1 / 7 Log-Based Streaming & Kafka

Eight Idle Consumers

A topic has 12 partitions. Your consumer group is falling behind, so you scale it from 12 to 20 instances. CPU on the new pods sits near zero and throughput does not move. Why?

A

Consumer parallelism is capped by the replication factor of 3, so scaling brokers is the only lever

B

Each partition goes to at most one consumer in a group, so 8 instances get no assignment; you need more partitions to go wider

C

Members of a group split each partition by offset range, so the extra instances need a longer poll timeout to claim their slice

D

The group coordinator hands records to all 20 members evenly, so the bottleneck has to be broker side