My use-case is quite simple: I want to achieve Pulsar as a proper queue. Process individual keys in order Process. If there is a validation failure during processing, negatively ack the message to re-process it again later. I am using Key_shared subscription to achieve the same. But somehow message dispatch from broker to consumers seems flaky. Even if there is backlog, often most of the times except for 1-2 consumers rest remain idle. Restart works. See more even distribution of polled messages across consumers for some time. But it comes back in the same state within few hours. Before pipeline goes to a stuck state, last logs are always some consumer doing negative ack. Not sure how I am using negative ack is wrong or not. I am not using retry-letter topic. Relying upon the negatively acked messages to get redelivered by the broker to existing consumers. I am negatively ack upto 10 times, then handling it in the code to ack when this number is exhausted.