A Kafka producer is experiencing high p99 latency. List 5 potential causes and fixes.

Common causes: 1) Metadata fetch: send() can block for the first record sent to the cluster while the producer fetches metadata; a common approach is to warm it at startup by calling partitionsFor(). 2) Batching: linger.ms (default 5 ms since 4.0) makes the producer wait up to that long before sending, so it adds latency to each record, while very small batches mean many requests; tune linger.ms and batch.size together against your latency target. 3) Buffer full: when buffer.memory is exhausted send() blocks for up to max.block.ms; reduce the send rate, raise buffer.memory, or find why the broker is slow. 4) Broker request handler saturation: RequestHandlerAvgIdlePercent is ideally above 0.3, and a lower value means the handler threads are busy, so add broker capacity or reduce load. 5) Follower lag: with acks=all the leader waits for the in-sync replicas (ISR), so a slow follower raises latency; check replica lag and broker disk and network. 6) GC pauses in the producer JVM stall sends; tune the collector or reduce allocation. 7) Network latency between the producer and the brokers: place them close together; delivery.timeout.ms (default 120 s) and request.timeout.ms (default 30 s) bound how long a send can wait. The broker's request total time metric, split into queue, local, remote and response send time, shows where the time goes.

Expected key concepts: metadata fetch, partitionsFor, linger.ms, buffer.memory, max.block.ms, RequestHandlerAvgIdlePercent, follower lag, acks=all, ISR, GC pauses, network latency