[KAFKA-7738] Track partition leader epochs in client metadata

• Type: Improvement
• Status: Resolved
• Priority: Minor
• Resolution: Fixed
• Fix Version/s: 2.2.0
• Assignee: David Arthur
• Reporter: David Arthur
• Created: 14/Dec/18 17:23
• Updated: 17/Jan/19 18:15
• Resolved: 17/Jan/19 18:15

Description

The Metadata API now exposes the current leader epoch of each partition. We can leverage this information to be smarter in how we fetch metadata. For example, when we receive a NOT_LEADER_FOR_PARTITION, we know to look for an epoch bump before resuming whatever operation we were trying to do. Additionally, this gives us a way to detect stale metadata.

This requires a little more sophistication in how we track metadata in the client. For example, we may be not be able to assume metadata updates are monotonic. In other words, they may mix stale metadata for one partition and fresh metadata for another. I am not sure we have any strong guarantees on the order in which metadata updates are seen on each broker.

It may be helpful in this context to control the topics that we fetch metadata for at a finer granularity. Potentially we could even extend the Metadata API to specify a subset of the partitions that the client is interested in.

Issue Links

• links to: GitHub Pull Request #6045