Add "Any" scalar type

[IvanGoncharov]

In some of our tool, we need to be able to pass an arbitrary value as an argument of a directive, i.e. value which can be anything and can’t be described as a single type. The only workaround we found is to define scalar JSON which can accept any JSON value including objects and arrays.

Moreover, this as popular approach in the community:

• Graph.cool and Scaphold both allow user to use JSON scalar in their schema
• graphql-type-json implements such scalar as a NPM package with 20+ packages depending on it.
• Apollo uses scalar JSON as an example in their documentation.

AFAIK, this technique doesn’t contradict GraphQL spec and solves issues at hand. At the moment, it has very limited support from tooling since it’s not a built-in type and introspection provides you no clue beyond the name of a scalar.

The idea behind this PR is to formalize existing practice for creating scalar JSON but in protocol agnostic way, that’s why I propose to use Any as the name.

The last argument I have is that all type system that I know have some form of opt-out type: any in Flow and TypScript, void* in C/C++, sql_variant in SQL and Object class in C#/JAVA.

[leebyron]

I also think that this proposal needs some strong support from real product teams (perhaps someone from AirBnB or Github?) rather than just citing graph.cool/scaffold/Apollo since those are generalized tools instead of product APIs.

My primary concern (which I share with @dschafer) is that despite GraphQL being pervasive at Facebook there hasn't been a need for an Any type, and in fact if we had an Any type we may see it as an anti-pattern and put lint rules in place to ensure people aren't just skipping out of properly typing their API. Facebook could be a special case, I'd be interested in hearing some real examples of product teams defining these types.

[dschafer]

One interesting question that comes to mind for me is whether Any, in this context, is a Top Type to all object types. For example, can I do this?

interface I { 
  f: any
}

type T implements I {
  f: T
}

It would feel a bit odd to do that as it stands because any is a scalar.

---

But yeah, my primary concern is what @leebyron articulated: we really haven't seen a demand for this at Facebook. So when I think about potential use cases of this, there aren't a lot of positive examples that spring to mind... but the negative examples of someone lazily typing a field as Any instead of actually describing the types are certainly there.

I think there's a potential argument that the added friction that exists today to enable this functionality with a custom scalar is actively desirable. If this functionality were available as GraphQLAny or GraphQL::Any in a library, suddenly when there's an interesting-but-known return values, the easiest thing to do (just type it as any and move on) is no longer the right thing to do (figure out the shape of the return values and describe them), which is unfortunate.

That being said, the absence of evidence at Facebook isn't evidence of absence in general, so definitely interested in hearing more about the use cases!

[andimarek]

I am not sure about this change: I can attest that people ask and want something like that when they first learn GraphQL ("I just want to have JSON/Map"), but I always argue that this is against the fundamental idea of GraphQL and it is not really supported. If it is really needed a custom Scalar can be created.

I agree with @leebyron that I would like to see some real life examples.

And while Programming Languages need something like that I am not sure about GraphQL: it is about describing an API and therefore fundamentally different from a PL.

And as pointed out from @dschafer it adds complexity to the type system itself: Is any the super type of all types? Can I be more specific when I implement an Interface?

[stubailo]

I don't have an opinion on this - I think the current approach with a JSON scalar is okay, but it is something that comes up somewhat often so it might be nice to have a standard.

In Optics, we use this to return data from our payment system which allows arbitrary tags and data, and we don't want to put all of the options in the schema.

[IvanGoncharov]

I've left some comments here to help clean up the spec language and catch some ambiguities,

@leebyron I updated PR although it may still have some ambiguities as I'm not a native speaker.

however next steps should be to link a reference implementation PR, and to hold discussion.

I will start working on it.

One interesting question that comes to mind for me is whether Any, in this context, is a Top Type to all object types. For example, can I do this?

@dschafer my idea is to treat it as scalar Any so it will have no special meaning for objects, interfaces, etc. But I understand how based on previous interaction with any type in other languages like Flow or TS person can have such assumptions. Maybe it makes sense to rename Any to Mixed, Variant or something else?

when there's an interesting-but-known return values, the easiest thing to do (just type it as any and move on) is no longer the right thing to do (figure out the shape of the return values and describe them), which is unfortunate.

For that, you don't need Any or JSON since you can return String with JSON blob in it. It's something that people are already doing, for example, Graphene provides JSONString scalar as a built-in type.

I think there's a potential argument that the added friction that exists today

In JS it's a matter of two lines: npm install --save 'graphql-type-json import GraphQLJSON from 'graphql-type-json';

For Sangria you can use snippet discouraged but provided by it's author. And probably similar packages exist for many popular GraphQL implementations.

Also, support for any in TS and Flow doesn't prevent developers from writing proper typings.

---

We use JSON type in our projects and we learned that it provides totally different experience: no autocompletion and error checking in GraphiQL, no automatic error checking on the server, you need to figure out how to document fields/args using it, etc.

IMHO, using JSON/Any doesn't fall into the category of workarounds that goes unnoticed for a long time and then suddenly shoot you in the leg. It's like you draw borderline where GraphQL ends and JSON blob starts. And with Any scalar this borderline is clearly visible both on backend and frontend.

[OlegIlyenko]

Just my 5 cents. I would agree with others. I do see demand for JSON/Map but I'm yet to see a good compelling use-case for it (if you look at sangria's snippet, you will see a big disclaimer. :) ). For the most part, I heard about it in a context of the transition from untyped API. So I would rather prefer to avoid introduction of this feature unless we have a very compelling real-world use-case.

[robzhu]

In Optics, we use this to return data from our payment system which allows arbitrary tags and data, and we don't want to put all of the options in the schema.

@stubailo, can you please go into more detail? What are some examples for tags/data? Do they take on any standard shape? Why aren't they useful to include in the schema? How are you parsing this data and what assurances do you have about the shape of the deserialized object?

[BjRo]

I'll take a stab at adding some real product team input 😄

In the context of adopting GraphQL at XING SE and transitioning an untyped API to it (I think this is what @OlegIlyenko was referring to), we're also using an untyped JSON scalar and would like to express support for considering this proposal.

Our current use-case is two-fold:

1. Ease the transition of untyped APIs to GraphQL.

For example we have a backend specific JSON error response that we did not yet model in a typed way but would like to already gradually adopt a typed API and pass this yet unmodelled part through at a certain node.

2. Provide custom directives on the schema to support development

As the different product teams integrating via GraphQL do not all use the same programming language, we try to make it possible to work directly with the schema definition as IDL. Similar to what @IvanGoncharov was describing we also define directives that need to receive arbitrary input.

These include:

• stubbing backend responses via @const(...)
• declaratively defining REST resolvers via @httpGet(...), @httpPut(...), etc.

HTH, Björn

[machineghost]

I might just be a GraphQL n00b, but my use case for this scalar is a dynamic form. My client needs to be able to ask my server "give me all the values for form X for user Y" (where "form X" could have any number of fields). I could represent that in totally uncontroversial structured GraphQL by storing each value as an object in an array:

[{ field: 'foo', value: 1 }, { field: 'bar', value: 2 }, ...]

or I could just have a JSON field with the value:

{ foo: 1, bar: 2 }

The latter is less bandwidth and seems easier to understand to me.

[ofersadgat]

I am trying to do this as well. I have two use cases:

1. I have a swagger endpoint which already exposes type data and I'm building a system which auto-converts from swagger.json to graphql. This helps transition technologies from one to the other in parts rather than all at once. Since swagger currently supports an 'Any' type, I'm stuck.
2. there are a few APIs that Swagger is exposing which use the Any type:

• an endpoint in which a database row is the object parameterized by tableName
  • returns an array of json objects with the keys as the column names and values as the retrieved value.
• a generic "execute sql" and return the result.
  • returns arbitrarily structured data because it depends on what query you execute.

I can see no way of implementing these endpoint without completely bypassing the graphql type system either via a JSON type or stringifying the data and passing another 'type' field which feels even worse.

Given the difficulty of implementing these endpoints with GraphQL you might ask why I'm even trying to do so. The reason there is that there are many other APIs that can be easily typed with GraphQL and I'm trying to move our stack the direction of GraphQL + Relay Modern / Apollo and if not all of the endpoints can be expressed with GraphQL that means that I have to expose both a REST server + GraphQL server to the UI as opposed to exposing only GraphQL but backing it with the REST + Swagger server. Furthermore, it also means that I cannot migrate endpoints from REST to GraphQL seamlessly because the UI has to know about which endpoints are coming from REST and which are coming from GraphQL.

P.S. I tried to implement this instead using a Union which includes Scalar types, but that isnt supported either. In swagger, you can specify that an api returns anyOf: typelist which isnt currently possible to express in graphql if typelist includes a scalar. That might be a way to allow an any type to exist in the system for people who want it (or not have it if you dont). This is further useful for apis which return an object on success or a plain string on error.

[serle]

I would prefer to implement the "any" concept by relaxing the restriction on unions and allowing a union of scalars and/or non-scalars as mentioned previously by @ofersadgat. Our primary user-case is the ability to create arrays of key-value pairs for property maps.

I feel that the whole JSON scalar concept is an anti-pattern as it allows one to bypass the benefits of a strongly types API, instead of addressing the shortfalls directly.

[dmythro]

Why not adding this feature for development-only mode at least, with warnings etc?

[svilendobrev]

i also prefer allowing unions to have both objects and scalars (example: field "user" can be int-id, uuid or whole User object, depending on some settings/context/needs/permissions). This way is less prone to "make everything a void*" while still gives a lot of freedom, and allows for more cleaner APIs. (currently this can be done with 3 fields user_obj, user_id, user_uuid - but if its for each and every type around, that's alot of clutter)