FAIRVISOR Docs
⌘K
GitHub fairvisor.com

Benchmarks

All benchmarks are reproducible. Run them on your hardware — see Reproduce below.

Test setup (March 4, 2026 run)

  • Hardware: AWS c7i.2xlarge (8 vCPU, 16 GB RAM)
  • OS: Ubuntu 24.04.3 LTS
  • Fairvisor: latest main from github.com/fairvisor/edge
  • OpenResty: 1.29.2.1 (native, no Docker)
  • Load generator: k6 v0.54.0 (constant-arrival-rate)
  • Profile: 10,000 RPS steady state, 60s run, 10s warmup
  • Topology: single-host benchmark (SUT + k6 on one VM), CPU-isolated with taskset
    • OpenResty/backend: cores 0-3
    • k6: cores 4-7
  • Decision API method: POST /v1/decision with X-Original-Method and X-Original-URI
  • Reverse proxy path: includes enforcement + upstream proxy hop to backend nginx

Latency

Results at 10,000 RPS (steady state):

Percentile Fairvisor (decision only) Fairvisor (reverse proxy) Raw Nginx (no rate limiting)
p50 112μs 241μs 71μs
p90 191μs 376μs 190μs
p99 426μs 822μs 446μs
p99.9 2.99ms 2.98ms 1.61ms

For this run, decision_service reached sub-millisecond p99 (426μs) on single-host setup with CPU pinning.

Throughput

Max sustained throughput (single edge):

Configuration Max RPS
Simple rate limit (1 rule) 110,500
Complex policy (5 rules, JWT parsing, loop detection) 67,600
With token estimation 49,400

Memory (reference sizing)

The following memory profile is retained as a sizing reference for active limiter keys:

Active limit keys Memory usage
1,000 48 MB
10,000 120 MB
100,000 680 MB
1,000,000 5.2 GB

Memory scales roughly linearly with active limit keys because each key keeps a small fixed token-bucket state.

Reproduce

git clone https://github.com/fairvisor/benchmark
cd benchmark
./run-all.sh

Results are published with every release.

MD