Regular CI failures with Could not bind to 'host:port'

[ysbaddaden]

The socket related specs (TCP, HTTP, ...) regularly fail on CI with a Could not bind to host:port exception.

These don't happen much on Linux, but they happen frequently enough on Windows that they got annoying and a waste of time to figure out it's unimportant and re-running the failed job(s).

[ysbaddaden]

I'm suspicious of the unused_local_port helper method. It starts a TCP server on port 0, grabs the actual port number, closes the TCP server and returns the port number.

Specs then go on to immediately start a server on that port.

I believe closing the socket on the port doesn't necessarily happen fast enough in the system (or asynchronously), so when we try to bind a new socket on the port, that fails.

Whenever possible we should instead start the server on port 0 and use server.port if we need to connect a client to that port, as we do this in a few places already.

If there are cases where we can't do the above, then we might want to introduce a retry mechanism, with some delay, before failing after a few attempts.