feat: dispatch

[ronag]

Makes our internal API public. This is the API which our public API is implemented on top.

There is need for some internal cleanup and refactor post this PR. I've tried to keep the changes as small as possible for easy of review.

I am not 100% satisfied with the API. However, I believe it does fullfil functional and performance requirements. Would appreciate some feedback and maybe ideas. I suspect this API might evolve/improve over time and be less stable than the other ones.

Some problems that might or might not need to be resolved:

• There is now two AsyncResources, the handler object and Request object. Depending on whether the implementation needs to asynchronously propagate async scope.

• It's a bit weird that _onConnect takes a resume function.

• Not sure about naming for _onConnect.

• It's a bit weird that _onData should return null to indicate an aborted message.

Refs: #293

_onUpgrade(statusCode, headers, socker): Void
_onHeaders(statusCode, headers, resume): Void
_onData(chunk): Boolean|Null
_onComplete(trailers): Void
_onErrror(err): Void

[ronag]

@szmarczak

[ronag]

pipeline performance regression is from a previous commit. Haven't bisected yet.

[szmarczak]

The new API looks good IMO.

Instead of returning true / false / null I think it would be clearer to return a constant e.g.

• undici.RESUME
• undici.PAUSE
• undici.ABORT

Just an idea.

[ronag]

Just an idea.

That's kind of nice. Less magical.

[ronag]

Another alternative is something like:

_onConnect(controller)

Where controller has the following methods:

pause()
resume()
abort()
runInAsyncScope()

Which would resolve the async resource issue as well... at the cost of an extra object allocation.

EDIT: the runInAsyncScope won't work since we would need it before _onConnect.

[ronag]

@mcollina I think perf diff is within margin of error now...

[ronag]

master:

undici - pipeline x 7,225 ops/sec ±1.49% (273 runs sampled)
undici - request x 11,760 ops/sec ±0.90% (277 runs sampled)
undici - stream x 12,568 ops/sec ±0.79% (280 runs sampled)
undici - simple x 12,949 ops/sec ±0.59% (278 runs sampled)
undici - noop x 19,087 ops/sec ±1.19% (259 runs sampled)

dispatch2:

undici - pipeline x 7,281 ops/sec ±1.68% (273 runs sampled)
undici - request x 11,700 ops/sec ±0.56% (278 runs sampled)
undici - stream x 12,684 ops/sec ±0.72% (280 runs sampled)
undici - dispatch x 13,446 ops/sec ±0.37% (276 runs sampled)
undici - noop x 19,590 ops/sec ±1.06% (255 runs sampled)

[mcollina]

lgtm