2 / 6 Backpressure, DLQ & Retries

Depth Versus Lag

A team moves a pipeline from SQS to Kafka and ports its alarm on approximate queue depth to an alarm on topic size in bytes. During a real consumer outage the new alarm never fires. Why not?

A

Size metrics are emitted only when a group commits an offset, so a stalled group emits none

B

Kafka keeps records for its whole retention window whether or not any consumer has read them

C

Topic size is reported per broker, so one stalled consumer is averaged away across the cluster

D

Kafka compacts the topic as consumers fall behind, keeping reported size flat during a backlog