Because links are resource metadata, not the resource itself let alone entities represented by a resource. The relationships between resources are independent of the resources themselves.
> If data-attributes have to be parsed out of some json-body, but link-attributes have to be parsed out of the header, that's weird.
It really isn't. In fact, it's the other way around. Link relations depend on the request, not the resource, just like the particular version of the resource (see ETag header) or the date and time at which the origin server believes the resource was last modified (see Last-Modified header).
Considering the petstore example, it would be weird if the pet would include a version/hash or the last time it was modified. A pet is a pet. It has a name, a species, a breed, and an owner. The pet resource is the pet information that you received as a response to the request you've made in a specific moment in time.
Not really. Since it's mother's day, the fact that I'm related to my mother is an attribute of me, not an attribute of the request you make for information about me. I am related to my mom.
In a conventional database, link relations are defined on the tables themselves, but versioning information and last-modified information is normally defined in metadata tables. So I don't see any prior art for this. It just doesn't match how most people normally represent objects.
> and the pet resource is the pet information that you received as a response to the request you've made in a specific moment in time.
And at version X, the pet has a specific owner, defined as a relation on the pet itself. The request provides versioning information, but the owner is keyed on the pet, just like it would be in the server's database.
You are related to your mother, but that relationship is represented through link relations. You are represented by your resource, and your mother is represented through her resource. Where and how those resources are represented or found is an entirely different concern that has absolutely no relation with the relationship between you and your mother.
> In a conventional database, link relations are defined on the tables themselves
Actually, they are not. I mean, there are entity tables (the resources) and then there are the relationship tables (the... relations). Although some techniques involve using non-normalized databases, that does not mean that entities and entity metadata, such as relations, are or should be conflated.
Now, imagine that the database tables weren't fixed properties of the system, which is a basic design issue in resource-oriented architectures. Would it make any sense to hard-code references to other tables if they could change at any point in time, specially if you're operating a system that could cache table rows?
> And at version X, the pet has a specific owner, defined as a relation on the pet itself.
Yes, and that's also the wrong way to go about specifying the relationship between resources. The correct way would be to express the relationship between resources as a link relation. The identity of neither the pet or the owner depend on each other.
Foreign keys are put in entity tables in normalized databases.
> The identity of neither the pet or the owner depend on each other
So then the only thing that belongs on the pet itself is an arbitrary id? Because the identity of the pet doesn't depend on it's age or when it was born, and a timestamp is just a timestamp, so putting "time of birth" could just be a link relationship to another entity.
Same for breed. Location too, that's not a tool for identifying the pet. Heck, even the name is often defined by the pet-owner relationship but not the pet itself.
So we're left with a pet object that has only a uuid, and a set of link relations in the headers.
Why is that better? You've said my way is "wrong", but you haven't actually explained what your way does to improve the situation. I just demonstrated that any arbitrary piece of info can be represented as a link relationship, so the whole thing is arbitrary anyway.
So the simple question, the only question is, as a designer or user of an API, how does sticking some relationships in headers make my life easier? How do I decide which relationships those should be?
You clearly have opinions on how this should be done, but neither you nor the RFC explains what I gain from your way of doing things. So again, I'm not asking what I should do, I'm asking why?