Work out kafka partition count instantly with clear inputs, formula shown and shareable results.
Partition count is the larger of two requirements: enough partitions to carry the throughput at per-partition limits, and at least as many as consumer instances since a partition is consumed by exactly one member of a group. Replication multiplies the write load the brokers must sustain, so a factor of three triples disk and network write volume.
Partition count
throughput partitions = ceil(target / per-partition throughput); partitions = max(that, consumer instances); broker writes = target x replication factor
Kafka only supports increasing partition count, and increasing it changes key-to-partition mapping, which breaks ordering guarantees for existing keys. Over-provision modestly rather than under-provisioning.
Yes. Each partition costs file handles, memory for buffers and leader-election work, and end-to-end latency rises. Tens of thousands per cluster is the practical ceiling.