Side-by-side benchmarks. Same API.
Real databases. No toy endpoints.

Results (200 Concurrency, 5s per endpoint) 📊

Stack Map 📦

Inside every container the server listens on the canonical port 8080 (baked as PORT in the Dockerfile, PLAN §6 rule 1); the benchmark client maps a dynamic host port. The port below is each server's assigned 2LRFF host port (dev_port in its bench.json, PLAN §6), used by its .env.example and as the code default in the Go/Rust/Kotlin/Zig servers; Node/Deno just dev binds the canonical 8080 unless PORT is set (e.g. via cp .env.example .env).

Folder	Runtime	Framework	Dev port
benchmark	Go 1.27rc1	-	-
servers/go-stdlib	Go 1.27rc1	net/http (stdlib)	21001
servers/go-chi	Go 1.27rc1	Chi 5.3.0	21002
servers/go-gin	Go 1.27rc1	Gin 1.12.0	21003
servers/go-fiber	Go 1.27rc1	Fiber 3.4.0	21004
servers/go-echo	Go 1.27rc1	Echo 4.13.4	21005
servers/ts-express	Node 26.4.0	Express 5.2.1	22001
servers/ts-nestjs	Node 26.4.0	NestJS 11.1.27	22002
servers/ts-fastify	Node 26.4.0	Fastify 5.9.0	22003
servers/ts-honojs	Node 26.4.0	Hono 4.12.27	22005
servers/ts-bun-honojs	Bun 1.3.14	Hono 4.12.27	22105
servers/ts-bun-elysia	Bun 1.3.14	Elysia 1.4.29	22106
servers/ts-deno-oak	Deno 2.9.1	Oak 17.2.0	22204
servers/ts-deno-honojs	Deno 2.9.1	Hono 4.12.27	22205
servers/py-fastapi	Python 3.14.6	FastAPI >=0.128	23001
servers/py-django	Python 3.14.6	Django >=6.0	23002
servers/py-flask	Python 3.14.6	Flask 3.x	23003
servers/rs-axum	Rust 1.96.1	Axum 0.8.9	24001
servers/rs-actix	Rust 1.96.1	Actix Web 4.14.0	24002
servers/kt-ktor	Kotlin 2.3.21	Ktor 3.5.1	25001
servers/kt-spring-boot	Kotlin 2.3.21	Spring Boot 4.1.0	25002
servers/zig	Zig 0.16	http.zig	26001

Pinned dependencies 📌

Deliberate prerelease pins (PLAN §10), exempt from the blanket just update bump; each tracks the newest release of its dist-tag channel via scripts/update.mts, never the stable latest line:

Package	Pinned at	Channel	Why
typescript	7.0.1-rc	rc	TypeScript 7 native tsc (latest is still 6.x)
drizzle-orm	1.0.0-rc.4	rc	drizzle 1.0 RC (latest is still 0.45.x)

Held back: bson is pinned to 7.2.0 for the whole workspace via pnpm-workspace.yamloverrides. bson 7.3.x calls v8.startupSnapshot.isBuildingSnapshot() at import time, which Bun 1.3.14 ships as a throwing stub (NotImplementedError), crashing the server on boot. Since PR #26 mongodb is a single shared dependency (@bench/shared), and pnpm resolves one bson version for every consumer (overrides are graph-path scoped, not per-member), so the Node servers pin to 7.2.0 too — pnpm cannot scope an override to just the Bun members. 7.2.0 sits inside mongodb's own ^7.2.0 range; drop the override once Bun implements the stub.

NestJS dev mode tradeoff: under TS7 nest start --watch is gone, so ts-nestjs's dev compiles with tsc then runs node --watch dist/main.js. That restarts the server on rebuild but does not watch-recompile sources — edit, re-run tsc (or just dev ts-nestjs) to pick up changes.

Quick Start 🚀

just benchmark # Run benchmark (interactive mode)
just benchmark --servers=a,b # Run benchmark for specific servers only
just dev ts-bun-honojs # Start dev server (ts-bun-honojs, ts-express, go-chi, etc.)

First pnpm install over a checkout that predates the pnpm workspace can want to purge the old top-level node_modules. pnpm-workspace.yaml sets confirmModulesPurge: false, so the install proceeds non-interactively (no CI=true or TTY needed) instead of aborting with ERR_PNPM_ABORTED_REMOVE_MODULES_DIR_NO_TTY.

Configuration ⚙️

Variable	Default	Description
ENV	dev	dev enables logger, prod disables it
HOST	0.0.0.0	IP or localhost (mapped to 0.0.0.0)
PORT	See Stack Map	Server port

Benchmark config is JSON-only and lives at config/config.json.

API Surface 🌐

Base Routes

Method	Route	Response
GET	/	{ "hello": "world"}
GET	/health	OK (text/plain)

Params Routes (/params/*)

Method	Route	Description
GET	/params/search	Query q (trim, default none) and limit (safe int, default 10)
GET	/params/url/:val	Returns { "dynamic": "<val>" }
GET	/params/header	Reads X-Custom-Header (trim, default none)
POST	/params/body	Validates JSON object (no array/null), returns { "body": <parsed> }
GET	/params/cookie	Reads cookie foo (trim, default none), sets cookie bar
POST	/params/form	Supports urlencoded/multipart, returns { "name": "<trim>", "age": <int> }
POST	/params/file	Multipart file (max 1MB, text/plain only), returns { "filename", "size", "content" }

Database Routes (/db/{database}/*)

Supported databases: postgres, mongodb, redis, cassandra

Method	Route	Description
GET	/db/{database}/health	Database health check
POST	/db/{database}/users	Create user { "name", "email", "favoriteNumber?" } (201)
GET	/db/{database}/users/{id}	Get user by ID (200 or 404)
PATCH	/db/{database}/users/{id}	Update user fields (200 or 404)
DELETE	/db/{database}/users/{id}	Delete user by ID (200 or 404)
DELETE	/db/{database}/users	Delete all users (200)
DELETE	/db/{database}/reset	Reset database (200)

Error Responses ⚠️

All errors return JSON { "error": "<message>" }.

Status	Messages
400	invalid JSON body, invalid form data, invalid multipart form data, file not found in form data
404	not found
413	file size exceeds limit
415	only text/plain files are allowed, file does not look like plain text
500	internal error

Databases 🗄️

All servers connect to all 4 databases with the same user schema.

Database	ID Type
PostgreSQL	UUID v7
MongoDB	ObjectId (native)
Redis	UUID v7
Cassandra	UUID v7

User schema:id, name, email, favoriteNumber (optional)

Grafana 📊

Metrics are exported to a dedicated metrics PostgreSQL (separate from the benchmarked postgres) and visualized in Grafana during benchmarks. Run history is durable: the metrics volume survives stack restarts, so runs can be compared across weeks.

Service	URL	Username	Password
Grafana	http://localhost:20090	admin	123456
metrics-postgres	postgres://localhost:20091	benchmark	benchmark

Exported Metrics

Aggregate tables carry exact numbers computed from the full in-memory result set before any sampling; request_events is sampled raw drilldown only. Canonical queries live in infra/grafana/queries/.

Table	Contents	Key columns
runs	one row per run: sample rate + write accounting	run_id, started_at, finished_at
endpoint_stats	exact per-endpoint aggregates (rps, avg/p50/p95/p99/p99.9, open-mode fields)	run_id, server, endpoint, method, source
sequence_stats	exact per-sequence aggregates (full-sequence durations)	run_id, server, sequence_id, database
resource_samples	container memory/CPU min/avg/max per server run	run_id, server, source, database (DB only)
request_events	sampled raw request/sequence events with real timestamps	run_id, server, endpoint, source, database

Development 🛠️

Per-Target Commands

All verification commands accept an optional target (default: all).

just typecheck go-chi # Type check only go-chi
just fmt ts-bun-honojs # Format only ts-bun-honojs
just lint py-fastapi # Lint only py-fastapi
just verify ts-express # Full verification for ts-express

Valid targets: ts-express, ts-nestjs, ts-fastify, ts-deno-oak, ts-bun-honojs, ts-bun-elysia, go-chi, go-gin, go-fiber, py-fastapi, zig, benchmark, root (the servers/*/bench.jsonname fields, plus benchmark and root)

Full Command Reference

just benchmark # Run benchmark (interactive mode)
just install # Install all dependencies
just typecheck # Type/compile check all projects
just fmt # Format all code
just lint # Lint all code
just verify # Non-mutating gate: typecheck -> format-check -> lint
just images # Build all Docker images
just clean # Remove build artifacts and node_modules
just remove-images # Remove Docker images
just grafana-up # Start Grafana/metrics-postgres stack
just grafana-down # Stop Grafana/metrics-postgres stack (volumes are kept)
just db-up # Start database stack
just db-down # Stop database stack