Test appending to file

[davidbrochart]

Just copying a test here from jupyter-server/jupyter_server#783, where we see it failing on MacOS.

[codecov]

Codecov Report

Merging #118 (a087347) into main (c35fa79) will not change coverage.
The diff coverage is n/a.

@@            Coverage Diff            @@
##              main      #118   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files            6         6           
  Lines          358       358           
  Branches        73        73           
=========================================
  Hits           358       358           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update c35fa79...a087347. Read the comment docs.

[samuelcolvin]

I'm afraid this is not going to work.

MacOS filesystem notifications can be significantly delayed, sometimes by seconds. You can also get an "added" or "modified" when you delete a file

The unit tests in test_rust_notify.py have to do a lot to work around this - e.g. instead of using the tmp_path fixture, we have to have a pre-existing path which we reset after tests - otherwise the creation of the directory from tmp_path is received when the watcher starts.

You can see this by enabling debug and running watchfiles on MacOS - events are significantly confused.

[samuelcolvin]

We already have some logic to try and clean events which is for macos, but it can't cover all cases.

watchfiles/src/lib.rs

Lines 67 to 76 in c35fa79

	// these events sometimes happen when creating files and deleting them, hence these checks
	let changes = changes_clone.lock().unwrap();
	if changes.contains(&(CHANGE_DELETED, path.clone()))
	|| changes.contains(&(CHANGE_ADDED, path.clone()))
	{
	// file was already deleted or file was added in this batch, ignore this event
	return;
	} else {
	CHANGE_MODIFIED
	}

[davidbrochart]

Thanks for the feedback, I think I'll just simplify tests in jupyter-server for now. We don't want to test watchfiles again there. If the PR is merged we'll just trust whatever watchfiles is doing.

[samuelcolvin]

I agree that's the best way. You could even mock the rust code like we do in many of these tests.

[samuelcolvin]

#113 will also help a lot with tests - avoiding tests hanging if no change is detected.

I implemented that partly to help with the uvicorn tests for watchfiles, see encode/uvicorn#1437.

[davidbrochart]

Great, thanks!