Benchmarks

How much does Basalt's neutral core cost over the bare HTTP server?

Short answer: on Fastify, roughly 10–17% throughput overhead for the container, plugin boot, request context, route metadata and validation pipeline. Meaningful but modest — you're trading it for the whole @basaltkit/* stack. The reproducible harness lives in apps/bench.

Method

The same two routes are defined once with the adapter-agnostic route() from @basaltkit/http and served four ways — three Basalt adapters plus a hand-written Fastify baseline:

• GET /health — trivial handler, measures pure framework overhead
• POST /echo — a Zod-validated body, measures the validation path

// src/routes.ts — one definition, every adapter
export const routes = [
  route({ method: 'GET', url: '/health', async handler() { return { ok: true } } }),
  route({
    method: 'POST', url: '/echo',
    body: z.object({ name: z.string(), n: z.number() }),
    async handler({ body }) { return { hello: body.name, doubled: body.n * 2 } },
  }),
]

Fair comparison is easy to get wrong, so the harness:

• Runs each server in its own process — no sockets, GC, JIT or event-loop state carries over, so the server measured last isn't penalised by the ones before it. (Running everything in one process skews the mean and can even make a wrapper look faster than the server it wraps.)
• Warms up at the measurement concurrency and discards it.
• Runs several iterations and reports the median, with cooldowns and a clean shutdown between servers.
• Reports p50 (median) and p99 latency, never the arithmetic mean — the mean is skewed by a handful of outliers. An errors count flags any untrustworthy run.

Results

50 connections · 10s · median of 3 iterations on Apple Silicon / Node 24. Absolute numbers are machine-specific — read the gaps between rows, not the digits. Zero errors across all runs.

GET /health

server	req/sec	vs #1	p50	p99
plain fastify	59,943	100%	<1ms	1ms
Basalt · fastify	49,884	83%	<1ms	2ms
Basalt · express	32,136	54%	1ms	3ms
Basalt · hono	26,665	44%	1ms	4ms

POST /echo (Zod-validated)

server	req/sec	vs #1	p50	p99
plain fastify	39,114	100%	1ms	5ms
Basalt · fastify	34,474	88%	1ms	5ms
Basalt · express	27,228	70%	1ms	3ms
Basalt · hono	18,987	49%	2ms	5ms

Reading the numbers

• Basalt on Fastify keeps ~83–88% of raw Fastify throughput — a ~10–17% overhead for the container, DI, plugin lifecycle, request context and metadata routing. Plain Fastify is the fastest row, as it should be: Basalt wraps it, so it can never beat it.
• Express and Hono are slower because those runtimes are slower — not because Basalt costs more on them. The Basalt layer is byte-for-byte identical across all three adapters (same route() objects); the gap you see is each adapter's own per-request cost. Choose an adapter for its ecosystem, not for a number Basalt would change.
• Even the slowest row serves ~19k validated requests/second on a laptop — comfortably above what a typical multi-tenant SaaS API needs. Your database and network are the ceiling long before the framework is.

Run it yourself

pnpm --filter @basaltkit/bench bench
# knobs (with defaults):
BENCH_CONNECTIONS=50 BENCH_DURATION=10 BENCH_ITERATIONS=3 BENCH_WARMUP=3 \
  pnpm --filter @basaltkit/bench bench