Measured latency
Measured on October 27, 2026 from a c7gn.2xlarge in AWSus-east-1, against a disk in the same region.
A client cache miss is served from the file server’s cache. Data that has to come from the data
source takes longer, typically 10 to 30 ms for Amazon S3. An
O_SYNC write returns once the file server has made it
durable. Checkpoint and branch times don’t depend on disk size.
Throughput
The production figures are the highest one-minute averages recorded in production. They are
observations, not caps. See Limits for every other limit.
Compared with Amazon EFS
In a side-by-side benchmark on one host, Archil ran git, npm, and SQLite transaction workloads about 5x to 190x faster than Amazon EFS. EFS was faster for sequential writes, 4 KiB random writes, and one large SQLite insert. The page lists every result, including the ones Archil lost, and the method behind it.Caching
If your application requests data that’s not in the cache, Archil fetches it from the backing bucket and caches it for future requests. Archil keeps this data in cache for up to 1 hour after the last read or write. Archil may read up to 10 MiB more than requested from the backing bucket.How we measure
- Public interfaces only. The tests use the public API, the released
archilCLI, and standard tools such asfio,curl,git, andnpm. No internal access or special builds. - Same region. The client runs in the same cloud region as the disk. A cross-region round trip is larger than most of these figures, so run your test where your workload runs.
- Percentiles, not best runs. Latency rows report p50 and p95 over every sample. Benchmark rows report the median of 3 runs.
- Losses stay in. When another system wins a test, the result stays in the table.
- Dated and versioned. Each result lists the date, host, and tool versions, so you can tell when a number is stale.