test_socketserver.test_UnixDatagramServer() times out on GHA iOS CI

[vstinner]

Bug report

Bug description:

Example: https://github.com/python/cpython/actions/runs/18868398194/job/53840759567?pr=140558

   ======================================================================
   ERROR: test_UnixDatagramServer (test.test_socketserver.SocketServerTest.test_UnixDatagramServer)
   ----------------------------------------------------------------------
   Traceback (most recent call last):
     File "/Users/runner/Library/Developer/CoreSimulator/Devices/7EDFF154-B6AB-44D3-BA15-EBE53F1F1878/data/Containers/Bundle/Application/F7FB9E8B-0A4F-4BDD-A041-3CB5C1B883AE/iOSTestbed.app/python/lib/python3.15/test/test_socketserver.py", line 227, in test_UnixDatagramServer
       self.run_server(socketserver.UnixDatagramServer,
       ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                       socketserver.DatagramRequestHandler,
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                       self.dgram_examine)
                       ^^^^^^^^^^^^^^^^^^^
     File "/Users/runner/Library/Developer/CoreSimulator/Devices/7EDFF154-B6AB-44D3-BA15-EBE53F1F1878/data/Containers/Bundle/Application/F7FB9E8B-0A4F-4BDD-A041-3CB5C1B883AE/iOSTestbed.app/python/lib/python3.15/test/support/threading_helper.py", line 66, in decorator
       return func(*args)
     File "/Users/runner/Library/Developer/CoreSimulator/Devices/7EDFF154-B6AB-44D3-BA15-EBE53F1F1878/data/Containers/Bundle/Application/F7FB9E8B-0A4F-4BDD-A041-3CB5C1B883AE/iOSTestbed.app/python/lib/python3.15/test/test_socketserver.py", line 135, in run_server
       testfunc(svrcls.address_family, addr)
       ~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
     File "/Users/runner/Library/Developer/CoreSimulator/Devices/7EDFF154-B6AB-44D3-BA15-EBE53F1F1878/data/Containers/Bundle/Application/F7FB9E8B-0A4F-4BDD-A041-3CB5C1B883AE/iOSTestbed.app/python/lib/python3.15/test/test_socketserver.py", line 162, in dgram_examine
       buf = data = receive(s, 100)
                    ~~~~~~~^^^^^^^^
     File "/Users/runner/Library/Developer/CoreSimulator/Devices/7EDFF154-B6AB-44D3-BA15-EBE53F1F1878/data/Containers/Bundle/Application/F7FB9E8B-0A4F-4BDD-A041-3CB5C1B883AE/iOSTestbed.app/python/lib/python3.15/test/test_socketserver.py", line 44, in receive
       raise RuntimeError("timed out on %r" % (sock,))
   RuntimeError: timed out on <socket.socket fd=33, family=1, type=2, proto=0, laddr=./test_python_x3rw1qn_.sock>

The function uses test.support.SHORT_TIMEOUT which is 30 seconds by default:

def receive(sock, n, timeout=test.support.SHORT_TIMEOUT):
    r, w, x = _real_select([sock], [], [], timeout)
    if sock in r:
        return sock.recv(n)
    else:       
        raise RuntimeError("timed out on %r" % (sock,))

CPython versions tested on:

CPython main branch

Operating systems tested on:

Other

Linked PRs

• gh-140702: Diagnose iOS datagram timeout bug #140736
• gh-140702: Add test skip for Unix Datagram tests on iOS when on Github Actions #140740
• [3.14] gh-140702: Add test skip for Unix Datagram tests on iOS when on Github Actions (GH-140740) #140742
• gh-140702: Log color and GHA env vars in test.pythoninfo #140747
• [3.14] gh-140702: Log color and GHA env vars in test.pythoninfo (GH-140747) #140753
• [3.13] gh-140702: Log color and GHA env vars in test.pythoninfo (GH-140747) #140754

[vstinner]

cc @hugovk @freakboy3742

[freakboy3742]

I've spent the day trying to diagnose this, without much luck.

I can't reproduce the problem locally; and looking at the buildbot, it has maybe happened once in the last six months.

Although it manifests as a timeout, it doesn't look like it's a situation where bumping the timeout window will work. When the test passes, it does so in <1s. There's clearly something getting stalled with Unix sockets... but I can't even begin to think what that might be.

At this point, a test skip (isolated to GitHub Actions runners) seems the pragmatic solution.

[hugovk]

At this point, a test skip (isolated to GitHub Actions runners) seems the pragmatic solution.

This sounds reasonable, especially as we have coverage of the test on a buildbot.