rjbs forgot what he was saying

not logged in (root) | by date | tagcloud | help | login

validation is hard

by rjbs, created 2008-07-17 23:26
tagged with: @markup:md journal programming

More and more, I'm dealing with lots of mostly-JSON web service APIs. I enjoy this. It's simple, and tends to work well. I think JSON is a really nice format to work with. I had a discussion with Ingy and a bit of the #yaml gang about things I don't like about YAML (largely it's implicit types) and I ended up thinking that YAML wasn't quite as insane as I'd thought. Still, from the perspective of clarity, JSON blows it out of the water.

I want to add more formalized validation to our JSON APIs, so I was looking at ways to quickly write validation for data structures. I'd seen Kwalify and its Perl implementation out of the corner of my eye a number of times, and I looked into it more today. It started out looking quite reasonable, but then I started to find things that made me unhappy.

The first thing that should've made me wary was the fact that you can declare a pattern for a node, saying that the string must match a regex. That's not so crazy if you're writing a library for one language, but Kwalify seems like it's multilanguage. What regex library does it use? You can't specify it. Presumably it's the "native" library, but that means that your schemata are already not cross-language. That didn't jump out at me until later.

There's an assert, too, which is basically a string of code that gets templated and evaluated. This made me grumpy, but it's marked as "experimental," so I pretended that it would be ditched or replaced.

The mechanism for reusable schema parts is the YAML reference/anchor schema, which is equivalent to Perl references. In order to say that both "masters" and "servants" are the same kind of data, you might write a schema like this:

---
type: map
mapping:
  masters:
    type: seq
    sequence:
      - &person
        type: map
        mapping:
          name: { type: str }
          age:  { type: int }
  servants:
    type: seq
    sequence:
      - *person

...or something like that. I'd rather write something like:

---
schema: map/person
type: map
mapping:
  name: { type: str }
  age:  { type: int }
...
---
type: map
mapping:
  masters:
    type: seq
    sequence:
      - type: map/person
  servants:
    type: seq
    sequence:
      - type: map/person

Then I could have some better registry of types than just "references to other parts of one document." For example, the first document in the above section could actually be resolved by Some Unspecified Mechanism like "look up in filesystem." Without that, each schema is an island.

Kwalify also has a "data-binding" feature, which really means that it can generate a class from a schema and transform input into objects. That's okay, but I don't see how the fact that you might want to do that belongs in the schema. Doesn't it belong in some loader layer? Every time more features are added that do runtime things, the schema system becomes less cross-platform.

The two container types allowed in Kwalify are seq and map, which correspond to YAML's sequence and mapping. Unfortuantely, the sequence type can only define sequences of zero or more elements each matching the same constraint. In other words, you can say "this is an array of zero or more integers" but not "this is an array of between 2 and 8 integers" or "this is an array of an integer, then a string, then a map."

Validation systems always disappoint. So far my experience has been that everybody has his own particular fetishes for what he wants from a validator, and the more validation he has done, the more perverse his desires become. If you're really satisfied with your data validator (and you only use the one library for it) you're probably not validating enough.

Some time ago, I looked at BDFOY's Brick library. It looks great, but it has no non-developer releases, and the impression I got is that while it's really, really cool, no one actually uses it. I'm wary of being the first one to try to use it in production.

Beyond that, it's pretty clearly Perl. Of course, it would be possible to write something to compile simpler schema definitions into Bricks, so you could hand your compiler something like a Kwalify YAML file and get a Brick.

I've spent the last two hours or so trying to decide how I'd want to write this, with an eye toward being able to say things like, "and then you should have an array with one or more things matching the Pobox/Spam/Item schema." I'm also keen on being able to have things defined in terms of semantics that can be implemented in multiple languages, rather than being specified in regex. Maybe that just means you have to say "to be a valid Email/Address, you meet these criteria" and people are responsible for their own testing.

Of course, with a schema in a format like JSON, and a cross-platform testing system like FIT, this could be practicable. Further, it's not the worst thing in the world if parts of the schema can't be validated everywhere by everyone. It lets the client say "well, I tried to validate before sending, but I do not have all of the schema to be sure I am okay." This might not be acceptable for two-way transmission, but it's okay as long as the receiving end can validate enough to be happy.

I'm going to keep thinking about it while I clean the guinea pig cage.