Incorrect SchemaError for enum schema with unique elements

[Zac-HD]

# Draft7 for convenience; the same applies to 3, 4, and 6.
# We can have a unique list of things that compare equal, e.g. this works:
jsonschema.Draft7Validator.check_schema({'enum': [False, 0.0]})

# But wrap them in a one-element list, and we get a `SchemaError` instead!
jsonschema.Draft7Validator.check_schema({'enum': [[False], [0.0]]})

There are IMO two bugs here:

1. JSON values are being tested for equality using Python object semantics, rather than comparing them as JSON (e.g. via a canonical encoding).

2. The json-schema specification says

    The value of this keyword MUST be an array.
    This array SHOULD have at least one element.
    Elements in the array SHOULD be unique.
   

   Validation therefore ought to allow schema authors to use empty or non-unique arrays, because the RFC 2119 definition of SHOULD explicitly permits ignoring such clauses after considering the circumstances.

[Zac-HD]

I also just want to say thanks for the jsonschema library - the interface is easy to use, and the fantastic error messages when I misused or misunderstood something have saved so much time debugging.

I stumbled over this when a build failed for my hypothesis-jsonschema library, which generates arbitrary instances matching a schema for testing. In this case it was also generating arbitrary schemas!

[Julian]

Woohoo, thanks for writing the hypothesis thing, been on my list for awhile but never gotten to it (oddly got to it for protobuf first).

That first one does look like a bug, though jsonschema, despite its name, doesn't deal at all with JSON, it does use Python semantics and objects, and just avoids the Python behavior when needed (because it'd clash with JSON). E.g., it evades issubclass(bool, int) for doing type checks that JSON Schema defines. Seems like uniqueItems is likely missing the same sort of evasion.

The second one also looks like a bug but not here :), because all jsonschema does is use the official meta schemas. E.g. here for enum you'll see that despite the language of the spec, it does refuse any instances with non-unique items (or an empty array). That metaschema is yeah just the official one essentially (copied from e.g. https://github.com/json-schema-org/json-schema-spec/blob/draft-07/schema.json#L139-L144). The right thing to do here then is probably for us to file a bug in the spec repo, and then once we merge it there jsonschema will just stop raising those.

I also just want to say thanks for the jsonschema library - the interface is easy to use, and the fantastic error messages when I misused or misunderstood something have saved so much time debugging.

Thank you! That's very very good to hear :) particularly about the error messages. There's still lots to do to make them even better I suspect, but yeah glad they helped.

[Julian]

OK, have pushed a local fix for the metaschema bugs here... Gonna have to set aside some time to start moving tests upstream to help us move it into the spec repo too.

[Julian]

Hee hee I also just realized that the second bug is likely impossible to fix for arbitrary types, so will need to think about what to do about that.

[Zac-HD]

I just checked this, and it's fixed as of jsonschema==3.0.2 🎉

[Julian]

I mentioned this briefly on json-schema-org/json-schema-spec#717 (comment) but just for archaeological purposes, this change was just partially "backed out" for drafts 3 and 4 where indeed the original behavior (MUST) was correct, those drafts contain stronger language. It's definitely correct on newer ones though (so thanks again for filing it!)

[erohmensing]

@Julian thanks for the "for archeological purposes", rest assured it's worth adding as people still run into this strange mismatch years later 😄