Question

A real-time pricing consumer (Kafka, Protobuf payloads) crash-loops on one partition starting 10:30. Dashboards: partition-7 lag climbs while partitions 0–6 are healthy; the consumer throws `InvalidProtocolBufferException` then exits and is restarted by the orchestrator, re-reads the same offset, and crashes again every few seconds; the consumer group keeps rebalancing because the crashing member leaves and rejoins. A producer team deployed a change that emits a NEW message type onto the SAME topic without a type discriminator, and these new messages only landed on partition-7 due to keying. How do you triage, recover partition-7 throughput, and handle the incompatible messages?