Hey @benjamingr thank you for the response! I agree that that seems like a solution at first. But I think there's a deeper problem, which is that the further from the initial createServer call the logic is, the harder it gets to have the user do the linking manually. Let me explain...

Your own utility

Say you're writing an HTTP utility yourself. For this example, let's go with setStatus for simply setting a status code. (This is just one example, but there are lots of utilities that run into this problem for handling HTTP edge cases.) You might write it like this:

import statuses from 'statuses'

function setStatus(res, code) {
  if (res.headersSent) return
  this.res.statusCode = code
  if (statuses.empty[code]) {} // redacted: handle empty statuses…
  if (req.httpVersionMajor < 2) res.statusMessage = statuses[code] // oops!
}

Notice how you end up needing req at the last line for plumbing-related reasons. But req isn't something that's really in the user's mental model for setStatus because they are just updating the response—the need for req is an implementation detail.

But oh well, since you control the code you add the res.req = req yourself and update your own utility, and it works.

A library of utilities

Going further, consider now that you're not writing the utility yourself, but using a Lodash-like package for HTTP instead. This is ideal, because writing these is tedious (and error prone) and it would be nice if this was solved ecosystem-wide. To use it:

import { setStatus } from 'httpdash'

// Somewhere in your server...
setStatus(res, 500)

But you run into the same issue. There are three potential solutions that technically solve the problem, but none is great:

1. The library could require users to res.req = req in the original http.createServer. This solution is technically possible, but very leaky! It will be hard to remember to do and result in confusion for first-time users, all to handle an edge case.

2. Or the library could change the signature to setStatus(req, res, code). This also technically works, but now you've introduced a new type of API confusion. Some response-related utilities require just res and some require req, res. To the user there's no real way to tell which, because the req is only ever used for HTTP plumbing, ideally they wouldn't need to care and everything would work.

3. Or the library could bite the bullet and add (req, res, ...) to all of its functions. This again technically solves the problem, but now the API has been complicated in all cases. You get further away from the niceness that monolithic frameworks are able to offer because even simple things like setHeader(res, key, value) become setHeader(req, res, key, value) for no technical reason.

What this boils down to, is that without the reference from res -> req, it becomes very hard for non-monolithic frameworks to offer the clean, easy to understand APIs that the frameworks can.

But it actually gets worse still…

A library using utilities

Finally consider the case where you aren't the user writing the server, and you aren't the library of HTTP utilities either… you're now a separate library that does some sort of HTTP-related behavior—maybe you're a pretty printing library that works across all HTTP frameworks.

You have access to res from when you were initialized. You go to use the cool httpdash package of utilities, but realize that for plumbing reasons you also need to change your API to have the user pass in req as well. Depending on how or where you are initialized this can be a non-trivial change to make!

Because the link (that all frameworks do by default) isn't present in core, it makes it harder for 3rd-hand libraries in the ecosystem to use 2nd-hand libraries like the hypothetical httpdash, and further cements the monopoly of frameworks like Express. People feel like making libraries framework-agnostic is too much work, so they just tightly couple everything to whatever their personal preference for framework is. And the whole ecosystem suffers.

In summary…

Yes, it is technically possible to solve this by the user themselves writing res.req = req, I agree. But the issue is that the entire downstream ecosystem of utilities then requires the user to do this for any smarter utilities to take advantage of it, which results in a catch-22 where it never happens and we never get nice framework-independent utilities.

The link between res and its original req is inherent to the needs of HTTP. This is evidenced by Node core’s internal logic relying on it, and by all of the server frameworks in Node.js having made this link implicit for this exact reason.

Right now these nice HTTP utilities are only available to frameworks because they take control of the entire server-side stack. To make writing this sort of nice utilities in a framework-independent way Node.js needs to make the link in core.

@nodejs/http This could use some comments/feedback/review.

I'm not opposed to this.

@ianstormtaylor are you still interested to continue working on this PR?

@ronag yes! I should be able to get to checking out the HTTP2 compat shim after the holidays to add the same property to it.

@ronag I am still -0.5 as I think this conflates things but not -1 since requests and responses are already very intertwined and the use case sounds reasonable.

Not blocking, a week has passed, as far as I'm concerned this can land.

CI: https://ci.nodejs.org/job/node-test-pull-request/35615/

CI: https://ci.nodejs.org/job/node-test-pull-request/35623/

Landed in fc3f1c3

I guess we missed it, but this should've been flagged as a semver-major change since it breaks code that was previously attaching req to res (because req is a getter, and JavaScript on strict mode doesn't allow assigning to getters). A simplified example of the issue:

'use strict'

class Foo {
}

const foo = new Foo();
foo.req = {};

class Bar {
  get req() {
    return "bar";
  }
}

const bar = new Bar();
bar.req = {} // code will crash here

$ node example.js
bar.req = {}
        ^

TypeError: Cannot set property req of #<Bar> which has only a getter
    at Object.<anonymous> (/dev/fd/11:16:9)
    at Module._compile (node:internal/modules/cjs/loader:1092:14)
    at Object.Module._extensions..js (node:internal/modules/cjs/loader:1121:10)
    at Module.load (node:internal/modules/cjs/loader:972:32)
    at Function.Module._load (node:internal/modules/cjs/loader:813:14)
    at Function.executeUserEntryPoint [as runMain] (node:internal/modules/run_main:76:12)
    at node:internal/main/run_main_module:17:47

An example of this in the wild is broken support for http2 on restify with Node.js v15: https://github.com/restify/node-restify/runs/2081960519?check_suite_focus=true#step:5:242

I don't know what we should do here: reverting is also a breaking change. Maybe we should just add a note on the changelog that this was an accidental breaking change?

@aduh95 good call