KIP-1358 — Gradual and gated preferred-leader election in the KRaft controller

Discussion KRaftBroker

Gates and throttles preferred-leader election in the KRaft controller, holding leadership back from a broker that still has global catch-up work outstanding and making the per-run election batch configurable instead of the hardcoded 1000. Today the controller returns leadership per partition the instant that partition re-enters the ISR, so a restarted or rebuilt broker takes leadership in one burst while its disk and network are still saturated by replica fetches, spiking produce P99 and prolonging recovery of __consumer_offsets, __transaction_state, and __share_group_state.

Details