Fix PermissionError for inaccessible database path

[JonathanPlasse]

• Fix PermissionError when Hypothesis database dir path is not accessible #3764

[JonathanPlasse]

Should I move the permission database tests to cover to avoid the coverage failure?

[Zac-HD]

Tests look good, let's move them to cover/ so that we don't need the pragma. Implementation also looks good though I'd put it into the existing function - see below.

[Zac-HD]

Nice work! I've cherry-picked in our weekly dependency update, since you'd already grabbed the new CPython alpha 🙂

[Zac-HD]

Looks like I underestimated how tricky the edge cases here would be - I think I've got it now; the idea is that we error out if the user has explicitly asked for a directory-based DB, but do the fallback logic if the type of DB was never specific.

[JonathanPlasse]

Looks like I underestimated how tricky the edge cases here would be - I think I've got it now; the idea is that we error out if the user has explicitly asked for a directory-based DB, but do the fallback logic if the type of DB was never specific.

Yes, that is the idea. Also, if an explicit path is given instead of not_set with ExampleDatabase it will also fall back to the in-memory database.

[Zac-HD]

Also, if an explicit path is given instead of not_set with ExampleDatabase it will also fall back to the in-memory database.

I don't want to do that - if the user has given us a path, we should either use it or error. Falling back to an in-memory DB is equivalent to silently deleting the database at the end of their session, which would be a pretty nasty surprise.

[JonathanPlasse]

I missed the change when you removed _usage_dir next to the ":memory:" condition and uselessly disable the auto-merge to avoid merging it.

[JonathanPlasse]

This PR was interesting. Thank you again for your help.

[Zac-HD]

My pleasure 🙂

It looks like the remaining CI issues are just flaky tests, so I expect we'll merge on retry.

[JonathanPlasse]

What do you usually do when you encounter flaky tests in CI?

[Zac-HD]

Some combination of "hit the retry button" (since I have maintainer permissions to do that), and opening an issue to make the test not flaky. Generally our tests are solid, but the sheer number of changes happening for #3921 have raised the baseline a bit lately 😢