Kafka Improvement Proposals

2.2.0 (partially implemented)

/ 2.3.0 (WIP)

KIP-164 - Add UnderMinIsrPartitionCount and per-partition UnderMinIsr metrics

KIP-163: Lower the Minimum Required ACL Permission of OffsetFetch

0.11.0.0

0.11.0.0

KIP-106 - Change Default unclean.leader.election.enabled from True to False

KIP-185: Make exactly once in order delivery the default producer setting

KIP-291: Separating controller connections and requests from the data plane

Discussion

KIP-379: Multiple Consumer Group Management

Voting in progress (restarted 18th January 2019, due to no votes in first attempt)

Discussion

KIP-396: Add Commit/List Offsets Operations to AdminClient

Discussion

KIP-409: Allow creating under-replicated topics and partitions

KIP-410: Add metric for request handler thread pool utilization by request type

KIP-411: Make default Kafka Connect worker task client IDs distinct

Discussion

KIP-432: Additional Broker-Side Opt-In for Default, Unsecure SASL/OAUTHBEARER Implementation

Functionality can be realized with existing API although at a lower level. 

KIP-34 Add Partitioner Change Listener to Partitioner Interface for Multiple Use Case

Duplicated by KIP-160

• KIP-253 - partition expansion: We discussed a few things. (1) Is it useful to backfill a compacted topic? The main use case is to rebuild the application states. If the new partition has the existing data, rebuilding the state can be done easily by reading from a single partition. Otherwise, an application has to read both the child and the parent partition to rebuild the state. This is possible, but can be complicated. Jan will do an exercise to see how complicated this is. (2) What's the best way to add the backfilling support if we want to do it? We can do this on the server side or on the client side. The former potentially makes the coordination easier. The latter potentially reduces the memory footprint on the server for reshuffling. We need to think through how to support EOS message format and how to throttle the process to avoid overwhelming the cluster. (3) Linear hashing vs doubling partitions. It seems that Linear hashing is more general. (4) Partition splitting in Kinesis. This is done differently since it doesn't allow customized partitioning. It doesn't support compacted topic either. (5) Sticky partition assignment. It could be useful to support a partition assignment strategy where the child partition is assigned together with the parent partition to a consumer instance so that the local state doesn't have to be moved immediately. (6) Consumer callback on partition splitting. This could still be useful if the states are maintained globally.

• KIP-112 - Handle disk failure for JBOD: We discussed whether we need to support JBOD directly in Kafka or just rely on the 1 disk per broker model. The general consensus is that direct JBOD support in Kafka is needed. There is some concern on the complexity added to Kafka. So, we have to be careful with the implementation details. We discussed how directory failure should be detected, where the failure state is kept, and whether the state should be reset on broker restart. There is a bit confusing on what's written in the wiki. Dong is going to clarify the proposal based on the feedback and we will follow up on the details in the mailing list.

• KIP-82 - add record header: We agreed that there are use cases for third-party vendors building tools around Kafka. We haven't reached the conclusion whether the added complexity justifies the use cases. We will follow up on the mailing list with use cases, container format people have been using, and details on the proposal.

• KIP-54 (Sticky Partition Assignment): aims to minimise partition movement so that resource reinitialisation (e.g. caches) is minimised. It is partially sticky and partially fair. Some concerns around the fact that user code for partitionsRevoked and partitionsAssigned would have to be changed to work correctly with this assignment strategy. Good: more complex usage of an assigner that takes advantage of the user data field. Vahid will start the vote.

• KIP-72 (Allow Sizing Incoming Request Queue in Bytes): large requests can kill the broker, no control over how much memory is allocated. Client quotas don't help as damage may already have been done by the time they kick in. There was a discussion on whether it was worth it to avoid the immediate return from select when there was no memory available in the pool. Radai will update the KIP to describe this aspect in more detail as well as the config validation that is performed.

• KIP-79 (ListOffsetRequest/ListOffsetResponse v1 and add timestamp search methods to the new consumer): we discussed the option of passing multiple timestamps for the same partition in the same request. Becket thinks it's a rare use case and not worth supporting. Gwen said that it would be nice to have, but not essential. We talked about validation of duplicate topics. Becket will check the approach taken by the create topics request and evaluate if it can be adopted here too. PR will be available today and Jason will evaluate if it's feasible to include it in the next release once it's available.

• KIP48 (delegation tokens): Harsha will update the wiki with more details on how to use delegation tokens and how to configure it.
• KIP-78 (cluster id): There was discussion on adding human readable tags later. No major concerns.

• time-based release: No one seems to have objections. Ismael will follow up with a release wiki.
• KIP-4: We discussed having separate ACL requests of add and delete. No one seems to object to it. We discussed the admin client. Grant will send a PR. We discussed how KStream can use the ACL api.  It seems that we will need some kind of regex or namespace support in ACL to make the authorization convenient in KStream.
• KIP-50: There is some discussion for further changes in the PR. Ashish will reply to the KIP email thread with the recommended changes. Ashish/Grant plan to look into whether it's possible to make the authorizer api change backward compatible. However, it seems that people are in general ok with a non-compatible api change.
• KIP-74: No objections on the current proposal.
• Java 7 support timeline: The consensus is to defer dropping the Java 7 support until the next major release (which will be next year). Ismael will follow up on the email thread.
• KIP-48 delegation token : Ashish will ping Harsh to see if this is still active.
• Some of the KIPs have been idle. Grant will send a proposal on tagging them properly (e.g., blocked, inactive, no resource, etc).

• KIP-58 - Make Log Compaction Point Configurable: We want to start with just a time-based configuration since there is no good usage for byte-based or message-based configuration. Eric will change the KIP and start the vote.
• KIP-4 - Admin api: Grant will pick up the work. Initially, he plans to route the write requests from the admin clients to the controller directly to avoid having the broker forward the requests to the controller.
• KIP-48 - Delegation tokens: Two of the remaining issues are (1) how to store the delegation tokens and (2) how token expiration works. Since Parth wasn't able to attend the meeting. We will follow up in the mailing list.

• KIP-4: There is a slight debate on the metadata request schema, as well as the internal ZK based implementation, which we will wait for Jun to comment on the mailing list thread.
• KIP-52: We decided to start a voting process for this.
• KIP-35: Decided on renaming ApiVersionQuery api to ApiVersion. Consensus on using the api in java client to only check for availability of current versions. ApiVersion api's versions will not be deprecated. Update KIP-35 wiki will be updated with latest info and vote thread will be initiated.

• KIP-33 - Add a time based log index to Kafka: We decided NOT to include this in 0.10.0 since the changes may have performance risks.
• KIP-45 - Standardize all client sequence interaction on j.u.Collection: There is no consensus in the discussion. We will just put it to vote.
• KIP-35 - Retrieving protocol version: This gets the longest discussion. There is still no consensus. Magnus thinks the current proposal of maintaining a global protocol version won't work and will try to submit a new proposal.
• KIP-43 - Kafka SASL enhancements: Rajini will modify the KIP to only support native SASL mechanisms and leave the changes to Login and CallbackHandler to KIP-44 instead.

• KIP-33 and KIP-47: No issues. Will start the voting thread.
• KIP-43: We discussed whether there is a need to support multiple SASL mechanisms at the same time and what's the best way to implement this. Will discuss this in more details in the email thread.
• KIP-4: Grant gave a comprehensive summary of the current state. We have gaps on how to make the admin request block on the broker, how to integrate admin requests with ACL (especially with respect to client config changes for throttling and ACL changes), how to do the alter topic request properly. Grant will update the KIP with an interim plan and a long term plan.
• KIP-43: We briefly discussed on to support multiple sasl mechanisms on the broker. Harsha will follow up with more details on the email thread.
• Everyone seems to be in favor of making the next major release 0.10.0, instead of 0.9.1.

• KIP-42: We agreed to leave the broker side interceptor for another KIP. On the client side, people favor the 2nd option in Anna's proposal. Anna will update the wiki accordingly.
• KIP-43: We discussed whether there is a need to support multiple SASL mechanisms at the same time and what's the best way to implement this. Will discuss this in more details in the email thread.
• Jiangjie brought up an issue related to KIP-32 (adding timestamp field in the message). The issue is that currently there is no convenient way for the consumer to tell whether the timestamp in a message is the create time or the server time. He and Guozhang propose to use a bit in the message attribute to do that. Jiangjie will describe the proposal in the email thread.

• KIP-41: Discussed whether the issue of long processing time between poll calls is a common issue and whether we should revisit the poll api. Also discussed whether the number of records returned in poll calls can be made more dynamic. In the end, we feel that just adding a config that controls the number records returned in poll() is the simplest approach at this moment.
• KIP-36: Need to look into how to change the broker JSON representation in ZK w/o breaking rolling upgrades. Otherwise, ready for voting.

• KIP-38: No concerns with this KIP. Flavio will initiate the voting on this.
• KIP-37: There are questions on how ACL, configurations, etc will work, and whether we should support "move" or not. We will discuss the details more in the mailing list.
• KIP-32/KIP-33: Jiangjie raised some concerns on the approach that Jay proposed. Guozhang and Jay will follow up on the mailing list.

• 0.9.0 release: We discussed if KAFKA-2397 should be a blocker in 0.9.0. Jason and Guozhang will follow up on the jira.
• KIP-32 and KIP-33: We discussed Jay's alternative proposal of just keeping CreateTime in the message and having a config to control how far off the CreateTime can be from the broker time. We will think a bit more on this and Jiangjie will update the KIP wiki.
• KIP-36: We discussed an alternative approach of introducing a new broker property to designate the rack. It's simpler and potentially can work in the case when the broker to rack mapping is maintaining externally. We need to make sure that we have an upgrade plan for this change. Allen will update the KIP wiki

• We only had the time to go through KIP-35. The consensus is that we will add a BrokerProtocolRequest that returns the supported versions for every type of requests. It's up to the client to decide how to use this. Magnus will update the KIP wiki with more details.

• KIP-31: Need to figure out how to evolve inter.broker.protocol.version with multiple protocol changes within the same release, mostly for people who are deploying from trunk. Becket will update the wiki.
• KIP-32/KIP-33: Having both CreateTime and LogAppendTime per message adds significant overtime. There are a couple of possibilities to improve this. Becket will follow up on this.
• LinkedIn has been testing SSL in MirrorMaker (SSL is only enabled in the producer). So far, MirrorMaker can keep up with the load. LinkedIn folks will share some of the performance results.

• KIP-28: Discussed the improved proposal including 2 layers of API (the higher layer is for streaming DSL), and stream time vs processor time. Ready for review.
• KIP-31, KIP-32: (1) Discussed whether the timestamp should be from the client or the broker. (2) Discussed the migration path and whether this requires all consumers to upgrade before the new message format can be used. (3) Since this is too big a change, it will NOT be included in 0.9.0 release. Becket will update the wiki.

• client-side assignment strategy: We discussed concerns about rebalancing time due to metadata inconsistency, especially when lots of topics are subscribed. Will discuss a bit more on the mailing list. 
• CopyCat data api: The discussions are in KAFKA-2367 for people who are interested.
• 0.8.2.2: We want to make this a low risk bug fix release since 0.8.3 is coming. So, will only include a small number of critical and small fixes.
• 0.8.3: The main features will be security and the new consumer. We will be cutting a release branch when the major pieces for these new features have been committed.

• KIP-29: we will do a quick fix for unblocking production issues with hard-coded interval values, still needs further discussion on KIP itself.
• KIP-26: RP-99 (https://github.com/apache/kafka/pull/99) pending for reviews.
• KIP-28: RP-130 (https://github.com/apache/kafka/pull/130) needs further discussion.
• KIP-4: KAFKA-1695 / 2210 pending for reviews.