In an earlier PR, I added calls to `disable_socket()` from
`pytest_socket` where I thought they were needed to prevent some tests
from accessing the network, in case they weren't monkeypatched properly.
Today, I discovered that `disable_socket()` disables sockets globally
for all tests, which means that the tests that use remote data cannot
run if they are executed after another test calls `disable_socket()`.
This change calls `disable_socket()` once from `conftest.py`, so that
no tests are allowed to use network data unless they are marked as ok
to use the network, with `@pytest.mark.enable_socket`. See example of
usage in `tests/test_connection.py`.
Changed return code for unexpected exceptions:
This allows us to write tests that can discover whether an unexpected
exception occurred just by checking the return code, rather than reading
stderr. This will allow us to write less friable tests that don't break
every time some insignificant output details change.
This change catches exceptions derived from Exception and
KeyboardInterrupt raised by `installer`, while run by multiple
processes, and propagates them back to earlier stack entries. This will
prevent any OSError and BrokenPipe exceptions that would otherwise be
raised when one process has an exception while the other processes are
still running.
This also handles the MemoryError exception we saw in #416, and offers
some suggestions for solving the issue.
The test test_cli_unknown_version, in `tests/test_connection.py`, is
never run in CI because it is marked as requiring remote data. When I
run `pytest` locally, this test always fails because the version string
"5.12" cannot be parsed into a valid `semantic_version.Version` object.
This change modifies that version string to "5.16.0", so that it will
make a valid Version object that will never be a valid version of Qt.
This also modifies the assertion code in test_cli_unknown_version.
On my machine, the new logger prints directly to `stderr` and not
`stdout` when run via `pytest`, and the `aqt - WARNING` and `aqt -ERROR`
prefixes are omitted. The new assertion code uses a regex to pick out
the required parts of the output and ignore anything that might not show
up.
This change also accounts for what happens during a redirect. In my
testing, I noticed that redirects would occasionally insert some
additional lines of output regarding a redirect. This change describes
this output in a comment, and swallows up the additional lines using a
Kleene star.
