Use xfail consistently for known failing tests
#867
Conversation
Codecov Report
@@ Coverage Diff @@
## main #867 +/- ##
==========================================
+ Coverage 77.69% 80.63% +2.93%
==========================================
Files 75 75
Lines 4218 4218
Branches 766 766
==========================================
+ Hits 3277 3401 +124
+ Misses 775 651 -124
Partials 166 166
📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more |
| # FIXME: handling is needed for httpx-based fastapi>=0.87.0 | ||
| - fastapi>=0.69.0,<0.87.0 |
There was a problem hiding this comment.
fastapi>=0.87.0 pulls in starlette>=0.21.0, which switched to using httpx instead of requests for its TestClient API (encode/starlette#1376); it's not immediately obvious to me how to support both versions of the API, so this pins us to the older starlette packages for now.
| @@ -332,15 +332,15 @@ def gpu_client(gpu_cluster): | |||
| # client for all computations. otherwise, only connect to a client | |||
| # when specified. | |||
| @pytest.fixture( | |||
| scope="function" if SCHEDULER_ADDR is None else "session", | |||
| scope="function", | |||
There was a problem hiding this comment.
When we use a session-wide client for external cluster testing, it ends up getting blown away with dask.config.refresh() while running test_config.py. Since there wasn't any immediate way to avoid this, I made this change so that we reconnect to the external cluster on each test, so that tests run after test_config.py are still able to use the external cluster as expected.
There was a problem hiding this comment.
I also wonder if test_config can be improved somehow to save the state of configs before the tests start and restore that state at the end of the test.
Either way I didn't see a noticeable change in overall test runtimes with this change so it might be okay to improve later.
The intent of this PR is to make the way we indicate failing tests more consistent, by:
Closes #899