🐛 BUG: Only one instance of Wrangler can be started

[penalosa]

Which Cloudflare product(s) does this pertain to?

Wrangler core

What version(s) of the tool(s) are you using?

main

What version of Node are you using?

18

What operating system and version are you using?

macOS

Describe the Bug

Observed behavior

Checkout & build Wrangler locally, and then run npx wrangler dev in workers-sdk/fixtures/service-bindings-app/a and then in workers-sdk/fixtures/service-bindings-app/b. The second command will fail with Address already in use (127.0.0.1:9229)

Expected behavior

Multiple instances of Wrangler to be startable without error.

This first started regressing in #4830, and hasn't yet been published in a Wrangler release.

Please provide a link to a minimal reproduction

N/A

Please provide any relevant error logs

N/A

[petebacondarwin]

This is not reproducible in node 16. I think it is a reoccurence of the old node 18 switching to IPV6 by default problem.

[petebacondarwin]

OK so looking into this I think the problem is this line where we use localhost for the host in the getPort() call:

workers-sdk/packages/wrangler/src/dev.tsx

Line 715 in 48f9085

const getInspectorPort = memoizeGetPort(DEFAULT_INSPECTOR_PORT, "localhost");

I believe that what is happening here is that localhost gets resolve to both 127.0.0.1 and ::1.
The first call to this (by Worker a) checks localhost:9229 and is told it is free (because 127.0.0.1:9229 is free. So it uses that port and starts listening on 127.0.0.1:9229.
The second call to this (by Worker b) checks localhost:9229 and is told it is free (because [::1]:9229 is free). So it uses that port and tries to start listening on 127.0.0.1:9229which of course fails because that is already in use by Worker a.

If you change the line above to use 127.0.0.1 as the host explicitly then the second call (Worker b) gets a completely new port and that seems to solve the problem.

I think this does not fail in node 16 because it does not have support for IPV6 so calling getPort() with localhost only ever checks 127.0.0.1.

cc @Lekensteyn

[petebacondarwin]

An alternative solution is to ask workerd to listen on localhost explicitly, rather than leaving it undefined, which means that workerd is able to pick the other IP address, here:

workers-sdk/packages/wrangler/src/api/startDevWorker/ProxyController.ts

Line 126 in 884e418

unsafeDirectHost: this.latestConfig.dev?.inspector?.hostname,

e.g.

unsafeDirectHost:
	this.latestConfig.dev?.inspector?.hostname ?? "localhost",

But I am worried that we then have two different servers listening on localhost:9229 and it is not clear how to distinguish between them. So I think the explicit use of IPv4 127.0.0.1 across all these internal servers is the best approach.

[penalosa]

Fixed in #4907