acks, Retries and Durability

Key points

• acks=0 is fire and forget. acks=1 waits for the leader only, so the write dies with the leader.
• acks=all waits for every in-sync replica. It has been the client default since Kafka 3.0, and it is the one you want.
• acks=all on its own is not enough. Pair it with min.insync.replicas=2 on a topic with 3 replicas.
• Without that, an ISR shrunk to one replica still satisfies acks=all. With it, the write fails loudly instead.
• Retries are effectively infinite, bounded by delivery.timeout.ms (2 minutes). When that expires it is a real failure — handle it.
• linger.ms trades a few milliseconds of latency for much bigger batches. On a busy topic that is nearly free throughput.

Example

# ---- producer ----
acks=all
enable.idempotence=true
delivery.timeout.ms=120000
request.timeout.ms=30000
max.in.flight.requests.per.connection=5
compression.type=lz4
linger.ms=10
batch.size=65536

# ---- topic side (kafka-configs.sh --alter --entity-type topics) ----
# min.insync.replicas=2   with replication.factor=3

acks=all plus min.insync.replicas=2 is the durable baseline. Anything weaker is a deliberate trade.