The identity "Me while I'm visiting nytimes.com" is distinct from the identity "Me while visiting cnn.com".
To me, this means all client-triggered requests while my address bar ends in "nytimes.com" will have a distinct set of cookies that is only drawn from while I'm on nytimes.com.
I prefer a rules-based solution to an ML-based solution. ML is more opaque and it's harder for developers to know when their solution will run afoul of the classifier.
Are there significant compatibility problems this would cause?
Anyway, the reason we moved away from cookie partitioning is that many third party embeds got confused when presented with non-empty cookies, but ones that weren’t the user’s canonical cookie for that site. Bugs were reported to us by embed providers, including Google web properties. This is a bummer because partitioning is elegant and has been WebKit’s go to approach since 2013, when we partitioned all storage besides cookies (LocalStorage, the HTTP cache, etc). A move that, incidentally was opposed by Chrime Engineers back when we shared an engine code base, though they are coming around now.
(I should mention also that partitioning cookies made many social embeds stop working at all until we provided Storage Access API and got embed providers to adopt them, which is why ITP originally had a 24 hour user interaction exception. Fortunately the web has adapted now.)
I should also mention, WebKit ML solution is deterministic. It’s trained offline. And it doesn’t substitute for enforcement actions, it substitutes for a block list of known trackers, which other tracking prevention systems use. One downside of strict rule based systems is that it may allow malicious players to game the system. So there has to be an out for changing the rules or ratcheting up enforcement.
I do still wish the ML stuff weren't there, though. As a developer for a small company that's trying to become a big company (and develops on these boundaries), I fear we'll build something that works one day, but the next will be classified and the rules will change.
I'd rather have a policy that forces big and small sites to follow the same rules. Between storage access and link decoration detection, WebKit seems quite close to being able to detect cross-site tracking as it happens rather than via classifier, and I'd hope that could one day replace the ML component.