Frequently Asked Questions

My schema specifies format validation. Why do invalid instances seem valid?

The format keyword can be a bit of a stumbling block for new users working with JSON Schema.

In a schema such as:

{"type": "string", "format": "date"}

JSON Schema specifications have historically differentiated between the format keyword and other keywords. In particular, the format keyword was specified to be informational as much as it may be used for validation.

In other words, for many use cases, schema authors may wish to use values for the format keyword but have no expectation they be validated alongside other required assertions in a schema.

Of course this does not represent all or even most use cases – many schema authors do wish to assert that instances conform fully, even to the specific format mentioned.

In drafts prior to draft2019-09, the decision on whether to automatically enable format validation was left up to validation implementations such as this one.

This library made the choice to leave it off by default, for two reasons:

  • for forward compatibility and implementation complexity reasons – if format validation were on by default, and a future draft of JSON Schema introduced a hard-to-implement format, either the implementation of that format would block releases of this library until it were implemented, or the behavior surrounding format would need to be even more complex than simply defaulting to be on. It therefore was safer to start with it off, and defend against the expectation that a given format would always automatically work.

  • given that a common use of JSON Schema is for portability across languages (and therefore implementations of JSON Schema), so that users be aware of this point itself regarding format validation, and therefore remember to check any other implementations they were using to ensure they too were explicitly enabled for format validation.

As of draft2019-09 however, the opt-out by default behavior mentioned here is now required for all implementations of JSON Schema.

Difficult as this may sound for new users, at this point it at least means they should expect the same behavior that has always been implemented here, across any other implementation they encounter.

See also

Draft 2019-09’s release notes on format

for upstream details on the behavior of format and how it has changed in draft2019-09

Validating Formats

for details on how to enable format validation


the object which implements format validation

Can jsonschema be used to validate YAML, TOML, etc.?

Like most JSON Schema implementations, jsonschema doesn’t actually deal directly with JSON at all (other than in relation to the $ref keyword, elaborated on below).

In other words as far as this library is concerned, schemas and instances are simply runtime Python objects. The JSON object {} is simply the Python dict {}, and a JSON Schema like {"type": "object", {"properties": {}}} is really an assertion about particular Python objects and their keys.


The $ref keyword is a single notable exception.

Specifically, in the case where jsonschema is asked to resolve a remote reference, it has no choice but to assume that the remote reference is serialized as JSON, and to deserialize it using the json module.

One cannot today therefore reference some remote piece of YAML and have it deserialized into Python objects by this library without doing some additional work. See Resolving References to Schemas Written in YAML for details.

In practice what this means for JSON-like formats like YAML and TOML is that indeed one can generally schematize and then validate them exactly as if they were JSON by simply first deserializing them using libraries like PyYAML or the like, and passing the resulting Python objects into functions within this library.

Beware however that there are cases where the behavior of the JSON Schema specification itself is only well-defined within the data model of JSON itself, and therefore only for Python objects that could have “in theory” come from JSON. As an example, JSON supports only string-valued keys, whereas YAML supports additional types. The JSON Schema specification does not deal with how to apply the patternProperties keyword to non-string properties. The behavior of this library is therefore similarly not defined when presented with Python objects of this form, which could never have come from JSON. In such cases one is recommended to first pre-process the data such that the resulting behavior is well-defined. In the previous example, if the desired behavior is to transparently coerce numeric properties to strings, as Javascript might, then do the conversion explicitly before passing data to this library.

Why doesn’t my schema’s default property set the default on my instance?

The basic answer is that the specification does not require that default actually do anything.

For an inkling as to why it doesn’t actually do anything, consider that none of the other keywords modify the instance either. More importantly, having default modify the instance can produce quite peculiar things. It’s perfectly valid (and perhaps even useful) to have a default that is not valid under the schema it lives in! So an instance modified by the default would pass validation the first time, but fail the second!

Still, filling in defaults is a thing that is useful. jsonschema allows you to define your own validator classes and callables, so you can easily create an jsonschema.protocols.Validator that does do default setting. Here’s some code to get you started. (In this code, we add the default properties to each object before the properties are validated, so the default values themselves will need to be valid under the schema.)

from jsonschema import Draft202012Validator, validators

def extend_with_default(validator_class):
    validate_properties = validator_class.VALIDATORS["properties"]

    def set_defaults(validator, properties, instance, schema):
        for property, subschema in properties.items():
            if "default" in subschema:
                instance.setdefault(property, subschema["default"])

        for error in validate_properties(
            validator, properties, instance, schema,
            yield error

    return validators.extend(
        validator_class, {"properties" : set_defaults},

DefaultValidatingValidator = extend_with_default(Draft202012Validator)

# Example usage:
obj = {}
schema = {'properties': {'foo': {'default': 'bar'}}}
# Note jsonschema.validate(obj, schema, cls=DefaultValidatingValidator)
# will not work because the metaschema contains `default` keywords.
assert obj == {'foo': 'bar'}

See the above-linked document for more info on how this works, but basically, it just extends the properties keyword on a jsonschema.validators.Draft202012Validator to then go ahead and update all the defaults.


If you’re interested in a more interesting solution to a larger class of these types of transformations, keep an eye on Seep, which is an experimental data transformation and extraction library written on top of jsonschema.


The above code can provide default values for an entire object and all of its properties, but only if your schema provides a default value for the object itself, like so:

schema = {
    "type": "object",
    "properties": {
        "outer-object": {
            "type": "object",
            "properties" : {
                "inner-object": {
                    "type": "string",
                    "default": "INNER-DEFAULT"
            "default": {} # <-- MUST PROVIDE DEFAULT OBJECT

obj = {}
assert obj == {'outer-object': {'inner-object': 'INNER-DEFAULT'}}

…but if you don’t provide a default value for your object, then it won’t be instantiated at all, much less populated with default properties.

del schema["properties"]["outer-object"]["default"]
obj2 = {}
assert obj2 == {} # whoops

How do jsonschema version numbers work?

jsonschema tries to follow the Semantic Versioning specification.

This means broadly that no backwards-incompatible changes should be made in minor releases (and certainly not in dot releases).

The full picture requires defining what constitutes a backwards-incompatible change.

The following are simple examples of things considered public API, and therefore should not be changed without bumping a major version number:

  • module names and contents, when not marked private by Python convention (a single leading underscore)

  • function and object signature (parameter order and name)

The following are not considered public API and may change without notice:

  • the exact wording and contents of error messages; typical reasons to rely on this seem to involve downstream tests in packages using jsonschema. These use cases are encouraged to use the extensive introspection provided in jsonschema.exceptions.ValidationErrors instead to make meaningful assertions about what failed rather than relying on how what failed is explained to a human.

  • the order in which validation errors are returned or raised

  • the contents of the jsonschema.tests package

  • the contents of the jsonschema.benchmarks package

  • the specific non-zero error codes presented by the command line interface

  • the exact representation of errors presented by the command line interface, other than that errors represented by the plain outputter will be reported one per line

  • anything marked private

With the exception of the last two of those, flippant changes are avoided, but changes can and will be made if there is improvement to be had. Feel free to open an issue ticket if there is a specific issue or question worth raising.