core: Define how to apply subschemas to child instances

[handrews]

In #157 @awwright noted that:

Technically, "properties" and "items" are just keywords for validation. All they do is return true or false.
People have been reading in that we can recursively apply the semantics, and that's fine too. Maybe we should make that more explicit.

Without ways to declare which subschemas apply to which child instances, JSON Schema as a media type is very limited.

In theory, it should be possible to use some JSON Schema vocabulary without validation, which is not possible at the moment. We don't have such a vocabulary right now, but to demonstrate the idea, here is a hyper-schema that imposes no validation constraints:

{
    "links": [
        {
            "rel": "self",
            "href": "/foos"
        }
    ],
    "items": {
        "links": [
            {
                "rel": "item",
                "href": "/foos/{id}"
            }
        ]
    }
}

This schema uses items from the validation vocabulary, but only to apply the "item" link to each element of an array instance. The URI template implies that the elements of the array are objects with at least an "id" property, but this schema does not validate that at all. Obviously, this is not the most realistic schema, but we don't have other vocabularies right now and I can't come up with one off the top of my head.

---

I recently rewrote the items, additionalItems, properties, patternProperties, and additionalProperties to make it clear that they do not directly validate the instance, but instead control which subschemas are applied to which child instances. While I expressed that in terms of validation, reading back over the language shows that it is not necessary.

We could move those keywords into the core specification and simply reference their impact on validation in the validation specification. There would not be any changes to the current functionality, but it would allow for much more flexible definitions of new vocabularies.

This is all a bit abstract, and probably hard to grasp from just eyeballing the current validation spec wording. So I will write up a trial PR to demonstrate how that would look.

[deuberger]

I second this issue. Reading over the latest schema at http://json-schema.org/latest/json-schema-validation.html is confusing to me now compared to what was there in early October. It almost seems like it would be valid to have invalid child instances since the specification doesn't require that child instances be validated it seems. Maybe I'm misunderstanding something, but it's definitely confusing to me one way or another.

[handrews]

@deuberger I rewrote the sections for items, additionalItems, properties, patternProperties, and additionalProperties within the validation spec in PR #111 , could you take a look and see if that is easier to understand? You can read it in the diff or check out the current master here and build the HTML form of the docs if you install xml2rfc.

[deuberger]

@handrews thanks. I took a look at what's in master as of commit e784d5d. Comments:

• The "items" section is still a little confusing. In particular, it's not very clear that the keyword only effects child instances of arrays. "This keyword controls child instance validation," makes it sound like it applies to any child instance (that of arrays or objects). The next paragraph, explicitly mentions arrays, but the last paragraph is only implies it by saying "elements".
• "additionalItems" has a similar issue in that nothing explicitly says it applies only to arrays even though it can be inferred.
• Maybe I'm missing something, but it seems that setting "additionalItems" to false is no longer allowed? So how does one disallow additional items? By setting maxItems?
• "properties", "patternProperties", and "additionalProperties" also have the same issue in that nothing explicitly says they apply only to child instances of objects.
• The last paragraph in the "properties" section is a bit confusing. It says "the instance value" needs to validate against the corresponding schema, but its actually the value at that name within the instance (though maybe there's something better to call it). Maybe it should be "child instance for that name" like "patternProperties" uses.
• "properties" and "patternProperties" use the term "empty object" instead of "empty schema" like elsewhere
• It seems like the language for default values (e.g. "empty schema" or 0) varies from one keyword to the next. It might make sense to normalize all of them.

In general, it does seem that this version is much clearer with respect to how child instances are validated. Thanks!

[handrews]

@deuberger thanks! I've submitted PR #193 incorporating most of this. Note also the "Validation of primitive types and child values" section setting up the general context.

For the stuff I did not change:

Maybe I'm missing something, but it seems that setting "additionalItems" to false is no longer allowed? So how does one disallow additional items? By setting maxItems?

booleans are now allowed anywhere a schema is allowed- this is covered in the core spec rather than validation. In addition to being inconsistent, the old approach made it seem like setting "additionalProperties" or "additionalItems" to false had some sort of special effect. It does not, it is simply that false always fails validation, so applying a schema of false to every property or item not directly specified has the effect of preventing any such property or item from being defined, as any defined value will fail validation.

It seems like the language for default values (e.g. "empty schema" or 0) varies from one keyword to the next. It might make sense to normalize all of them.

All five of the child validation-controlling keywords now say "empty schema" (not "empty object"). I'm not sure where you were seeing the 0. Keywords with numeric values take a zero default in some cases, but those must be 0 while schema keywords must take an object (or a boolean).

[deuberger]

@handrews: changes look great. Your explanation about false for a schema value makes sense.

Regarding my comment:

It seems like the language for default values (e.g. "empty schema" or 0) varies from one keyword to the next. It might make sense to normalize all of them.

I meant that in reference to all keywords, not just those related to arrays and schemas. Many of them have a line about the default value and the exact language changes slightly from one to the next. It's not a huge deal because I didn't notice any inaccuracies. I just meant that it might make sense to normalize as much as possible for consistency sake.

Thanks for the improvements!

[handrews]

@deuberger

I just meant that it might make sense to normalize as much as possible for consistency sake.

See #171 for work on that.