A poison record can stop a connector task even while the brokers and other records remain healthy. An error policy needs to say which errors can be tolerated, where rejected records go and who investigates them. Setting tolerance to all without observable routing can hide data loss. A dead-letter topic preserves the failed record for repair, but it does not fix the original data or promise every possible connector failure is recoverable. These tasks contain one valid record and one deliberately invalid record. Configure the local sink policy, run it and inspect both the destination table and the dead-letter topic. The original Kafka input remains available for an agreed replay.
Inspect the topics, partition logs and group offsets before changing the cluster.