Signaling when preconditions are not met on a reftest

[frivoal]

No description provided.

[jgraham]

As written the proposal is not to skip the reftest (which would require us to statically determine that it shouldn't be run), but to enable reftests to return the PRECONDITION_FAILED status.

I think that's fine, but it should really be for behaviour that's really optional, or really depends on a precondition (i.e. a test for Y in "you must do X if Y"). Historically this distinction wasn't made clear in testharness.js tests and it caused problems, see https://github.com/web-platform-tests/rfcs/blob/master/rfcs/assert_precondition_rename.md

I will also say that in general I'm not convinced that this extra status has turned out to be as useful as it sounds like it ought to be. For example we can't include tests that might return PRECONDITION_FAILED in Interop because it's not clear how they should be counted. Consider e.g. tests depending on specific codecs that return PRECONDITION_FAILED if the codec isn't supported. In that case if you consider the precondition failing a pass then an implemention with no actual supported codecs looks perfectly interoperable. On the other hand if you consider it a fail then implementations can't actually get 100% without supporting all the optional codecs.

[frivoal]

Is it not possible to report a test as skipped if it hits PRECONDITION_FAILED? In the example of codecs, that would enable you to say that a browser did pass 100% of the tests it did run, but that it didn't run all of them (as they don't apply when some optional codecs are not supported).

[frivoal]

For an extra bit of background, I'm thinking of spec test coverage in addition to interop/regression testing. If I know I have a test suite that covers all my normative statements in a spec, and I want to two if I have at least two implementations of each feature in the spec, I don't want to count "PRECONDITION FAILED" as a pass. But for interop purposes, counting it as a FAIL isn't nice, since the browser did nothing wrong and there's nothing to be fixed.

[jgraham]

Skipped already means something specific (we didn't try to run the test). I'd be opposed to reusing it for "we ran the test, but the test decided it wasn't relevant". I think PRECONDITION_FAILED is as close to the right status as we have.

Allowing different numbers of tests for different browsers also doesn't work well for Interop. You again have the possibility of some browser having 0 tests (e.g. if every test returns PRECONDITION_FAILED), plus it makes it impossible (or at least more difficult) to compute the overall interop score ("the fraction of tests passed by all browsers").

So, again, I think this is broadly fine if you update it to not talk about "skipped" tests but instead to allow a PRECONDITION_FAILED status from reftests. But a risk here is that any tests using this status are (currently and with no current plans to change) ineligible for Interop focus areas.

[frivoal]

I think this is broadly fine if you update it to not talk about "skipped" tests but instead to allow a PRECONDITION_FAILED status from reftests.

Done. Thanks for the feedback.

[Ms2ger]

This seems reasonable; the issue with interpreting this result already exists in th.js tests, so shouldn't block reusing it for reftests, IMO

[Ms2ger]

This actually seems like a benefit to me :)

[Ms2ger]

Risk: would this work reliably / consistently if these two calls were reversed? Do we already define the exact timing of test finishing vs removing the class? (E.g., sync, on a (micro)task, by polling)