JSON (Schema) Referencing

The JSON Schema $ref and $dynamicRef keywords allow schema authors to combine multiple schemas (or subschemas) together for reuse or deduplication.

The referencing library was written in order to provide a simple, well-behaved and well-tested implementation of this kind of reference resolution [1]. It has its own documentation which is worth reviewing, but this page serves as an introduction which is tailored specifically to JSON Schema, and even more specifically to how to configure referencing for use with Validator objects in order to customize the behavior of the $ref keyword and friends in your schemas.

Configuring jsonschema for custom referencing behavior is essentially a two step process:

The examples below essentially follow these two steps.

Introduction to the referencing API

There are 3 main objects to be aware of in the JSON (Schema) Referencing API:

As a concrete example, the simple schema {"type": "integer"} may be interpreted as a schema under either Draft 2020-12 or Draft 4 of the JSON Schema specification (amongst others); in draft 2020-12, the float 2.0 must be considered an integer, whereas in draft 4, it potentially is not. If you mean the former (i.e. to associate this schema with draft 2020-12), you’d use referencing.Resource(contents={"type": "integer"}, specification=referencing.jsonschema.DRAFT202012), whereas for the latter you’d use referencing.jsonschema.DRAFT4.

See also

the JSON Schema $schema keyword

Which should generally be used to remove all ambiguity and identify internally to the schema what version it is written for.

A schema may be identified via one or more URIs, either because they contain an $id keyword (in suitable versions of the JSON Schema specification) which indicates their canonical URI, or simply because you wish to externally associate a URI with the schema, regardless of whether it contains an $id keyword. You could add the aforementioned simple schema to a referencing.Registry by creating an empty registry and then identifying it via some URI:

from referencing import Registry, Resource
from referencing.jsonschema import DRAFT202012
schema = Resource(contents={"type": "integer"}, specification=DRAFT202012)
registry = Registry().with_resource(uri="http://example.com/my/schema", resource=schema)
print(registry)
<Registry (1 uncrawled resource)>

Note

referencing.Registry is an entirely immutable object. All of its methods which add schemas (resources) to itself return new registry objects containing the added schemas.

You could also confirm your schema is in the registry if you’d like, via referencing.Registry.contents, which will show you the contents of a resource at a given URI:

print(registry.contents("http://example.com/my/schema"))
{'type': 'integer'}

For further details, see the referencing documentation.

Common Scenarios

Making Additional In-Memory Schemas Available

The most common scenario one is likely to encounter is the desire to include a small number of additional in-memory schemas, making them available for use during validation.

For instance, imagine the below schema for non-negative integers:

{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "type": "integer",
  "minimum": 0
}

We may wish to have other schemas we write be able to make use of this schema, and refer to it as http://example.com/nonneg-int-schema and/or as urn:nonneg-integer-schema.

To do so we make use of APIs from the referencing library to create a referencing.Registry which maps the URIs above to this schema:

from referencing import Registry, Resource
schema = Resource.from_contents(
    {
        "$schema": "https://json-schema.org/draft/2020-12/schema",
        "type": "integer",
        "minimum": 0,
    },
)
registry = Registry().with_resources(
    [
        ("http://example.com/nonneg-int-schema", schema),
        ("urn:nonneg-integer-schema", schema),
    ],
)

What’s above is likely mostly self-explanatory, other than the presence of the referencing.Resource.from_contents function. Its purpose is to convert a piece of “opaque” JSON (or really a Python dict containing deserialized JSON) into an object which indicates what version of JSON Schema the schema is meant to be interpreted under. Calling it will inspect a $schema keyword present in the given schema and use that to associate the JSON with an appropriate specification. If your schemas do not contain $schema dialect identifiers, and you intend for them to be interpreted always under a specific dialect – say Draft 2020-12 of JSON Schema – you may instead use e.g.:

from referencing import Registry, Resource
from referencing.jsonschema import DRAFT202012
schema = DRAFT202012.create_resource({"type": "integer", "minimum": 0})
registry = Registry().with_resources(
    [
        ("http://example.com/nonneg-int-schema", schema),
        ("urn:nonneg-integer-schema", schema),
    ],
)

which has the same functional effect.

You can now pass this registry to your Validator, which allows a schema passed to it to make use of the aforementioned URIs to refer to our non-negative integer schema. Here for instance is an example which validates that instances are JSON objects with non-negative integral values:

from jsonschema import Draft202012Validator
validator = Draft202012Validator(
    {
        "type": "object",
        "additionalProperties": {"$ref": "urn:nonneg-integer-schema"},
    },
    registry=registry,  # the critical argument, our registry from above
)
validator.validate({"foo": 37})
assert not validator.is_valid({"foo": -37})  # Uh oh!

Resolving References from the File System

Another common request from schema authors is to be able to map URIs to the file system, perhaps while developing a set of schemas in different local files. If you have a set of fixed or static schemas in a few files, you still likely will want to follow the above in-memory instructions, and simply load all of your files by reading them in-memory from your program. If however you wish to dynamically read files off of the file system, perhaps because they may change during the lifetime of your process, then the referencing library supports doing so fully dynamically by configuring a callable which can be used to retrieve any schema which is not already pre-loaded in-memory.

Here we resolve any schema beginning with http://localhost to a directory /tmp/schemas on the local filesystem (note of course that this will not work if run directly unless you have populated that directory with some schemas):

from pathlib import Path
import json

from referencing import Registry, Resource
from referencing.exceptions import NoSuchResource

SCHEMAS = Path("/tmp/schemas")

def retrieve_from_filesystem(uri: str):
    if not uri.startswith("http://localhost/"):
        raise NoSuchResource(ref=uri)
    path = SCHEMAS / Path(uri.removeprefix("http://localhost/"))
    contents = json.loads(path.read_text())
    return Resource.from_contents(contents)

registry = Registry(retrieve=retrieve_from_filesystem)

Such a registry can then be used with Validator objects in the same way shown above, and any such references to URIs which are not already in-memory will be retrieved from the configured directory.

We can mix the two examples above if we wish for some in-memory schemas to be available in addition to the filesystem schemas, e.g.:

from referencing.jsonschema import DRAFT7
registry = Registry(retrieve=retrieve_from_filesystem).with_resource(
    "urn:non-empty-array", DRAFT7.create_resource({"type": "array", "minItems": 1}),
)

where we’ve made use of the similar referencing.Registry.with_resource function to add a single additional resource.

Resolving References to Schemas Written in YAML

Generalizing slightly, the retrieval function provided need not even assume that it is retrieving JSON. As long as you deserialize what you have retrieved into Python objects, you may equally be retrieving references to YAML documents or any other format.

Here for instance we retrieve YAML documents in a way similar to the above using PyYAML:

from pathlib import Path
import yaml

from referencing import Registry, Resource
from referencing.exceptions import NoSuchResource

SCHEMAS = Path("/tmp/yaml-schemas")

def retrieve_yaml(uri: str):
    if not uri.startswith("http://localhost/"):
        raise NoSuchResource(ref=uri)
    path = SCHEMAS / Path(uri.removeprefix("http://localhost/"))
    contents = yaml.safe_load(path.read_text())
    return Resource.from_contents(contents)

registry = Registry(retrieve=retrieve_yaml)

Note

Not all YAML fits within the JSON data model.

JSON Schema is defined specifically for JSON, and has well-defined behavior strictly for Python objects which could have possibly existed as JSON.

If you stick to the subset of YAML for which this is the case then you shouldn’t have issue, but if you pass schemas (or instances) around whose structure could never have possibly existed as JSON (e.g. a mapping whose keys are not strings), all bets are off.

One could similarly imagine a retrieval function which switches on whether to call yaml.safe_load or json.loads by file extension (or some more reliable mechanism) and thereby support retrieving references of various different file formats.

Automatically Retrieving Resources Over HTTP

In the general case, the JSON Schema specifications tend to discourage implementations (like this one) from automatically retrieving references over the network, or even assuming such a thing is feasible (as schemas may be identified by URIs which are strictly identifiers, and not necessarily downloadable from the URI even when such a thing is sensical).

However, if you as a schema author are in a situation where you indeed do wish to do so for convenience (and understand the implications of doing so), you may do so by making use of the retrieve argument to referencing.Registry.

Here is how one would configure a registry to automatically retrieve schemas from the JSON Schema Store on the fly using the httpx:

from referencing import Registry, Resource
import httpx

def retrieve_via_httpx(uri: str):
    response = httpx.get(uri)
    return Resource.from_contents(response.json())

registry = Registry(retrieve=retrieve_via_httpx)

Given such a registry, we can now, for instance, validate instances against schemas from the schema store by passing the registry we configured to our Validator as in previous examples:

from jsonschema import Draft202012Validator
Draft202012Validator(
    {"$ref": "https://json.schemastore.org/pyproject.json"},
    registry=registry,
).validate({"project": {"name": 12}})

which should in this case indicate the example data is invalid:

Traceback (most recent call last):
    ...
jsonschema.exceptions.ValidationError: 12 is not of type 'string'

Failed validating 'type' in schema['properties']['project']['properties']['name']:
    {'pattern': '^([a-zA-Z\\d]|[a-zA-Z\\d][\\w.-]*[a-zA-Z\\d])$',
    'title': 'Project name',
    'type': 'string'}

On instance['project']['name']:
    12

Retrieving resources from a SQLite database or some other network-accessible resource should be more or less similar, replacing the HTTP client with one for your database of course.

Warning

Be sure you understand the security implications of the reference resolution you configure. And if you accept untrusted schemas, doubly sure!

You wouldn’t want a user causing your machine to go off and retrieve giant files off the network by passing it a $ref to some huge blob, or exploiting similar vulnerabilities in your setup.

Migrating From RefResolver

Older versions of jsonschema used a different object – _RefResolver – for reference resolution, which you a schema author may already be configuring for your own use.

_RefResolver is now fully deprecated and replaced by the use of referencing.Registry as shown in examples above.

If you are not already constructing your own _RefResolver, this change should be transparent to you (or even recognizably improved, as the point of the migration was to improve the quality of the referencing implementation and enable some new functionality).

Rough equivalence between _RefResolver and referencing.Registry APIs

Old API

New API

RefResolver.from_schema({"$id": "urn:example:foo", ...}

Registry().with_resource(uri="urn:example:foo", resource=Resource.from_contents({"$id": "urn:example:foo", ...}))

Overriding RefResolver.resolve_from_url

Passing a callable to referencing.Registry‘s retrieve argument

DraftNValidator(..., resolver=_RefResolver(...)) ``

DraftNValidator(…, registry=Registry().with_resources(…))``

Here are some more specifics on how to migrate to the newer APIs:

The store argument

_RefResolver‘s store argument was essentially the equivalent of referencing.Registry‘s in-memory schema storage.

If you currently pass a set of schemas via e.g.:

from jsonschema import Draft202012Validator, RefResolver
resolver = RefResolver.from_schema(
    schema={"title": "my schema"},
    store={"http://example.com": {"type": "integer"}},
)
validator = Draft202012Validator(
    {"$ref": "http://example.com"},
    resolver=resolver,
)
validator.validate("foo")

you should be able to simply move to something like:

from referencing import Registry
from referencing.jsonschema import DRAFT202012

from jsonschema import Draft202012Validator

registry = Registry().with_resource(
    "http://example.com",
    DRAFT202012.create_resource({"type": "integer"}),
)
validator = Draft202012Validator(
    {"$ref": "http://example.com"},
    registry=registry,
)
assert not validator.is_valid("foo")

Handlers

The handlers functionality from _RefResolver was a way to support additional HTTP schemes for schema retrieval.

Here you should move to a custom retrieve function which does whatever you’d like. E.g. in pseudocode:

from urllib.parse import urlsplit

def retrieve(uri: str):
    parsed = urlsplit(uri)
    if parsed.scheme == "file":
        ...
    elif parsed.scheme == "custom":
        ...

registry = Registry(retrieve=retrieve)

Other Key Functional Differences

Whilst _RefResolver did automatically retrieve remote references (against the recommendation of the spec, and in a way which therefore could lead to questionable security concerns when combined with untrusted schemas), referencing.Registry does not do so. If you rely on this behavior, you should follow the above example of retrieving resources over HTTP.