > > > > > > > > > > Thanks for the KIP! I have a few questions, mostly clarification at > > this > > > > > point. > > > > > > > > > > RS1: There is a `CreateMirror` request but no corresponding > > `DeleteMirror` > > > > > request. Is that intentional? > > > > > > > > > > > > > > > > > > > > RS2: It will be good to define the format of data going into the > > internal > > > > > mirror state topic. There is an example under kafka-dump-logs, > which > > > > > shows partition-level state in the payload and the mirror name as > > key. I > > > > > guess that is not what we expect it to be. Do we delete this > > information > > > > > when a topic is deleted or a mirror is deleted? > > > > > > > > > > > > > > > > > > > > RS3: KIP currently says mirror name cannot end with .removed. I > > guess it > > > > > cannot also end with .paused. Have we considered storing state and > > > > > mirror name separately, but updated together for a topic? Since new > > > > > states may be added in future, name restrictions may become hard to > > > > > implement. > > > > > > > > > > > > > > > > > > > > RS4: The KIP says *“mirroring must fetch only up to the LSO to > > maintain > > > > > transactional consistency”* and it also says *“During the mirror > > stopping > > > > > transition, the MirrorCoordinator performs a log truncation > > operation that > > > > > resets each mirror partition to its LSO.”* I guess the plan is to > > fetch > > > > > up to high watermark and truncate to locally computed LSO on > > failover? > > > > > Details of the sequence here will be useful. How does > > MirrorCoordinator > > > > > perform truncation? > > > > > > > > > > > > > > > > > > > > RS5: The KIP says “*On the destination cluster, mirror-related > > operations > > > > > (creating mirrors, adding/removing topics from mirrors, managing > > mirror > > > > > configurations) require the CLUSTER_ACTION permission on the > cluster > > > > > resource.*” The `Cluster:ClusterAction` ACL is currently used for > > broker > > > > > service account, e.g. local replication is authorized using this. > It > > seems > > > > > odd to grant this permission to users managing a resource on the > > cluster. > > > > > Have we considered adding a new resource type `ClusterMirror` and > > define > > > > > ACLs like `ClusterMirror:Create`, `ClusterMirror:Alter` and ` > > > > > ClusterMirror:AlterConfigs`? > > > > > > > > > > RS6: The KIP talks about three entities: Cluster Mirror, Mirror > > Topic and Mirror > > > > > Partition, with Cluster Mirroring as the feature name. Since we > > already > > > > > have MirrorMaker that also refers to mirrors, it will be nice if we > > can > > > > > refer to the entities using their full name in the CLI and public > > APIs. > > > > > That will enable us to add more mirror topic and mirror partition > > APIs in > > > > > the future if needed. For example: > > > > > > > > > > > > > > > - `kafka-cluster-mirrors.sh` to manage cluster mirrors > > > > > - createClusterMirrors(), listClusterMirrors(), > > > > > describeClusterMirrors() etc on the Admin API and Kafka > Protocol. > > > > > - KIP proposes pauseMirrorTopics(), resumeMirrorTopics() which > > are > > > > > good. > > > > > > > > > > > > > > > > > > > > RS7: The KIP proposes to store mirror configs in the internal > mirror > > state > > > > > topic. This includes sensitive credentials of another cluster. Have > > we > > > > > considered other options? Can a user with read access read the data > > from > > > > > the state topic using a consumer? > > > > > > > > > > > > > > > > > > > > Thanks, > > > > > > > > > > > > > > > Rajini > > > > > > > > > > > > > > > > > > > > On Sun, Mar 8, 2026 at 8:58 PM Andrew Schofield < > > [email protected]> > > > > > wrote: > > > > > > > > > >> Hi Fede and friends, > > > > >> I've re-read in detail and have quite a lot of comments, mostly > > minor > > > > >> clarifications, but as it approaches a vote, it's good to get the > > details > > > > >> nailed down. > > > > >> > > > > >> AS6: Could we have a diagram which shows which RPCs are served by > > which > > > > >> components? This will help illustrate the authorisation > > requirements for > > > > >> the various components, which is an aspect of the KIP that I don't > > think is > > > > >> completely specified yet. > > > > >> > > > > >> AS7: Please could you include a table of the operations and > > resources > > > > >> which will be checked for authorisation of each of the RPCs > > introduced. > > > > >> Also, please could you document the permissions which the > > destination > > > > >> cluster will require to mirror data and ACLs (for example, I think > > it will > > > > >> need ALTER on the CLUSTER resource to manipulate ACLs)? It's going > > to need > > > > >> Metadata, DescribeConfigs, DescribeAcls, ListGroups, OffsetFetch, > > > > >> LastMirrorOffset and Fetch RPCs I think, possibly others too. The > > user is > > > > >> probably going to want to give as little permission as possible to > > the > > > > >> destination cluster to get its job done. > > > > >> > > > > >> AS8: You include AuthorizedOperations in DescribeMirrorsResponse, > > but I > > > > >> don't know what the operations are. I think the implies MIRROR is > a > > new > > > > >> resource type in the Kafka security model and DescribeMirrors can > > be used > > > > >> to enquire the authorised operations for the client making the > > Admin API > > > > >> request. > > > > >> > > > > >> AS9: I think you're going to need some new error codes in the > Kafka > > > > >> protocol, as least: > > > > >> > > > > >> * INVALID_MIRROR_NAME or similar if the mirror name doesn't meet > the > > > > >> rules for a topic name > > > > >> * UNKNOWN_MIRROR if the mirror doesn't exist > > > > >> > > > > >> And probably some more for logical inconsistencies such as this > > topic > > > > >> isn't in that mirror, that topic is already in another mirror, and > > so on. > > > > >> > > > > >> AS10: Could you add the usage information for kafka-mirrors.sh > (the > > > > >> intended output from kafka-mirrors.sh --help) so all of the > options > > are > > > > >> documented together? For example, I see that --replication-factor > is > > > > >> included in one of the examples, which seems a bit surprising and > > I'm not > > > > >> sure whether it's a mistake or a feature. I can probably use > > --describe > > > > >> with a specific --mirror but it's not specified. > > > > >> > > > > >> AS11: I would expect the signature for Admin.addTopicsToMirror to > be > > > > >> Admin.addTopicsToMirror(String mirrorName, Set<String> topics, > > > > >> AddTopicsToMirrorOptions options) because it's for adding topics > to > > a > > > > >> mirror, as the counterpart to Admin.removeTopicsFromMirror(String > > > > >> mirrorName, Set<String> topics, RemoveTopicsFromMirrorOptions > > options). > > > > >> > > > > >> AS12: I don't think ignorable RPC fields in version 0 RPCs make > > sense > > > > >> because they're not trying to be compatible with a previous > version. > > > > >> > > > > >> AS13: I would have expected AddTopicsToMirrorRequest to have > mirror > > name > > > > >> above the list of topics because the same mirror name applies to > > all of the > > > > >> topics being added. As specified, you repeat the mirror name for > > all of the > > > > >> topics. > > > > >> > > > > >> AS14: I suggest adding ErrorMessage to the responses in all cases > > to make > > > > >> it easier to give more descriptive exception messages than just > the > > default > > > > >> for the error codes. > > > > >> > > > > >> AS15: I may have the wrong end of the stick here, but I expected > > > > >> RemoveTopicsFromMirrorRequest to remove the topics from a specific > > named > > > > >> mirror as implied by the example of the kafka-mirrors.sh command. > > In fact, > > > > >> I was expecting the mirror to contain the topics in the admin RPC > > requests > > > > >> and responses, and that's only true for about half of them. > > > > >> > > > > >> AS16: Can I change the mirror.name config using > > IncrementalAlterConfigs? > > > > >> If I attempt it, what's the error? > > > > >> > > > > >> AS17: If I attempt mirror RPCs when the mirror is in the wrong > > state, the > > > > >> error is specified as INVALID_REQUEST. That's usually kept for > > badly formed > > > > >> requests, as opposed to logically invalid ones. Maybe > > MIRROR_NOT_STOPPED or > > > > >> MIRRORING_ACTIVE or similar would be more expressive. > > > > >> > > > > >> AS18: Should the LastMirroredOffsetsResponse, > > ReadMirrorStatesResponse > > > > >> and WriteMirrorStatesRequest include LeaderEpoch? I suspect so. > > > > >> > > > > >> AS19: In DescribeMirrorsResponse, I suspect you will want "null" > > values > > > > >> for some fields which don't have values during initialisation and > > so on, > > > > >> such as lag. > > > > >> > > > > >> AS20: Do you need to add new versions of the DescribeConfigs and > > > > >> IncrementalAlterConfigs RPCs to support mirror resources? > > > > >> > > > > >> AS21: The topic configuration > mirror.replication.throttled.replicas > > is > > > > >> described as a list, but the default is MAX_LONG. > > > > >> > > > > >> AS22: By including mirror.name as a topic config, a client which > > has > > > > >> permission to describe configs for the topic is able to discover > > the name > > > > >> of the mirror, whether they are permitted to list the mirrors or > > describe > > > > >> that particular mirror. Generally, the Kafka authorisation model > > does not > > > > >> permit this kind of unauthorised information disclosure. For > > example, when > > > > >> a client describes the committed offsets for a consumer group, the > > list of > > > > >> topics returned is filtered to only those topics which the client > is > > > > >> permitted to describe, even though that may results in an > > incomplete set of > > > > >> topic partitions being returned. Is there an alternative way in > > which this > > > > >> information could be stored so Kafka only reveals mirror > > information to > > > > >> principals authorised to see it? > > > > >> > > > > >> AS23: I observe that there are situations in which a `.removed` > > suffix is > > > > >> added to the mirror name. Is it permitted for the user to define a > > mirror > > > > >> called "my.nasty.mirror.removed" and does it break anything? > > > > >> > > > > >> > > > > >> Thanks, > > > > >> Andrew > > > > >> > > > > >> On 2026/03/06 13:41:52 Paolo Patierno wrote: > > > > >> > Hi Fede, > > > > >> > something more ... > > > > >> > > > > > >> > Is there any migration path for users who want to migrate from > > using > > > > >> Mirror > > > > >> > Maker 2 to the cluster mirroring? > > > > >> > I mean, something like a tool useful to create a corresponding > > cluster > > > > >> > mirroring configuration starting from a MM2 one. Nothing that > > runs the > > > > >> > migration automatically but something that can be provided to > the > > users > > > > >> as > > > > >> > output to be validated and put in place by them. > > > > >> > > > > > >> > The Admin Client is missing methods to pause and stop mirroring > > (but we > > > > >> > have corresponding protocol messages). Is it on purpose? Any > > specific > > > > >> > reasons? They would be important from an automatic operator > > perspective > > > > >> use > > > > >> > case. > > > > >> > Also a method to provide the LastMirroredOffset from the source > > cluster > > > > >> > could be useful for progress and tracking purposes. > > > > >> > Finally, what about a method to get the mirror states? I don't > > think the > > > > >> > describe method provides such information. > > > > >> > In general, I think that the Admin Client section needs to cover > > in more > > > > >> > details the new classes definition like CreateMirrorOptions, > > > > >> > CreateMirrorResult, ... and so on for all the defined new > methods. > > > > >> > > > > > >> > > AddTopicsToMirrorResult addTopicsToMirror(Map<String, String> > > > > >> > topicToMirrorName, AddTopicsToMirrorOptions options); > > > > >> > > > > > >> > Isn't it missing the mirrorName (as you have in the > > > > >> removeTopicsFromMirror > > > > >> > counterpart)? > > > > >> > What's the topicToMirrorName parameter if it's defined as a Map? > > The > > > > >> method > > > > >> > is also plural using "topics" so comparing to the > > removeTopicsFromMirror > > > > >> > method, I would assume the parameter really is Set<String> > topics? > > > > >> > Comparing to the corresponding protocol message > > > > >> AddTopicsToMirrorRequest, I > > > > >> > see a list of topics but each of them has id, name and > > corresponding > > > > >> > mirror. So it's unclear how the addTopicsToMirror is defined. > > > > >> > > > > > >> > > RemoveTopicsFromMirrorResult removeTopicsFromMirror(String > > mirrorName, > > > > >> > Set<String> topics, RemoveTopicsFromMirrorOptions options); > > > > >> > > > > > >> > This method gets a mirrorName but if I look at the corresponding > > > > >> protocol > > > > >> > message RemoveTopicsFromMirrorRequest, it says "Allows users to > > detach > > > > >> > topics from their associated mirror" so the mirror is actually > not > > > > >> provided > > > > >> > and it's exactly what I see in the JSON definition (only topics > > list > > > > >> with > > > > >> > id and name). > > > > >> > > > > > >> > Finally, regarding the protocol change: > > > > >> > > > > > >> > * ListMirrorsResponse I would add the clusterId in the JSON > > definition > > > > >> > (it's related to my comments in the previous email when using > the > > tool). > > > > >> > * WriteMirrorStatesRequest has the following in the JSON which > > should > > > > >> not > > > > >> > be part of it "{ "name": "RemovedTopics", "type": "[]string", > > > > >> "versions": > > > > >> > "0+", "about": "The topic names to be removed." }" > > > > >> > > > > > >> > Thanks, > > > > >> > Paolo. > > > > >> > > > > > >> > On Fri, 6 Mar 2026 at 13:08, Paolo Patierno < > > [email protected]> > > > > >> > wrote: > > > > >> > > > > > >> > > Hi Fede, > > > > >> > > thank you for the proposal. I had a first pass with following > > > > >> thoughts and > > > > >> > > questions. > > > > >> > > > > > > >> > > > When the unclean.leader.election.enable is set to true, the > > broker > > > > >> will > > > > >> > > log a warning at every configuration synchronization period. > > > > >> > > Be more explicit about what the warning says. > > > > >> > > > > > > >> > > > This topic ID is not used by other topics in the current > > cluster > > > > >> > > In such a case, which should be very unlikely, what's going to > > happen? > > > > >> > > Isn't it possible to mirror the topic? > > > > >> > > > > > > >> > > > To enable it, all cluster nodes (controllers and brokers) > must > > > > >> > > explicitly enable unstable API versions and unstable feature > > versions > > > > >> in > > > > >> > > all configuration files. After starting the cluster with a > > minimum > > > > >> metadata > > > > >> > > version, operators can dynamically enable the mirror version > > feature > > > > >> to > > > > >> > > activate Cluster Mirroring. > > > > >> > > AFAIU there is going to be a dedicated feature flag for it, > > right? If > > > > >> yes > > > > >> > > can we state it clearly also specifying the exact name (i.e. > > > > >> mirror.version > > > > >> > > or something similar)? > > > > >> > > > > > > >> > > When running the kafka-mirrors.sh tool to list the mirrors, > > other than > > > > >> > > showing the SOURCE-BOOTSTRAP, it could be useful to have also > > the > > > > >> clusterId > > > > >> > > which, as a unique identifier, could be helpful in automated > > systems > > > > >> using > > > > >> > > the cluster mirroring. Of course, it would be important to > have > > in the > > > > >> > > ListMirrorsResponse as well as an additional field. > > > > >> > > > > > > >> > > What happens in case of Kafka downgrade from a version > > supporting > > > > >> > > mirroring to an older one not supporting it. > > > > >> > > The mirror won't be running but the topic configuration will > > still > > > > >> have > > > > >> > > config parameters like mirror.name and so on, right? Are they > > just > > > > >> > > ignored by the older Kafka version and the cluster will work > > without > > > > >> issues? > > > > >> > > > > > > >> > > Thanks, > > > > >> > > Paolo > > > > >> > > > > > > >> > > On Fri, 6 Mar 2026 at 10:43, Luke Chen <[email protected]> > > wrote: > > > > >> > > > > > > >> > >> Hi Andrew and all, > > > > >> > >> > > > > >> > >> About AS5, yes, I've created a sub-document > > > > >> > >> < > > > > >> > >> > > > > >> > > > https://cwiki.apache.org/confluence/display/KAFKA/Unclean+Leader+Election+in+Cluster+Mirroring > > > > >> > >> >to > > > > >> > >> explain the algorithm to support unclean leader election in > > cluster > > > > >> > >> mirroring. > > > > >> > >> Thanks for your comments, I'm inspired by that! :) > > > > >> > >> > > > > >> > >> About your idea, to store the owner of the leader epoch when > > > > >> leadership > > > > >> > >> change, I think it might not be needed because the most > > important > > > > >> thing > > > > >> > >> should be this: > > > > >> > >> > you might find that both ends have declared a local epoch > N, > > but > > > > >> someone > > > > >> > >> has to win. > > > > >> > >> > > > > >> > >> That is, as long as we have a way to declare who is the owner > > of > > > > >> leader > > > > >> > >> epoch N, then the 2 clusters can sync up successfully. > > > > >> > >> And that's why I proposed to the "last mirrored leader epoch" > > > > >> semantic in > > > > >> > >> the sub-proposal > > > > >> > >> < > > > > >> > >> > > > > >> > > > https://cwiki.apache.org/confluence/display/KAFKA/Unclean+Leader+Election+in+Cluster+Mirroring > > > > >> > >> >, > > > > >> > >> which is a solution to draw a line between these 2 clusters > to > > > > >> declare > > > > >> > >> records beyond the "last mirrored leader epoch" N, it belongs > > to > > > > >> who. I > > > > >> > >> think this should work well, as long as all replicas in the > > cluster > > > > >> can > > > > >> > >> truncate the log correctly. > > > > >> > >> > > > > >> > >> What do you think? > > > > >> > >> > > > > >> > >> Any feedback is appreciated. > > > > >> > >> > > > > >> > >> Thank you, > > > > >> > >> Luke > > > > >> > >> > > > > >> > >> On Fri, Mar 6, 2026 at 6:02 PM Andrew Schofield < > > > > >> [email protected]> > > > > >> > >> wrote: > > > > >> > >> > > > > >> > >> > Hi Fede, > > > > >> > >> > Thanks for your response. > > > > >> > >> > > > > > >> > >> > AS1: Thanks for the clarification. > > > > >> > >> > > > > > >> > >> > AS2: I expect you'll include a version bump of > > > > >> AlterShareGroupOffsets in > > > > >> > >> > this KIP, but that's a small matter compared with the rest > > of the > > > > >> > >> protocol > > > > >> > >> > changes. > > > > >> > >> > > > > > >> > >> > AS3: OK. > > > > >> > >> > > > > > >> > >> > AS4: Thanks for the details. My only comment is that it > > might be a > > > > >> bit > > > > >> > >> > laborious when you want to failover all topics. I suggest > > adding > > > > >> > >> > `--all-topics` so you could do: > > > > >> > >> > > > > > >> > >> > $ bin/kafka-mirror.sh --bootstrap-server :9094 --remove > > > > >> --all-topics > > > > >> > >> > --mirror my-mirror > > > > >> > >> > > > > > >> > >> > AS5: Thanks for the response. I understand there are good > > reasons > > > > >> for > > > > >> > >> the > > > > >> > >> > way epochs are handled in the KIP. I see that there is a > > > > >> sub-document > > > > >> > >> for > > > > >> > >> > the KIP about unclean leader election. I'll spend some time > > > > >> reviewing > > > > >> > >> that. > > > > >> > >> > > > > > >> > >> > Thanks, > > > > >> > >> > Andrew > > > > >> > >> > > > > > >> > >> > On 2026/02/18 13:27:07 Federico Valeri wrote: > > > > >> > >> > > Hi Andrew, thanks for the review. > > > > >> > >> > > > > > > >> > >> > > Let me try to answer your questions and then other > authors > > can > > > > >> join > > > > >> > >> > > the discussion. > > > > >> > >> > > > > > > >> > >> > > AS1 > > > > >> > >> > > ------ > > > > >> > >> > > > > > > >> > >> > > Destination topics are created with the same topic IDs > > using the > > > > >> > >> > > extended CreateTopics API. Then, data is replicated > > starting from > > > > >> > >> > > offset 0 with byte-for-byte batch copying, so destination > > offsets > > > > >> > >> > > always match source offsets. When failing over, we record > > the > > > > >> last > > > > >> > >> > > mirrored offset (LMO) in the destination cluster. When > > failing > > > > >> back, > > > > >> > >> > > the LMO is used for truncating and then start mirroring > the > > > > >> delta, > > > > >> > >> > > otherwise we start mirroring from scratch by truncating > to > > zero. > > > > >> > >> > > > > > > >> > >> > > Retention: If the mirror leader attempts to fetch an > > offset that > > > > >> is > > > > >> > >> > > below the current log start offset of the source leader > > (e.g. > > > > >> fetching > > > > >> > >> > > offset 50 when log start offset is 100), the source > broker > > > > >> returns an > > > > >> > >> > > OffsetOutOfRangeException that the mirror leader handles > by > > > > >> truncating > > > > >> > >> > > to the source's current log start offset and resuming > > fetching > > > > >> from > > > > >> > >> > > that point. Compaction: The mirror leader replicates > these > > > > >> compacted > > > > >> > >> > > log segments exactly as they exist in the source cluster, > > > > >> maintaining > > > > >> > >> > > the same offset assignments and gaps. > > > > >> > >> > > > > > > >> > >> > > Do you have any specific corner case in mind? > > > > >> > >> > > > > > > >> > >> > > AS2 > > > > >> > >> > > ------ > > > > >> > >> > > > > > > >> > >> > > Agreed. The current AlterShareGroupOffsetsRequest (v0) > only > > > > >> includes > > > > >> > >> > > PartitionIndex and StartOffset with no epoch field. When > > > > >> mirroring > > > > >> > >> > > share group offsets across clusters, the epoch is needed > to > > > > >> ensure the > > > > >> > >> > > offset alteration targets the correct leader generation. > > > > >> > >> > > > > > > >> > >> > > AS3 > > > > >> > >> > > ------ > > > > >> > >> > > > > > > >> > >> > > Right, the enum is now fixed. Yes, we will parse from the > > right > > > > >> and > > > > >> > >> > > apply the same naming rules used for topic name ;) > > > > >> > >> > > > > > > >> > >> > > AS4 > > > > >> > >> > > ------- > > > > >> > >> > > > > > > >> > >> > > Agreed. I'll try to improve those paragraphs because they > > are > > > > >> crucial > > > > >> > >> > > from an operational point of view. > > > > >> > >> > > > > > > >> > >> > > Let me shortly explain how it is supposed to work: > > > > >> > >> > > > > > > >> > >> > > 9091 (source) -----> 9094 (destination) > > > > >> > >> > > > > > > >> > >> > > The single operation that allows an operator to switch > all > > > > >> topics at > > > > >> > >> > > once in case of disaster is the following: > > > > >> > >> > > > > > > >> > >> > > bin/kafka-mirror.sh --bootstrap-server :9094 --remove > > --topic .* > > > > >> > >> > > --mirror my-mirror > > > > >> > >> > > > > > > >> > >> > > 9091 (source) --x--> 9094 (destination) > > > > >> > >> > > > > > > >> > >> > > After that, all mirror topics become detached from the > > source > > > > >> cluster > > > > >> > >> > > and start accepting writes (the two cluster are allowed > to > > > > >> diverge). > > > > >> > >> > > > > > > >> > >> > > When the source cluster is back, the operator can > failback > > by > > > > >> creating > > > > >> > >> > > a mirror with the same name on the source cluster (new > > > > >> destination): > > > > >> > >> > > > > > > >> > >> > > echo "bootstrap.servers=localhost:9094" > > > > > >> /tmp/my-mirror.properties > > > > >> > >> > > bin/kafka-mirrors.sh --bootstrap-server :9091 --create > > --mirror > > > > >> > >> > > my-mirror --mirror-config /tmp/my-mirror.properties > > > > >> > >> > > bin/kafka-mirrors.sh --bootstrap-server :"9091 --add > > --topic .* > > > > >> > >> > > --mirror my-mirror > > > > >> > >> > > > > > > >> > >> > > 9091 (destination) <----- 9094 (source) > > > > >> > >> > > > > > > >> > >> > > AS5 > > > > >> > >> > > ------- > > > > >> > >> > > > > > > >> > >> > > This is the core of our design and we reached that > > empirically by > > > > >> > >> > > trying out different options. We didn't want to change > > local > > > > >> > >> > > replication, and this is something you need to do when > > > > >> preserving the > > > > >> > >> > > source leader epoch. The current design is simple and > > keeps the > > > > >> epoch > > > > >> > >> > > domains entirely separate. Destination cluster is in > > charge of > > > > >> the > > > > >> > >> > > leader epoch for its own log. The source epoch is only > used > > > > >> during the > > > > >> > >> > > fetch protocol to validate responses and detect > divergence. > > > > >> > >> > > > > > > >> > >> > > The polarity idea of tracking whether an epoch bump > > originated > > > > >> from > > > > >> > >> > > replication vs. local leadership change is interesting, > > but adds > > > > >> > >> > > significant complexity and coupling between source and > > > > >> destination > > > > >> > >> > > epochs. Could you clarify what specific scenario polarity > > > > >> tracking > > > > >> > >> > > would address that the current separation doesn't handle? > > One > > > > >> case we > > > > >> > >> > > don't support is unclean leader election reconciliation > > across > > > > >> > >> > > clusters, is that the gap you're aiming at? > > > > >> > >> > > > > > > >> > >> > > I tried to rewrite the unclean leader election paragraph > > in the > > > > >> > >> > > rejected alternatives to be easier to digest. Let me know > > if it > > > > >> works. > > > > >> > >> > > > > > > >> > >> > > On Tue, Feb 17, 2026 at 2:57 PM Andrew Schofield > > > > >> > >> > > <[email protected]> wrote: > > > > >> > >> > > > > > > > >> > >> > > > Hi Fede and friends, > > > > >> > >> > > > Thanks for the KIP. > > > > >> > >> > > > > > > > >> > >> > > > It’s a comprehensive design, easy to read and has > clearly > > > > >> taken a > > > > >> > >> lot > > > > >> > >> > of work. > > > > >> > >> > > > The principle of integrating mirroring into the brokers > > makes > > > > >> total > > > > >> > >> > sense to me. > > > > >> > >> > > > > > > > >> > >> > > > The main comment I have is that mirroring like this > > cannot > > > > >> handle > > > > >> > >> > situations > > > > >> > >> > > > in which multiple topic-partitions are logically > > related, such > > > > >> as > > > > >> > >> > transactions, > > > > >> > >> > > > with total fidelity. Each topic-partition is being > > replicated > > > > >> as a > > > > >> > >> > separate entity. > > > > >> > >> > > > The KIP calls this out and describes the behaviour > > thoroughly. > > > > >> > >> > > > > > > > >> > >> > > > A few initial comments. > > > > >> > >> > > > > > > > >> > >> > > > AS1) Is it true that offsets are always preserved by > > this KIP? > > > > >> I > > > > >> > >> > *think* so but > > > > >> > >> > > > not totally sure that it’s true in all cases. It would > > > > >> certainly be > > > > >> > >> > nice. > > > > >> > >> > > > > > > > >> > >> > > > AS2) I think you need to add epoch information to > > > > >> > >> > AlterShareGroupOffsetsRequest. > > > > >> > >> > > > It really should already be there in hindsight, but I > > think > > > > >> this KIP > > > > >> > >> > requires it. > > > > >> > >> > > > > > > > >> > >> > > > AS3) The CoordinatorType enum for MIRROR will need to > be > > 3 > > > > >> because 2 > > > > >> > >> > is SHARE. > > > > >> > >> > > > I’m sure you’ll parse the keys from the right ;) > > > > >> > >> > > > > > > > >> > >> > > > AS4) The procedure for achieving a failover could be > > clearer. > > > > >> Let’s > > > > >> > >> > say that I am > > > > >> > >> > > > using cluster mirroring to achieve DR replication. My > > source > > > > >> cluster > > > > >> > >> > is utterly lost > > > > >> > >> > > > due to a disaster. What’s the single operation that I > > perform > > > > >> to > > > > >> > >> > switch all of the > > > > >> > >> > > > topics mirrored from the lost source cluster to become > > the > > > > >> active > > > > >> > >> > topics? > > > > >> > >> > > > Similarly for failback. > > > > >> > >> > > > > > > > >> > >> > > > AS5) The only piece that I’m really unsure of is the > > epoch > > > > >> > >> management. > > > > >> > >> > I would > > > > >> > >> > > > have thought that the cluster which currently has the > > writable > > > > >> > >> > topic-partition > > > > >> > >> > > > would be in charge of the leader epoch and it would not > > be > > > > >> > >> necessary to > > > > >> > >> > > > perform all of the gymnastics described in the section > > on epoch > > > > >> > >> > rewriting. > > > > >> > >> > > > I have read the Rejected Alternatives section too, but > I > > don’t > > > > >> fully > > > > >> > >> > grasp > > > > >> > >> > > > why it was necessary to reject it. > > > > >> > >> > > > > > > > >> > >> > > > I wonder if we could store the “polarity” of an epoch, > > > > >> essentially > > > > >> > >> > whether the > > > > >> > >> > > > epoch bump was observed by replication from a source > > cluster, > > > > >> or > > > > >> > >> > whether > > > > >> > >> > > > it was bumped by a local leadership change when the > > topic is > > > > >> locally > > > > >> > >> > writable. > > > > >> > >> > > > When a topic-partition switches from read-only to > > writable, we > > > > >> > >> should > > > > >> > >> > definitely > > > > >> > >> > > > bump the epoch, and we could record the fact that it > was > > a > > > > >> local > > > > >> > >> epoch. > > > > >> > >> > > > When connectivity is re-established, you might find > that > > both > > > > >> ends > > > > >> > >> have > > > > >> > >> > > > declared a local epoch N, but someone has to win. > > > > >> > >> > > > > > > > >> > >> > > > Thanks, > > > > >> > >> > > > Andrew > > > > >> > >> > > > > > > > >> > >> > > > > On 14 Feb 2026, at 07:17, Federico Valeri < > > > > >> [email protected]> > > > > >> > >> > wrote: > > > > >> > >> > > > > > > > > >> > >> > > > > Hi, we would like to start a discussion thread about > > > > >> KIP-1279: > > > > >> > >> > Cluster > > > > >> > >> > > > > Mirroring. > > > > >> > >> > > > > > > > > >> > >> > > > > Cluster Mirroring is a new Kafka feature that enables > > native, > > > > >> > >> > > > > broker-level topic replication across clusters. > Unlike > > > > >> > >> MirrorMaker 2 > > > > >> > >> > > > > (which runs as an external Connect-based tool), > Cluster > > > > >> Mirroring > > > > >> > >> is > > > > >> > >> > > > > built into the broker itself, allowing tighter > > integration > > > > >> with > > > > >> > >> the > > > > >> > >> > > > > controller, coordinator, and partition lifecycle. > > > > >> > >> > > > > > > > > >> > >> > > > > > > > > >> > >> > > > > > >> > >> > > > > >> > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-1279%3A+Cluster+Mirroring > > > > >> > >> > > > > > > > > >> > >> > > > > There are a few missing bits, but most of the design > is > > > > >> there, so > > > > >> > >> we > > > > >> > >> > > > > think it is the right time to involve the community > > and get > > > > >> > >> feedback. > > > > >> > >> > > > > Please help validating our approach. > > > > >> > >> > > > > > > > > >> > >> > > > > Thanks > > > > >> > >> > > > > Fede > > > > >> > >> > > > > > > > >> > >> > > > > > > >> > >> > > > > > >> > >> > > > > >> > > > > > > >> > > > > > > >> > > -- > > > > >> > > Paolo Patierno > > > > >> > > > > > > >> > > *Senior Principal Software Engineer @ IBM**CNCF Ambassador* > > > > >> > > > > > > >> > > Twitter : @ppatierno <http://twitter.com/ppatierno> > > > > >> > > Linkedin : paolopatierno < > > http://it.linkedin.com/in/paolopatierno> > > > > >> > > GitHub : ppatierno <https://github.com/ppatierno> > > > > >> > > > > > > >> > > > > > >> > > > > > >> > -- > > > > >> > Paolo Patierno > > > > >> > > > > > >> > *Senior Principal Software Engineer @ IBM**CNCF Ambassador* > > > > >> > > > > > >> > Twitter : @ppatierno <http://twitter.com/ppatierno> > > > > >> > Linkedin : paolopatierno < > http://it.linkedin.com/in/paolopatierno > > > > > > > >> > GitHub : ppatierno <https://github.com/ppatierno> > > > > >> > > > > > >> > > > > > > > > > >
