[cbrnr]

Fifth PR of the tooling series. It moves the test suite to plain pytest and makes tests that depend on PhysioNet independent from the local checks, so a slow or unreachable PhysioNet no longer fails (or hangs) CI. It consists of five commits (the last two were added after CI showed 502 Bad Gateway errors, and after a review comment about hangs):

1. Convert tests from unittest to pytest. Test classes are plain pytest classes, self.assert*() calls are plain asserts (pytest.raises for assertRaises), setUpClass() is replaced by a class-scoped temp_path fixture in conftest.py (tmp_path for the CSV conversion test, which used to write to test_output/), and the unittest.main() footers are gone. Nothing else changes: the same 82 tests are collected with the same IDs. (unittest.mock is still used in test_processing.py.)

2. Run local and PhysioNet checks as separate tests. See below.

3. Format section comments in test_record.py consistently. The four section banners had different widths and spacing.

4. Skip network tests on HTTP server errors, too. PhysioNet regularly answers with 502 Bad Gateway, which wfdb raises as NetFileError. The skip hook now also treats a NetFileError with a 5xx status or 429 as a network error (404 and 403 still fail).

5. Add timeouts to requests for remote files. This is the only change to library code. wfdb/io/_url.py did not set a timeout, so a server that accepts the connection but does not respond blocked forever (scripts hung, and CI jobs waited until the job timeout). Requests now time out after 10 seconds to connect and 60 seconds to wait for data (not for the whole transfer, so slow but steady downloads still work), and a requests.exceptions.Timeout is raised, which the hook above reports as a skip for tests marked network. The constant is wfdb.io._url.TIMEOUT. A new test uses a dummy server with a delayed response. There is also a changelog entry.

PhysioNet tests

Most tests that use PhysioNet first check the local sample data and then compare it with the same data streamed from PhysioNet. If PhysioNet is slow or down, the whole test failed (or hung), including the local checks.

• 15 such tests now take a source fixture that is parametrized with local and physionet, e.g. test_1a[local] and test_1a[physionet]. The local variant never uses the network. The sequence of operations within each test is unchanged (the PhysioNet read and its comparison are just conditional).
• The physionet variants and the two dl_database tests are marked with network (registered in pyproject.toml).
• A small hook in tests/conftest.py reports a network test as skipped (with the reason) if it fails with a connection error, a timeout or an interrupted transfer. Failed assertions, and the same errors in tests that are not marked network, still fail. The skip reasons are shown in the summary (-r s), so an outage is visible.
• pytest -m "not network" runs only the tests that work offline (documented in DEVELOPING.md). 80 of the 97 tests work offline, compared to 65 of 82 before.

Testing

• Without access to PhysioNet, the suite gives 80 passed and 17 skipped on Python 3.11 with the lowest allowed dependency versions and on 3.15 with the latest ones (before: 65 passed and 17 failed).
• I ran the 15 physionet variants against the real PhysioNet, and they all pass. I did not run the two dl_database tests (about 27 MB each).
• I checked the skip hook with a connection error and a timeout in marked tests (skipped, also with -n 2), and with a failed assertion in a marked test and a connection error in an unmarked test (both fail).
• ruff check and ruff format --check pass.

[tompollard]

Thanks, this looks good! One case the hook can't cover yet is a hang: wfdb doesn't set a timeout on its requests (session.request(...) in wfdb/io/_url.py), so if PhysioNet accepts the connection but doesn't respond, a physionet test waits until the 45-minute job timeout instead of being skipped.

[tompollard]

Looks good to me, thanks again!

[cbrnr]

Now let's wait for the tests to come back green. This time they should!