@Jordan474 I'm reverting this for now, since apparently 30 minutes of CI time is spent on getfqdn. Any clue where that overhead is coming from? Feel free to resubmit a PR.

We call socket.getfqdn() when writing the database, as well. It's used in the tempfile name. I wonder if that is also causing a performance penalty on macOS.

Indeed, there are several reports of issues with fqdn on macOS.

There seems to be a user-side workaround: scutil --set HostName $(scutil --get LocalHostName). However, I did not find a definitive solution suitable for Spack yet.

I'm wondering if the hostname -f is affected too ? (its implementation differs from socket.fqdn). I'll look into github macos runners to test this.

Note that even on Linux, I probably should have memoized fqdn, because fqdn could query the DNS if /etc/hosts does not contain the non-fqdn hostname.

I tested a few options on github latest runners:

"fast" means between 1 to 10 ms. I removed python cold start and overheads before comparing the results.

(1): on the windows runner, hostname -f fails with unknown option -- f

(2): On macos, the full hostname is just hostname plus .local (Mac-1727169450078 -> Mac-1727169450078.local). On ubuntu and windows runners, full hostnames have way more components.

EDIT: To clarify, on macos, socket.gethostname() == socket.getfqdn() == scutil --get HostName == Mac-123.local, but scutil --get LocalHostName == Mac-123.

The .local name seems special enough, so maybe the simplest workaround would be to handle it explicitly:

@memoized
def _getfqdn():
    name = socket.gethostname()
    # Local hostname, avoid slow getfqdn on MacOS
    if name.endswith(".local"):
        return name
    return socket.getfqdn(name)

I don't think it's worth it implementing a different subprocess for each platform.