dunx: NestJS-style DI on Bun, and what the layer costs

[petarzarkov]

I wanted NestJS's structure without the 244ms boot or reflect-metadata, so I built dunx on Bun.serve.

Constructor injection, no parameter decorators, no reflect-metadata, no experimentalDecorators. An oxc transform reads the constructor's parameter types at load time and records them under a symbol, so the declaration is just the class:

export class UsersService {
  constructor(private readonly repo: UsersRepository) {}
}

A parameter whose type is erased (an interface, a primitive, a type-only import) is a boot error naming that parameter, rather than a silent undefined at the first call.

Routing is Bun.serve({ routes }). I did not write a router, so path params, per-method dispatch and method-miss 404s are Bun's.

Since @dunx/http is a layer over that exact API, the gap to raw Bun.serve is framework overhead and nothing else:

	plaintext	zod body	db + cache	boot
Bun.serve raw	134k	90k	28k	24ms
dunx	134k	83k	28k	46ms
Elysia	133k	80k	28k	52ms
axum	127k	84k	35k	2ms
Nest + Fastify	37k	20k	13k	252ms

Ryzen 9 5950X, Bun 1.4.2, oha at 64 connections, 5s, median of 5 runs. The load generator shares the box with the subject, so read it as a ranking and not as capacity. axum is ahead on the db scenario and that gap is real.

The number I would want if I were reading this: turn structured logging on and dunx drops to 81k plaintext, 60% of raw Bun.serve. Per-request logging costs more than the framework does. That is true of every subject in the suite, it just usually is not the number anyone publishes.

Harness and raw JSON: https://github.com/petarzarkov/dunx/tree/main/internal/bench
Repo: https://github.com/petarzarkov/dunx