Benchmarks

Caddy Snake runs Python in worker subprocesses managed by the plugin over a Unix domain socket, avoiding a separately configured Gunicorn or Uvicorn server and the extra HTTP hop of a traditional reverse-proxy setup. The harness compares Flask (WSGI), FastAPI (ASGI), and ESGI (gevent) on a minimal JSON GET /hello endpoint. The table below matches the committed benchmarks/results.json, produced on a 4 vCPU / 16 GB cloud VM (linux/amd64) via the Docker harness (benchmarks/Dockerfile).

Test configurations

Configuration	Description
Flask + Gunicorn + Caddy	Gunicorn (1 worker, 4 threads) behind Caddy reverse proxy
Flask + Caddy Snake	Flask served directly by Caddy via caddy-snake (process worker)
FastAPI + Uvicorn + Caddy	Uvicorn (1 worker) behind Caddy reverse proxy
FastAPI + Caddy Snake	FastAPI served directly by Caddy via caddy-snake (process worker)
ESGI (gevent) + Caddy reverse proxy	The same caddysnake.py gevent ESGI StreamServer used by workers, listening on a Unix domain socket with Caddy proxying HTTP to that socket
ESGI + Caddy Snake	The ESGI app served via module_esgi / runtime gevent inside the plugin

All configurations return the same JSON body: {"message":"Hello, World!"}.

Results

Configuration	Requests/sec	Avg Latency (ms)	P99 Latency (ms)
Flask + Gunicorn + Caddy	3,052	32.70	37.68
Flask + Caddy Snake	4,878	20.49	28.25
FastAPI + Uvicorn + Caddy	11,502	8.70	91.46
FastAPI + Caddy Snake	17,423	5.72	8.85
ESGI (gevent) + Caddy reverse proxy	29,193	3.43	9.97
ESGI + Caddy Snake	34,077	2.95	8.10

Highlights on this hardware:

• Flask: Caddy Snake serves ~60% more requests/sec than Gunicorn behind Caddy, with ~25% lower P99 latency.
• FastAPI: Caddy Snake serves ~51% more requests/sec than Uvicorn behind Caddy. The P99 gap is the most striking: 8.85 ms vs 91.46 ms — the embedded path avoids the upstream TCP hop and Uvicorn's accept-queue jitter under 100-connection load.
• ESGI: the pair is closest because both rows run the same gevent gateway code from caddysnake.py; the embedded wiring is still ~17% faster than proxying HTTP over a Unix socket.

Methodology

• Tool: hey
• Concurrency: 100 connections
• Duration: 10 seconds per test run
• Warmup: 200 requests at 10 concurrency before each configuration
• Repeats: 10 runs per configuration, averages reported
• Platform (published numbers): 4 vCPU / 16 GB cloud VM, linux/amd64
• Python: 3.13 (inside Docker Ubuntu 22.04 image from benchmarks/Dockerfile)
• Go: 1.26
• Workers: Caddy Snake uses one process worker in the benchmark Caddyfiles; Gunicorn uses 1 worker with 4 threads; Uvicorn uses 1 worker; the standalone ESGI line uses one caddysnake.py ESGI server process

Reproduce

Dedicated instance (Scaleway POP2-2C-8G)

For stable numbers on a clean dedicated instance, see benchmarks/scaleway_bench.sh (Scaleway CLI + SSH). Example:

BENCH_RSYNC_LOCAL=1 BENCH_SSH_IDENTITY="$HOME/.ssh/id_ed25519_scw" ./benchmarks/scaleway_bench.sh

Any machine (Docker)

docker build -t caddy-snake-bench -f benchmarks/Dockerfile .
docker run --rm -v $(pwd)/benchmarks:/workspace/benchmarks caddy-snake-bench

Results go to benchmarks/results.json and benchmarks/benchmark_chart.{png,svg}. Copy the SVG into the docs site when publishing updated figures:

cp benchmarks/benchmark_chart.svg docs/static/img/benchmark_chart.svg

What the benchmark does

1. Builds Caddy with the caddy-snake plugin from source
2. Creates a Python 3.13 virtual environment with Flask, FastAPI, Gunicorn, Uvicorn, uvloop, and gevent (for ESGI)
3. For each configuration: start servers, wait for GET /hello, warm up, then benchmark with hey (100 concurrent, 10 s) for 10 runs; record requests/sec, average latency, and P99
4. Writes results.json and generates the comparison chart