A producer batches records to amortize request overhead. batch.size is measured in bytes, while linger.ms bounds an intentional wait for a fuller batch. A larger setting is not automatically faster for every workload, and a latency-sensitive request can be harmed by waiting too long. Delivery configuration remains a separate requirement: removing acknowledgments to improve a benchmark changes the semantics being measured. These tasks request explicit bounded settings while preserving idempotence and acks=all. Inspect the units and the saved producer configuration. The model records these settings but does not claim to benchmark network throughput, compression ratios or real client memory. Measure those properties on a representative deployed workload.
Inspect the topics, partition logs and group offsets before changing the cluster.