Hi Lijun Tong: Thanks for your KIP which raise this critical issue.
what about just keep one topic instead of involve another topic. for existed topic data's migration. maybe we can use this way to solve the issue: (1) set the retention date > all of topic which enable remote storage's retention time (2) deploy new kafka version with feature: which send the message with key (3) wait all the message expired and new message with key coming to the topic (4) convert the topic to compact I don't test it. Just propose this solution according to code review result. just for your reference. The steps maybe a little complex. but it can avoiding add new topic. Regards Jian Lijun Tong <[email protected]> 于2026年1月8日周四 09:17写道： > Hey Kamal, > > > Thanks for your time for the review. > > > Here is my response to your questions: > > Q1 At this point, I don’t see a need to change > RemoteLogMetadataTopicPartitioner for this design. Nothing in the current > approach appears to require a partitioner change, but I’m open to > revisiting if a concrete need arises. > > Q2 I have some reservations about using SegmentId:State as the key. A > practical challenge we see today is that the same logical segment can be > retried multiple times with different SegmentIds across brokers. If the key > is SegmentId-based, it becomes harder to discover and tombstone all related > attempts when the segment eventually expires. The > TopicId:Partition:EndOffset:BrokerLeaderEpoch key is deterministic for a > logical segment attempt and helps group retries by epoch, which simplifies > cleanup and reasoning about state. I’d love to understand the benefits > you’re seeing with SegmentId:State compared to the offset/epoch-based key > so we can weigh the trade-offs. > > On partitioning: with this proposal, all states for a given user > topic-partition still map to the same metadata partition. That remains true > for the existing __remote_log_metadata (unchanged partitioner) and for the > new __remote_log_metadata_compacted, preserving the properties > RemoteMetadataCache relies on. > > Q3 It should be fine for ConsumerTask to ignore tombstone records (null > values) and no-op. > > Q4 Although TBRLMM is a sample RLMM implementation, it’s currently the only > OSS option and is widely used. The new __remote_log_metadata_compacted > topic offers clear operational benefits in that context. We can also > provide a configuration to let users choose whether they want to keep the > audit topic (__remote_log_metadata) in their cluster. > > Q4a Enabling compaction on __remote_log_metadata alone may not fully > address the unbounded growth, since we also need to emit tombstones for > expired keys to delete them. Deferring compaction and tombstoning to user > configuration could make the code flow complicated, also add operational > complexity and make outcomes less predictable. The proposal aims to provide > a consistent experience by defining deterministic keys and emitting > tombstones as part of the broker’s responsibilities, while still allowing > users to opt out of the audit topic if they prefer. But I am open to more > discussion if there is any concrete need I don't foresee. > > > Thanks, > > Lijun Tong > > Kamal Chandraprakash <[email protected]> 于2026年1月6日周二 > 01:01写道： > > > Hi Lijun, > > > > Thanks for the KIP! Went over the first pass. > > > > Few Questions: > > > > 1. Are we going to maintain the same RemoteLogMetadataTopicPartitioner > > < > > > https://sourcegraph.com/github.com/apache/kafka/-/blob/storage/src/main/java/org/apache/kafka/server/log/remote/metadata/storage/RemoteLogMetadataTopicPartitioner.java > > > > > for both the topics? It is not clear in the KIP, could you clarify it? > > 2. Can the key be changed to SegmentId:State instead of > > TopicId:Partition:EndOffset:BrokerLeaderEpoch if the same partitioner is > > used? It is good to maintain all the segment states for a > > user-topic-partition in the same metadata partition. > > 3. Should we have to handle the records with null value (tombstone) in > the > > ConsumerTask > > < > > > https://sourcegraph.com/github.com/apache/kafka/-/blob/storage/src/main/java/org/apache/kafka/server/log/remote/metadata/storage/ConsumerTask.java?L166 > > > > > ? > > 4. TBRLMM > > < > > > https://sourcegraph.com/github.com/apache/kafka/-/blob/storage/src/main/java/org/apache/kafka/server/log/remote/metadata/storage/TopicBasedRemoteLogMetadataManager.java > > > > > is a sample plugin implementation of RLMM. Not sure whether the community > > will agree to add one more internal topic for this plugin impl. > > 4a. Can we modify the new messages to the __remote_log_metadata topic to > > contain the key and leave it to the user to enable compaction for this > > topic if they need? > > > > Thanks, > > Kamal > > > > On Tue, Jan 6, 2026 at 7:35 AM Lijun Tong <[email protected]> > wrote: > > > > > Hey Henry, > > > > > > Thank you for your time and response! I really like your KIP-1248 about > > > offloading the consumption of remote log away from the broker, and I > > think > > > with that change, the topic that enables the tiered storage can also > have > > > longer retention configurations and would benefit from this KIP too. > > > > > > Some suggestions: In your example scenarios, it would also be good to > add > > > > an example of remote log segment deletion triggered by retention > policy > > > > which will trigger generation of tombstone event into metadata topic > > and > > > > trigger log compaction/deletion 24 hour later, I think this is the > key > > > > event to cap the metadata topic size. > > > > > > > > > Regarding to this suggestion, I am not sure whether Scenario 4 > > > < > > > > > > https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=406618613#KIP1266:BoundingTheNumberOfRemoteLogMetadataMessagesviaCompactedTopic-Scenario4:SegmentDeletion > > > > > > > has > > > covered it. I can add more rows in the Timeline Table like T5+24hour to > > > indicate the messages are gone by then to explicitly show that messages > > are > > > deleted, thus the number of messages are capped in the topic. > > > > > > Regarding whether the topic __remote_log_metadata is still necessary, I > > am > > > inclined to continue to have this topic at least for debugging purposes > > so > > > we can build confidence about the compacted topic change, we can > > > always choose to remove this topic in the future once we all agree it > > > provides limited value for the users. > > > > > > Thanks, > > > Lijun Tong > > > > > > > > > Henry Haiying Cai via dev <[email protected]> 于2026年1月5日周一 16:19写道： > > > > > > > Lijun, > > > > > > > > Thanks for the proposal and I liked your idea of using a compacted > > topic > > > > for tiered storage metadata topic. > > > > > > > > In our setup, we have set a shorter retention (3 days) for the tiered > > > > storage metadata topic to control the size growth. We can do that > > since > > > we > > > > control all topic's retention policy in our clusters and we set a > > uniform > > > > retention.policy for all our tiered storage topics. I can see other > > > > users/companies will not be able to enforce that retention policy to > > all > > > > tiered storage topics. > > > > > > > > Some suggestions: In your example scenarios, it would also be good to > > add > > > > an example of remote log segment deletion triggered by retention > policy > > > > which will trigger generation of tombstone event into metadata topic > > and > > > > trigger log compaction/deletion 24 hour later, I think this is the > key > > > > event to cap the metadata topic size. > > > > > > > > For the original unbounded remote_log_metadata topic, I am not sure > > > > whether we still need it or not. If it is left only for audit trail > > > > purpose, people can set up a data ingestion pipeline to ingest the > > > content > > > > of metadata topic into a separate storage location. I think we can > > have > > > a > > > > flag to have only one metadata topic (the compacted version). > > > > > > > > > > > > On Monday, January 5, 2026 at 01:22:42 PM PST, Lijun Tong < > > > > [email protected]> wrote: > > > > > > > > > > > > > > > > > > > > > > > > Hello Kafka Community, > > > > > > > > I would like to start a discussion on KIP-1266, which proposes to add > > > > another new compacted remote log metadata topic for the tiered > storage, > > > to > > > > limit the number of messages that need to be iterated to build the > > remote > > > > metadata state. > > > > > > > > KIP link: KIP-1266 Bounding The Number Of RemoteLogMetadata Messages > > via > > > > Compacted RemoteLogMetadata Topic > > > > < > > > > > > > > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-1266%3A+Bounding+The+Number+Of+RemoteLogMetadata+Messages+via+Compacted+Topic > > > > > > > > > > > > > Background: > > > > The current Tiered Storage implementation uses a > __remote_log_metadata > > > > topic with infinite retention and delete-based cleanup policy, > causing > > > > unbounded growth, slow broker bootstrap, no mechanism to clean up > > expired > > > > segment metadata, and inefficient re-reading from offset 0 during > > > > leadership changes. > > > > > > > > Proposal: > > > > A dual-topic approach that introduces a new > > > __remote_log_metadata_compacted > > > > topic using log compaction with deterministic offset-based keys, > while > > > > preserving the existing topic for audit history; this allows brokers > to > > > > build their metadata cache exclusively from the compacted topic, > > enables > > > > cleanup of expired segment metadata through tombstones, and includes > a > > > > migration strategy to populate the new topic during > upgrade—delivering > > > > bounded metadata growth and faster broker startup while maintaining > > > > backward compatibility. > > > > > > > > More details are in the attached KIP link. > > > > Looking forward to your thoughts. > > > > > > > > Thank you for your time! > > > > > > > > Best, > > > > Lijun Tong > > > > > > > > > >
