Tentative testdriver methods

[jcsteh]

Rendered.
CC @zcorpan, @cookiecrook.

[jcscottiii]

I agree that the tests for these tentative methods should be marked with tentative. But I think we should treat these tentative methods similarly to how we treat tentative tests for the sake of consistency. Given tentative tests do not require consensus currently, I would remove this requirement from tentative methods.

[gsnedders]

I'd argue that tentative testdriver API pose a greater risk, because removing tests that rely on undocumented API is painful and hard — see #172.

[gsnedders]

Most of my comments below are really "let's make this more explicit", rather than any real objection.

And mostly a question for the Core Team, but I do wonder what standard we thing we're holding things to based on #127: is an explainer enough for a testdriver API, or do we want a spec? Because if an explainer is enough, I'd question how often we'd actually add tentative API, because we probably do want something defining what the behaviour of the endpoint is, even if only at a high-level.

[gsnedders]

Let's be more explicit; we already have plenty of non-tentative things in testdriver which are not in either spec — #127 (final RFC) merely requires a WebDriver endpoint, not that it is in one of those specs.

[jcsteh]

Ah. I had assumed that the existence of a WebDriver endpoint implied it needed to be in the WebDriver spec. So how is WebDriver endpoint defined, given that Gecko uses Marionette for WPT and no other vendors implement this WebDriver endpoint yet?

[gsnedders]

For example, https://webbluetoothcg.github.io/web-bluetooth/#automated-testing defines a whole load of endpoints.

The policy is that it requires a spec for the endpoint, not that that spec be in any specific standard.

[jcsteh]

Thanks. Now I better understand your original question regarding whether an explainer is sufficient. I agree that at least having an explainer makes sense, so if the consensus is that an explainer is acceptable, we probably don't need this RFC. RFC 127 just says "an existing WebDriver endpoint", which isn't explicit about a formal spec, though I guess that could be inferred. So perhaps we still need an RFC making it clear that an explainer is sufficient.

That said, one difference between a spec and an explainer is that an explainer is far more likely to change than a spec. So that might still be one reason to mark these methods as tentative: to indicate that the details are still being ironed out.

[gsnedders]

For things like the Accessibility Investigation, is there any intention for there to be any, even high-level, explainer of what the proposed endpoints are? I'm mostly concerned about the risk of ending up somewhere where it is hard for anyone else to start passing the tests written using the tentative endpoint.

[jcsteh]

There is WICG/aom#203. However, that's an evolving issue, not a document. I could create an explainer document if that's what we need, though.

[gsnedders]

I'd argue that tentative testdriver API pose a greater risk, because removing tests that rely on undocumented API is painful and hard — see #172.

[gsnedders]

I wonder if we want there to be a tentative object, or whether we just want to prefix everything like this?

[jcsteh]

I wonder if having a tentative object would make it trickier to "graduate" things from tentative once the spec is final (and my thinking is that we want to incentivise finalising tentative things as soon as is feasible). With prefixing, you can basically just search/replace once the spec is final. With an object, things will need to be moved around. That's obviously not a show stopper if an object is preferred, just something I thought worth considering.

[Ms2ger]

I have no fundamental concerns with this, assuming we'd use this for experimentation, not things that stick around indefinitely.

[cookiecrook]

Perhaps the initial cross-engine prototype of the property bag implementation should include only the properties already exposed by other WebDriver methods, like computedrole and computedlabel.

[jcsteh]

This would make it easier to standardise on a first pass, for sure. However, that first pass would have no practical benefit, since we can already get those properties using the existing methods. The problem I see with this is that we're less likely to spot problems because the implementation is thus entirely "theoretical": we won't spot the kinds of problems we're likely to see in testing new properties, which is the entire point of this work. In contrast, the current prototype supports a few simple ARIA properties and parent/children, which allows us to test the kinds of things we really want to test with this work and thus spot problems we might hit as we move forward.

[jcscottiii]

We discussed this today during the RFC meeting. I'm still opposed to the wording in Detail 1. But the rest of the RFC looks fine. And in the absence of better alternatives for that one point, I will approve this.