We are intentionally breaking the system.

That is the essence of the CAP theorem.

You cannot have perfect consistency, availability, and partition tolerance simultaneously.

You have to pick your poison.

By requiring follower acknowledgments, you also introduce multiple network hops into your critical path.You're waiting for the leader to talk to the follower and the follower to talk to the leader before the producer gets unblocked.

This adds maybe five to fifteen milliseconds of produce latency per batch.

The monolith is completely unaware that Debezium exists, aside from a negligible CPU overhead required to maintain the logical replication slot.

Your newly minted microservices simply attach as consumers to those Kafka topics.