Kafka Producers enforce durability across replicas by setting acks=all (default since v3.0).
As enforcing this guarantee requires waiting for replicas to sync, this increases latency; and reducing it tends to give the impression that latency gets reduced overall.
When looking at the end-to-end latency (including production and consumption) this ends up not being the case:
Records are only available for consumption once the replicas have been updated and the leader can increase the high watermark.
So, acks only affect the latency at the producer side.
If the latency that matters to you is end-to-end, then reducing acks might not be the right trade-off.
References
- Tail latency at Scale: https://www.confluent.io/blog/configure-kafka-to-minimize-latency/
- Optimizing Kafka for Latency: https://docs.confluent.io/cloud/current/client-apps/optimizing/latency.html
- Kafka Internal https://developer.confluent.io/#kafka-internals, with requests lifecycle
