[KAFKA-16007] ZK migrations can be slow for large clusters

• Type: Improvement
• Status: Resolved
• Priority: Minor
• Resolution: Fixed
• Fix Version/s: 3.7.0, 3.6.2
• Component/s: controller, kraft
• Assignee: David Arthur
• Reporter: David Arthur
• Created: 13/Dec/23 22:04
• Updated: 26/Dec/23 20:00
• Resolved: 26/Dec/23 20:00

Description

On a large cluster with many single-partition topics, the ZK to KRaft migration took nearly half an hour:

[KRaftMigrationDriver id=9990] Completed migration of metadata from ZooKeeper to KRaft. 157396 records were generated in 2245862 ms across 67132 batches. The record types were {TOPIC_RECORD=66282, PARTITION_RECORD=72067, CONFIG_RECORD=17116, PRODUCER_IDS_RECORD=1, ACCESS_CONTROL_ENTRY_RECORD=1930}. The current metadata offset is now 332267 with an epoch of 19. Saw 36 brokers in the migrated metadata [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35].

This is a result of how we generate batches of records when traversing the ZK tree. Since we now using metadata transactions for the migration, we can re-batch these without any consistency problems.

Issue Links

• links to: GitHub Pull Request #15007