Consumer Groups & Rebalancing

Every pod stop and start is a separate membership change, and with eager rebalancing each one pauses the entire group, so a rolling deploy of N pods costs 2N stop-the-world pauses unless you switch to cooperative, sticky, and static-membership strategies.

Reassignment copies data online, so lag spikes come from disk and network saturation, throttling that trades rate for duration, leader moves that force metadata refresh, cold page cache on the destination, and coordination churn -- mitigate by monitoring replication throughput, avoiding peak hours, and throttling appropriately.

Within a group, partitions are divided so each is owned by one consumer; across groups, every group gets all partitions independently.

Kafka spreads partitions evenly across consumers, and a rebalance redistributes them whenever a consumer joins or fails.

Eager revokes all partitions and pauses the whole group; cooperative-sticky keeps existing assignments and only moves the partitions that truly need to move.