Question

A Kafka Streams app doing stateful aggregation (RocksDB state stores, 48 partitions, 8 instances) is deployed via a rolling restart for a routine config change. During the rollout, end-to-end lag spikes to several million and stays elevated for ~25 minutes per instance bounce; the app's `rebalance` JMX metrics fire on every pod restart, and instances spend minutes in a `RESTORING` state reading from the `__changelog` topic. The cluster is otherwise healthy. This rolling restart happens on every deploy and each one causes a multi-minute lag spike. Triage why the lag spikes during deploys and how you'd mitigate it.