[KAFKA-3810] replication of internal topics should not be limited by replica.fetch.max.bytes

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 0.10.1.0
• Assignee: Onur Karaman
• Reporter: Onur Karaman
• Created: 09/Jun/16 09:42
• Updated: 08/Aug/16 12:52
• Resolved: 19/Jun/16 18:42

Description

From the kafka-dev mailing list discussion:
[DISCUSS] scalability limits in the coordinator

There's a scalability limit on the new consumer / coordinator regarding the amount of group metadata we can fit into one message. This restricts a combination of consumer group size, topic subscription sizes, topic assignment sizes, and any remaining member metadata.

Under more strenuous use cases like mirroring clusters with thousands of topics, this limitation can be reached even after applying gzip to the __consumer_offsets topic.

Various options were proposed in the discussion:

1. Config change: reduce the number of consumers in the group. This isn't always a realistic answer in more strenuous use cases like MirrorMaker clusters or for auditing.
2. Config change: split the group into smaller groups which together will get full coverage of the topics. This gives each group member a smaller subscription.(ex: g1 has topics starting with a-m while g2 has topics starting with n-z). This would be operationally painful to manage.
3. Config change: split the topics among members of the group. Again this gives each group member a smaller subscription. This would also be operationally painful to manage.
4. Config change: bump up KafkaConfig.messageMaxBytes (a topic-level config) and KafkaConfig.replicaFetchMaxBytes (a broker-level config). Applying messageMaxBytes to just the __consumer_offsets topic seems relatively harmless, but bumping up the broker-level replicaFetchMaxBytes would probably need more attention.
5. Config change: try different compression codecs. Based on 2 minutes of googling, it seems like lz4 and snappy are faster than gzip but have worse compression, so this probably won't help.
6. Implementation change: support sending the regex over the wire instead of the fully expanded topic subscriptions. I think people said in the past that different languages have subtle differences in regex, so this doesn't play nicely with cross-language groups.
7. Implementation change: maybe we can reverse the mapping? Instead of mapping from member to subscriptions, we can map a subscription to a list of members.
8. Implementation change: maybe we can try to break apart the subscription and assignments from the same SyncGroupRequest into multiple records? They can still go to the same message set and get appended together. This way the limit become the segment size, which shouldn't be a problem. This can be tricky to get right because we're currently keying these messages on the group, so I think records from the same rebalance might accidentally compact one another, but my understanding of compaction isn't that great.
9. Implementation change: try to apply some tricks on the assignment serialization to make it smaller.
10. Config and Implementation change: bump up the __consumer_offsets topic messageMaxBytes and (from junrao) fix how we deal with the case when a message is larger than the fetch size. Today, if the fetch size is smaller than the fetch size, the consumer will get stuck. Instead, we can simply return the full message if it's larger than the fetch size w/o requiring the consumer to manually adjust the fetch size.
11. Config and Implementation change: same as above but only apply the special fetch logic when fetching from internal topics

Issue Links

• links to: GitHub Pull Request #1484
• mentioned in: Page Loading...