feat: spec compliant early hints

[climba03003]

Here is the implementation that should fit the spec.

1. We should not restrict to only Link header
2. We do not care about the result of Early Hints
3. Use undici to test

In this PR, I provide two function which is writeEarlyHints and writeEarlyHintsLink
writeEarlyHints can be used to write any header.
writeEarlyHintsLink is used to write Link header like previous version.

Checklist

• run npm run test and npm run benchmark
• tests and/or benchmarks are included
• documentation is changed or added
• commit message and code follows the Developer's Certification of Origin
  and the Code of conduct

[climba03003]

Do anyone have time to see why tap is not proceed to next test in node@14 and node@16?

[Uzlopak]

It seems that teardown is not called when the tests are finished because they still run because undici is still connected or so. So you have to explicitly close the client

[climba03003]

It seems that teardown is not called when the tests are finished because they still run because undici is still connected or so. So you have to explicitly close the client

Strange enough, I though it would be keepAliveTimeout issue and use setGlobalDispatcher. It doesn't work and it is actually the keepAliveTimeout issue which require to set in every Client.

No Idea, why undici test do not face the same problem.

[Uzlopak]

I have just two remarks:

1. I would like it more to call it writeEarlyHintsLinks. Or else people may think you can only pass one Link to that method. I actually think we should also add a segment regarding browser implementations and hint that google chrome uses only the first occurence of 103 and link to the chrome documentation. So maybe rename it to indicate it can handle multiple calls
2. If I do autocannon against the example, i get crazy high number of request/s and a very low number of 103 statuscodes. If I use curl while running autocannon i get always 103. So I think that it is working properly. But I have some doubts... it is so confusing...

[mcollina]

lgtm

[Uzlopak]

Strange. in node 14 we never get socket === null?

[Uzlopak]

LGTM

[Uzlopak]

I patched and pushed example.js to run my tests

[Uzlopak]

I opened an issue in autocannon regarding the wrong stats I get from early hints
mcollina/autocannon#464

I actually would have loved to have a stress test, where i would expect 50% 103 and 50% 200 status to have the confidence that it works properly.

But my manual testing was successful so I am confident that this plugin is working reliable.

[Eomm]

I would merge it and we could wait a sec before realising it.

Is ok for you @Uzlopak ?

[Uzlopak]

@Eomm

Of course ;)

[mcollina]

lgtm