Versions Compared

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

...

There is a subtle case where data is read from the log end offset and an append occurs immediately afterwards but before the lag calculation is complete. To make this more robust, the LogReadResult should return the LogOffsetMetadata corresponding to the LEO just before the read started.

Expected Scenarios

1. High volume topics - Follower for any high volume topic does not stay in ISR only if it cannot read from the log end offset for replica.lag.max.ms time i.e. it is genuinely not able to keep up. Previously it would fall out of ISR even if a single large batch was produced

2. Low volume topics - 

 Compatibility, Deprecation, and Migration Plan

This change is fully backward compatible. The only difference is that customers will no longer have to set the replica.lag.max.messages config.

...