Modelling Inheritance with JSON Schema
Probably the most common question that we get is, "How do I model an inheritance hierarchy in JSON Schema?" And most commonly, our answer to that question is, "You don't."
JSON Schema just isn't designed for that. It's a subtractive system, where more constraints means fewer matches, and data modelling tends to be additive, where more definition means more matches. The systems are inherently incompatible.
However, if we accept a few concessions, we might just be able to work something out.
Our Models
To get started, we're going to attempt to model some computer peripherals. In a strongly typed language, we may model this using a Peripheral
base class that defines a number of properties (and typically functions) that are common to all peripherals. Then, each device would be a subclass of this base class.
For our purposes, we're just going to define the name
property on the base class. That is, every peripheral needs to have a name.
I'm going to use TypeScript for the code samples, but the concepts will apply to other languages as well.
1abstract class Peripheral {
2 name: string;
3 // ...
4}
5
Now we can define other peripherals, Mouse
and Keyboard
, by inheriting from this base class.
1class Mouse extends Peripheral {
2 buttonCount: number;
3 wheelCount: number;
4 trackingType: "ball" | "optical";
5 // ...
6}
7
8class Keyboard extends Peripheral {
9 keyCount: number;
10 mediaButtons: boolean;
11 // ...
12}
13
This will suffice to get us started.
Representing Our Models Using Constraints
In JSON Schema, ideally, we'd want to have schemas for each of these. For peripheral, we might try something like this:
I'm using a schema:
URI for the schema identifiers since these schemas aren't accessible anywhere. It's a recommendation that we're considering for the upcoming version of JSON Schema. Let us know if you like this approach.
But that additionalProperties
keyword causes problems. Specifically, "inherited" schemas (like what we're going to build for Mouse
) can't define additional properties, which is something it definitely needs to do. That just won't work at all, and the solution is simply to omit it.
But now, any JSON object with a name
property is validated as a peripheral. While not quite right, we can live with it. This gives us our first concession:
Schemas that model base classes cannot verify that an instance represents a derivation of that class.
Modelling the derivations is pretty straightforward: we model what the derivation defines and add a $ref
back to the base schema.
{ "$schema": "https://json-schema.org/draft/2020-12/schema", "$id": "schema:keyboard", "$ref": "schema:peripheral", "properties": { "keys": { "type": "integer" }, "mediaButtons": { "type": "boolean" } }, "required": [ "keys", "mediaButtons" ], "unevaluatedProperties": false}
For the derived schemas, we can use unevaluatedProperties
because these don't have any schemas which derive from them. If the inheritance hierarchy is bigger and these classes serve as bases for others, we'd have to leave the unevaluatedProperties
off, just like we did for schema:peripheral
. Checking for extra properties can only be done for the leaves of the inheritance tree.
Additionally, we use unevaluatedProperties
instead of additionalProperties
because we need it to be able to "see inside" of the $ref
to identify that name
was evaluated as part of the base schema. With additionalProperties
, name
would be rejected.
That seems pretty simple, and we only had to make a single (and rather easy) concession.
Adding a Recursive Reference
What if one of our peripherals could itself have other peripherals attached? Say, a USB hub.
1class UsbHub extends Peripheral {
2 connectedDevices: Peripheral[];
3 // ...
4}
5
Let's try to model that in a schema:
This works, but remember that first concession we made? This schema would allow items that are anything with a string name
property. But that doesn't align with the TypeScript models. The TypeScript model says that connectedDevices
can only hold types that are derived from Peripheral
.
While this may be sufficient for some, in my opinion it doesn't work. I want to ensure that the items in the connectedDevices
array are only known peripheral types. To do this, we need another schema.
Supporting Only Known Derivations
Problem: we want a schema that identifies that some JSON represents one of our known device types.
Solution: define the schema using a oneOf
that references all of the known device type schemas.
This schema is pretty basic. It just says, "If the JSON matches one of these devices, then it's a known peripheral."
We can now reference this in schema:usbhub
.
Now the USB hub and its connected devices can be validated properly.
The catch is that because I can't dynamically add items to the oneOf
, I can only support devices I know about at dev time. For most cases, this isn't a problem. However, if I plan on publishing this in a package to be used by others, it wouldn't support devices that they create. (I do have a workaround for this, but it's not a good one, so I won't be sharing it here.) This gives us our second concession:
If we need references to the base class, we can only support derivations we know about ahead of time.
An Unexpected Benefit
To determine whether some JSON is a Mouse
or a Keyboard
or a UsbHub
, one might hold all three of those schemas and validate each one of them in turn to determine which one was received. But our solution to the referencing problem actually gives us a better option.
We know that schema:known-peripherals
can validate any known peripheral (because we designed it to do that), but if we use a more verbose output format, it can actually tell us which kind of peripheral we got.
First, we identify which oneOf
subschema passed validation by looking through its child output nodes for a valid: true
. We know that's going to be a $ref
schema (because it's a oneOf
that contains only $ref
schemas), which means that the child output node of that $ref
schema will represent the output of the peripheral schema, which contains the peripheral schema's $id
URI.
So in a single validation pass, we get whether it's a supported peripheral of any kind, and we can discern what kind it is. Two birds, one stone.
So is Inheritance in JSON Schema Possible?
No.
And yes, if we're okay that:
Schemas that model base classes cannot verify that an instance represents a derivation of that class.
If we need references to the base class, we can only support derivations we know about ahead of time.
I think these will be acceptable for most people, but I'm also sure that someone will inevitably run into a scenario where this approach won't work.
This is the best I've seen so far at modelling inheritance, and I'm fairly certain that JSON Schema can't get it 100% right without some new functionality.
If you have some other ideas about how to support polymorphism, or if you think polymorphism is overrated and JSON Schema doesn't need to support it, please join the conversation in our IDL Vocabulary repository.
Cover image by Gerd Altmann on Pixabay