Benchmarks
Benchmarks Overview
Benchmark-driven development: RandRW, metadata, scaling, object backends.
BrewFS treats RandRW, metadata-heavy access, multi-agent scaling, and object-backend performance as first-class benchmark scenarios — measured under published configurations, never as universal claims.
Scenarios
- RandRW — the primary benchmark; the July 24, 2026 Redis + RustFS snapshot records 92% faster fully drained mixed random I/O versus JuiceFS, with BrewFS leading 15 of 24 matched rows.
- Metadata-heavy — readdir / stat / lookup / rename / getattr.
- Multi-agent Scaling — 1 → 16 concurrent agents.
- Object backend — LocalFS, MinIO, RustFS, Ceph RGW, AWS S3.
Method and correctness gates
- Benchmark Methodology — environment, fio profiles, cache and mount boundaries, reproducible runners, and result interpretation.
- POSIX Testing — filesystem conformance gates that must be re-run alongside hot-path performance changes.
Principles
- Reproducible. Configurations and commit hashes are published.
- Transparent. Hardware, software versions, and mount options included.
- Workload-specific. Each claim is tied to a defined workload.
- Configuration-aware. Cache state and backend are always reported.
- No misleading universal claims. We never claim BrewFS is faster everywhere.