Search before asking
Description
Description
The current LeaderReplicaDistributionGoal balances leader counts across the whole cluster. However, leaders of an individual table can still be unevenly distributed.
For a hot table, this may cause significant CPU and request-load differences between TabletServers even when the cluster-wide leader count is balanced.
Proposed solution
Make leader balancing table-aware:
- Balance leaders independently within each table when replica placement permits.
- Prefer leadership movement between existing replicas.
- Move replicas when necessary to achieve table-level leader balance, such as during cluster expansion.
- Preserve rack awareness and the existing replica distribution constraints.
- Keep cluster-level leader balancing as a complementary goal.
Willingness to contribute
Search before asking
Description
Description
The current
LeaderReplicaDistributionGoalbalances leader counts across the whole cluster. However, leaders of an individual table can still be unevenly distributed.For a hot table, this may cause significant CPU and request-load differences between TabletServers even when the cluster-wide leader count is balanced.
Proposed solution
Make leader balancing table-aware:
Willingness to contribute