NATS JetStream under the same faults, with and without a sync on every write.

Jepsen showed in December 2025 that JetStream’s default settings can lose writes it has already acknowledged. We reproduced that as a comparison: the same cluster, the same workload and the same mix of faults, a hundred runs with the default and a hundred with sync_interval: always.

Default fsync every 2 min

31of 100 runs lost acknowledged writes

One power loss, replayed: 5 of 5 lose writes

Changed sync_interval: always

0of 100 runs lost an acknowledged write

One power loss, replayed: 0 of 5 lose writes

Each square is one run. A filled square lost a write the cluster had acknowledged. Squares count runs, not their order.

A known issue, used as a test

The behaviour was first published in Jepsen’s analysis of NATS 2.12.1 (8 December 2025) and is tracked upstream as nats-server issue #7564. By default JetStream acknowledges a write straight away and flushes data to disk every two minutes. If the nodes lose power together inside that window, writes the cluster said it had accepted are gone.

We ran it for a different reason. A tool that compares two configurations should be able to show a difference that is already known. If it cannot, its answer on an unknown one is not worth much.

The comparison

Both columns are the same three-node cluster and the same workload, under the same mix of faults. The counts below show that mix: the same kinds of fault in similar numbers, not identical schedules. Only one setting differs.

100 runs per configuration
DefaultAlways sync
Runs that lost acknowledged writes310
Writes acknowledged1,155 of 1,1901,167 of 1,190
Power losses117119
Operating-system crashes6758
Process kills8081
Pauses7480
Network partitions8888
Disk corruptions2329
A run counts as failing if at least one acknowledged write is missing at the end. The fault rows count faults injected across the hundred runs.

The smallest schedule that fails

After the hundred runs, Afterfault cut a failing schedule down to the faults it needs. One is enough: a power loss on the nodes at the same moment. Replayed five times against the default settings, it lost writes five times. Replayed five times with sync_interval: always, it lost none.

The setting

jetstream {
  sync_interval: always
}

It is not free. Jepsen’s report quotes the NATS documentation warning that this setting slows a server to a few hundred messages a second. We did not measure throughput in this comparison, so we have no number of our own for the cost.

What to take from it

  1. Find out what “acknowledged” means in your configuration. For many systems it means accepted in memory, not written to disk. The default is often the fast one.
  2. Test the fault that hits every node at once. Killing one process at a time rarely loses data in a replicated system. Losing power everywhere is the case replication does not cover.
  3. Compare settings under the same faults. Changing a durability setting and watching for trouble tells you little. Running both settings through the same faults and counting gives a number you can show someone.

Bring a system you need to trust.

Request early access Next: listmonk