Kafka share groups with share.acquire.mode=record_limit can exhibit severe performance degradation when fewer consumers than partitions are used alongside partition skew. The root cause is round-robin broker fetching combined with fetch.max.wait.ms (default 500ms): when lighter partitions drain, consumers waste up to 500ms per empty fetch, causing consumption rates to collapse from millions of records/s to just ~2000/s. Five benchmark cases illustrate the problem across balanced backlogs, skewed backlogs, batch production workloads, backlogs with producer load, and extreme max.poll.records=1 scenarios. The primary mitigation is to use partition count as the floor for consumer count when using record_limit. Secondary options include reducing fetch.max.wait.ms, increasing max.poll.records, or fixing partition skew.
