Start with the broker, not the symptom.
Under-replicated partitions, an offline broker or a controller that is not there explain more symptoms than anything a consumer does.
The Kafka overview reads whether the cluster is healthy or degraded, its brokers, its under-replicated partitions and the version it really runs.
Reading a Kafka cluster you did not build →With JMX set up, every broker and controller reports what it measures about itself — offline partitions, partitions under their minimum in-sync replicas, request times, how idle its handlers are.
Kafka broker health over JMX →Find the consumer that stopped, and why.
A backlog is a number until you know which reader is not reading and whether it is coming back.
Read a group's state before its lag, then the lag in messages and in time, the consume rate, and the one member that is actually stuck.
Consumer lag without guesswork →For RabbitMQ and Redis: whether to drain, divert or let a backlog run, and what each choice costs.
Queue backlogs →Which stream entries were delivered and never acknowledged, to whom, and for how long.
Redis pending entries →The history of the cluster is on the record.
Most incidents follow a change. The audit trail says what changed, who changed it, where and why.
Every write and every refusal, filtered by who, what, where and when — one trail for every platform.
The audit trail →Every incident the evaluator opened and closed, with the rule, the target and the reading — so "did this happen before" has an answer.
Alert rules you can trust →