jsonschema-rs

A high-performance JSON Schema validator for Python.

import jsonschema_rs

schema = {"maxLength": 5}
instance = "foo"

# One-off validation
try:
    jsonschema_rs.validate(schema, "incorrect")
except jsonschema_rs.ValidationError as exc:
    assert str(exc) == '''"incorrect" is longer than 5 characters

Failed validating "maxLength" in schema

On instance:
    "incorrect"'''

# Build & reuse (faster)
validator = jsonschema_rs.validator_for(schema)

# Iterate over errors
for error in validator.iter_errors(instance):
    print(f"Error: {error}")
    print(f"Location: {error.instance_path}")

# Boolean result
assert validator.is_valid(instance)

# Structured output (JSON Schema Output v1)
evaluation = validator.evaluate(instance)
for error in evaluation.errors():
    print(f"Error at {error['instanceLocation']}: {error['error']}")

⚠️ Upgrading from older versions? See the Migration Guide for breaking changes.

Migrating from jsonschema? See the jsonschema migration guide.

Highlights

• 📚 Drafts 4, 6, 7, 2019-09 and 2020-12
• 🔧 Custom keywords and format validators
• ⚡ Compile-time validators for your own extension modules, via the Rust crate
• 🌐 $ref resolution over HTTP and from files
• 📦 Schema bundling into Compound Schema Documents, and $ref dereferencing
• 🎨 Structured Output v1 reports (flag/list/hierarchical)
• ✨ Meta-schema validation for schema documents, including custom metaschemas
• 🧮 Experimental schema canonicalization

Supported drafts

Per-draft compliance results are on the Bowtie Report.

Playground

Try schemas in the browser with the WebAssembly playground.

Installation

pip install jsonschema-rs

Usage

Pass a schema as a JSON string to skip parsing it in Python:

import jsonschema_rs

validator = jsonschema_rs.validator_for('{"minimum": 42}')
...

validator_for detects the draft from $schema. To pick one yourself, use a draft-specific class:

import jsonschema_rs

# Automatic draft detection
validator = jsonschema_rs.validator_for({"minimum": 42})

# Draft-specific validators
validator = jsonschema_rs.Draft4Validator({"minimum": 42})
validator = jsonschema_rs.Draft6Validator({"minimum": 42})
validator = jsonschema_rs.Draft7Validator({"minimum": 42})
validator = jsonschema_rs.Draft201909Validator({"minimum": 42})
validator = jsonschema_rs.Draft202012Validator({"minimum": 42})

jsonschema-rs ships validators for the standard format values. To add your own, pass a function that takes a str and returns a bool via formats. Drafts 2019-09 and 2020-12 treat format as an annotation, so set validate_formats=True to get it checked:

import jsonschema_rs

def is_currency(value):
    # The input value is always a string
    return len(value) == 3 and value.isascii()


validator = jsonschema_rs.validator_for(
    {"type": "string", "format": "currency"}, 
    formats={"currency": is_currency},
    validate_formats=True  # Important for Draft 2019-09 and 2020-12
)
validator.is_valid("USD")  # True
validator.is_valid("invalid")  # False

Custom Keywords

A custom keyword is a class. The validator builds one instance per occurrence of the keyword, passing the parent schema, the keyword value and its schema path. Its validate method rejects a value by raising any exception, whose message becomes the error message:

import jsonschema_rs

class DivisibleBy:
    def __init__(self, parent_schema, value, schema_path):
        self.divisor = value

    def validate(self, instance):
        if isinstance(instance, int) and instance % self.divisor != 0:
            raise ValueError(f"{instance} is not divisible by {self.divisor}")


validator = jsonschema_rs.validator_for(
    {"type": "integer", "divisibleBy": 3},
    keywords={"divisibleBy": DivisibleBy},
)
validator.is_valid(9)   # True
validator.is_valid(10)  # False

The resulting ValidationError keeps that exception as __cause__:

try:
    validator.validate(instance)
except jsonschema_rs.ValidationError as e:
    print(type(e.__cause__))   # <class 'ValueError'>
    print(e.__cause__)         # original message

Other options:

• validate_formats: check format regardless of the draft default.
• ignore_unknown_formats: set to False to raise on a format value with no validator.
• base_uri: base URI for relative $refs in the schema.
• vocabularies: vocabularies your custom keywords implement.

import jsonschema_rs

validator = jsonschema_rs.Draft202012Validator(
    {"type": "string", "format": "date"},
    validate_formats=True,
    ignore_unknown_formats=False
)

# This will validate the "date" format
validator.is_valid("2023-05-17")  # True
validator.is_valid("not a date")  # False

# With ignore_unknown_formats=False, an unknown format raises an error
invalid_schema = {"type": "string", "format": "unknown"}
try:
    jsonschema_rs.Draft202012Validator(
        invalid_schema, validate_formats=True, ignore_unknown_formats=False
    )
except jsonschema_rs.ValidationError as exc:
    assert exc.message == (
        "Unknown format: 'unknown'. "
        "Adjust configuration to ignore unrecognized formats"
    )

Structured Output with evaluate

evaluate returns the JSON Schema Output v1 formats instead of a boolean:

import jsonschema_rs

schema = {
    "type": "array",
    "prefixItems": [{"type": "string"}],
    "items": {"type": "integer"},
}
evaluation = jsonschema_rs.evaluate(schema, ["hello", "oops"])
type_error = {"type": '"oops" is not of type "integer"'}

assert evaluation.flag() == {"valid": False}
assert evaluation.list() == {
    "valid": False,
    "details": [
        {
            "evaluationPath": "",
            "instanceLocation": "",
            "schemaLocation": "",
            "valid": False,
        },
        {
            "valid": True,
            "evaluationPath": "/type",
            "instanceLocation": "",
            "schemaLocation": "/type",
        },
        {
            "valid": False,
            "evaluationPath": "/items",
            "instanceLocation": "",
            "schemaLocation": "/items",
            "droppedAnnotations": True,
        },
        {
            "valid": False,
            "evaluationPath": "/items",
            "instanceLocation": "/1",
            "schemaLocation": "/items",
        },
        {
            "valid": False,
            "evaluationPath": "/items/type",
            "instanceLocation": "/1",
            "schemaLocation": "/items/type",
            "errors": type_error,
        },
        {
            "valid": True,
            "evaluationPath": "/prefixItems",
            "instanceLocation": "",
            "schemaLocation": "/prefixItems",
            "annotations": 0,
        },
        {
            "valid": True,
            "evaluationPath": "/prefixItems/0",
            "instanceLocation": "/0",
            "schemaLocation": "/prefixItems/0",
        },
        {
            "valid": True,
            "evaluationPath": "/prefixItems/0/type",
            "instanceLocation": "/0",
            "schemaLocation": "/prefixItems/0/type",
        },
    ],
}

hierarchical = evaluation.hierarchical()
assert hierarchical == {
    "valid": False,
    "evaluationPath": "",
    "instanceLocation": "",
    "schemaLocation": "",
    "details": [
        {
            "valid": True,
            "evaluationPath": "/type",
            "instanceLocation": "",
            "schemaLocation": "/type",
        },
        {
            "valid": False,
            "evaluationPath": "/items",
            "instanceLocation": "",
            "schemaLocation": "/items",
            "droppedAnnotations": True,
            "details": [
                {
                    "valid": False,
                    "evaluationPath": "/items",
                    "instanceLocation": "/1",
                    "schemaLocation": "/items",
                    "details": [
                        {
                            "valid": False,
                            "evaluationPath": "/items/type",
                            "instanceLocation": "/1",
                            "schemaLocation": "/items/type",
                            "errors": type_error,
                        }
                    ],
                }
            ],
        },
        {
            "valid": True,
            "evaluationPath": "/prefixItems",
            "instanceLocation": "",
            "schemaLocation": "/prefixItems",
            "annotations": 0,
            "details": [
                {
                    "valid": True,
                    "evaluationPath": "/prefixItems/0",
                    "instanceLocation": "/0",
                    "schemaLocation": "/prefixItems/0",
                    "details": [
                        {
                            "valid": True,
                            "evaluationPath": "/prefixItems/0/type",
                            "instanceLocation": "/0",
                            "schemaLocation": "/prefixItems/0/type",
                        }
                    ],
                }
            ],
        },
    ],
}

assert evaluation.errors() == [
    {
        "schemaLocation": "/items/type",
        "absoluteKeywordLocation": None,
        "instanceLocation": "/1",
        "error": '"oops" is not of type "integer"',
    }
]

# A failing schema produces no annotations
assert evaluation.annotations() == []

Arbitrary-Precision Numbers

Numbers keep their full precision on the way to Python:

• Integers, regardless of size, are returned as regular int objects.
• Floating-point literals that fit into IEEE-754 become Python floats.
• Floating-point literals that don't fit in float (for example 1e10000 or extremely precise decimals) fall back to decimal.Decimal using their original JSON string representation.

So ValidationError.kind attributes can hold Decimal values:

from decimal import Decimal
from jsonschema_rs import ValidationError, validator_for

validator = validator_for('{"const": 1e10000}')
try:
    validator.validate(0)
except ValidationError as exc:
    assert exc.kind.expected_value == Decimal("1e10000")

# Exponents beyond ~10^1_000_000 are clamped to keep parsing predictable

Schema Bundling and Dereferencing

Produce a Compound Schema Document (Appendix B) by embedding all external $ref targets into a draft-appropriate container. The bundle accepts the same values as the original.

import jsonschema_rs

address_schema = {
    "$schema": "https://json-schema.org/draft/2020-12/schema",
    "$id": "https://example.com/address.json",
    "type": "object",
    "properties": {"street": {"type": "string"}, "city": {"type": "string"}},
    "required": ["street", "city"]
}

schema = {
    "$schema": "https://json-schema.org/draft/2020-12/schema",
    "type": "object",
    "properties": {"home": {"$ref": "https://example.com/address.json"}},
    "required": ["home"]
}

registry = jsonschema_rs.Registry(
    [("https://example.com/address.json", address_schema)]
)
bundled = jsonschema_rs.bundle(schema, registry=registry)

dereference instead replaces each $ref with the schema it points to, for consumers that do not resolve references. Circular references are left in place.

dereferenced = jsonschema_rs.dereference(schema, registry=registry)

Schema Canonicalization

Experimental: the canonicalization API may change in minor releases.

canonicalize reduces a schema to a normal form that accepts the same values. Schemas that accept the same values reduce to equal CanonicalSchema objects, and a schema proven to accept nothing reduces to false:

import jsonschema_rs

canonical = jsonschema_rs.canonicalize({
    "allOf": [
        {"type": "integer", "minimum": 0},
        {"minimum": 10, "maximum": 100},
    ]
})
assert canonical.to_json_schema() == {
    "$schema": "https://json-schema.org/draft/2020-12/schema",
    "type": "integer", "minimum": 10, "maximum": 100,
}

# However they were written, equivalent schemas compare equal
same = jsonschema_rs.canonicalize(
    {"type": "integer", "maximum": 100, "minimum": 10}
)
assert canonical == same

# A schema no value can satisfy collapses
from jsonschema_rs.canonical import Satisfiability

empty = jsonschema_rs.canonicalize(
    {"type": "integer", "minimum": 10, "maximum": 5}
)
assert empty.satisfiability() == Satisfiability.NO

A schema the canonical form cannot model exactly comes back unchanged, with kind == CanonicalKind.RAW. Its view().reason names what stopped the run, and view().pointer the subschema at fault when a single one is.

Canonical schemas combine like sets of values. Every emitted schema carries $schema; the comments below leave it out:

positive = jsonschema_rs.canonicalize({"type": "integer", "minimum": 0})
bounded = jsonschema_rs.canonicalize({"type": "integer", "maximum": 100})

# Every value both admit
positive.intersect(bounded).to_json_schema()
# {"type": "integer", "minimum": 0, "maximum": 100}

# Every value either admits
positive.union(bounded).to_json_schema()
# {"type": "integer"}

# Every value `positive` admits and `bounded` rejects
positive.subtract(bounded).to_json_schema()
# {"type": "integer", "minimum": 101}

# Every value `positive` rejects: other types, negative integers
# and non-integer numbers
positive.negate().to_json_schema()
# {"anyOf": [{"type": ["null", "boolean", "string", "array", "object"]},
#            {"type": "integer", "maximum": -1},
#            {"type": "number", "not": {"multipleOf": 1}}]}

# Containment.NO: `bounded` takes negative integers, `positive` does not
positive.covers(bounded)
positive.satisfiability()  # Satisfiability.YES

Comparing two versions of a schema

subtract tells you what an edit did to a schema. old.subtract(new) accepts exactly the values old accepts and new rejects, so it accepts nothing when the edit lost nothing.

old = jsonschema_rs.canonicalize({"type": "string"})
new = jsonschema_rs.canonicalize({"type": "string", "maxLength": 50})

# What `new` stopped accepting, as a schema
assert old.subtract(new).to_json_schema() == {
    "$schema": "https://json-schema.org/draft/2020-12/schema",
    "type": "string", "minLength": 51,
}
# Nothing is accepted that was not accepted before, so the change only narrows
assert new.subtract(old).satisfiability() == Satisfiability.NO

The direction to check depends on who sends the value:

• For a request schema, check old.subtract(new). It accepts the payloads existing callers send that the new schema rejects.
• For a response schema, check new.subtract(old). It accepts the values the new schema lets a server return that callers never agreed to read.

Satisfiability.NO on the difference proves the edit safe in that direction. UNKNOWN means the canonicalizer could not decide, and proves nothing either way. Read it as the answer that keeps you safe:

• satisfiability(): only Satisfiability.NO proves a schema accepts nothing. Treat UNKNOWN like YES.
• a.covers(b): only Containment.YES proves a accepts every value b accepts. Treat UNKNOWN like NO.

Containment, Satisfiability, Distinctness, CanonicalKind, UnsatisfiableReason, Cause and the exceptions below live in jsonschema_rs.canonical.

The set operations raise IncompatibleOperands when the operands differ in draft, format assertion or regular-expression engine, or resolve # or one external resource to different schemas. They raise UnsupportedOperand when an operand is RAW, and UnsupportedResult when the canonical form cannot express the result exactly. All three subclass CanonicalizationError.

Finding the dead subschemas of a document

find_unsatisfiable walks a whole document and reports each subschema that accepts no value, with the keywords at fault and their pointers:

from jsonschema_rs.canonical import UnsatisfiableReason, find_unsatisfiable

reasons = find_unsatisfiable({
    "properties": {
        "tag": {"type": "string", "minLength": 5, "maxLength": 2},
        "name": {"type": "string"},
    }
})

match reasons["/properties/tag"]:
    case UnsatisfiableReason.Conflict(causes):
        assert [(cause.pointer, cause.keywords) for cause in causes] == [
            ("/properties/tag", ["type"]),
            ("/properties/tag", ["minLength", "maxLength"]),
        ]

# A live subschema is not reported
assert "/properties/name" not in reasons

A reason is Literal (written as false), Empty (one part admits nothing by itself) or Conflict (each part admits values, together they admit none). A missing pointer does not prove that subschema satisfiable. For a document the canonical form cannot model, find_unsatisfiable reports nothing, the same way satisfiability() answers UNKNOWN.

Meta-Schema Validation

jsonschema_rs.meta checks a schema against the meta-schema of its draft:

import jsonschema_rs

# Valid schema
schema = {
    "type": "object",
    "properties": {
        "name": {"type": "string"},
        "age": {"type": "integer", "minimum": 0}
    },
    "required": ["name"]
}

# Validate schema (draft is auto-detected)
assert jsonschema_rs.meta.is_valid(schema)
jsonschema_rs.meta.validate(schema)  # No error raised

# Invalid schema
invalid_schema = {
    "minimum": "not_a_number"  # "minimum" must be a number
}

try:
    jsonschema_rs.meta.validate(invalid_schema)
except jsonschema_rs.ValidationError as exc:
    assert 'is not of type "number"' in str(exc)

Regular Expression Configuration

pattern_options picks the regex engine for pattern and patternProperties and sets its limits:

import jsonschema_rs
from jsonschema_rs import FancyRegexOptions, RegexOptions

# Default fancy-regex engine with backtracking limits
# (supports advanced features but needs protection against DoS)
validator = jsonschema_rs.validator_for(
    {"type": "string", "pattern": "^(a+)+$"},
    pattern_options=FancyRegexOptions(backtrack_limit=10_000)
)

# Standard regex engine for guaranteed linear-time matching
# (prevents regex DoS attacks but supports fewer features)
validator = jsonschema_rs.validator_for(
    {"type": "string", "pattern": "^a+$"},
    pattern_options=RegexOptions()
)

# Both engines support memory usage configuration
validator = jsonschema_rs.validator_for(
    {"type": "string", "pattern": "^a+$"},
    pattern_options=RegexOptions(
        size_limit=1024 * 1024,   # Maximum compiled pattern size
        dfa_size_limit=10240      # Maximum DFA cache size
    )
)

• FancyRegexOptions: default engine, supports lookaround and backreferences

  • backtrack_limit: Maximum backtracking steps
  • size_limit: Maximum compiled regex size in bytes
  • dfa_size_limit: Maximum DFA cache size in bytes

• RegexOptions: matches in linear time, no lookaround or backreferences

  • size_limit: Maximum compiled regex size in bytes
  • dfa_size_limit: Maximum DFA cache size in bytes

If you validate against schemas from untrusted sources, use RegexOptions: a crafted pattern cannot make it backtrack.

Email Format Configuration

email_options makes {"format": "email"} stricter or looser than the spec default:

import jsonschema_rs
from jsonschema_rs import EmailOptions

# Require a top-level domain (reject "user@localhost")
validator = jsonschema_rs.validator_for(
    {"format": "email", "type": "string"},
    validate_formats=True,
    email_options=EmailOptions(require_tld=True)
)
validator.is_valid("user@localhost")     # False
validator.is_valid("user@example.com")   # True

# Disallow IP address literals and display names
validator = jsonschema_rs.validator_for(
    {"format": "email", "type": "string"},
    validate_formats=True,
    email_options=EmailOptions(
        allow_domain_literal=False,  # Reject "user@[127.0.0.1]"
        allow_display_text=False     # Reject "Name <user@example.com>"
    )
)

# Require at least 3 domain segments, e.g. user@sub.example.com
validator = jsonschema_rs.validator_for(
    {"format": "email", "type": "string"},
    validate_formats=True,
    email_options=EmailOptions(minimum_sub_domains=3),
)

• require_tld: Require a top-level domain (e.g., reject "user@localhost") (default: False)
• allow_domain_literal: Allow IP address literals like "user@[127.0.0.1]" (default: True)
• allow_display_text: Allow display names like "Name user@example.com" (default: True)
• minimum_sub_domains: Minimum number of domain segments required

External References

By default, jsonschema-rs fetches external $ref targets over HTTP and from the local file system. Pass a retriever to load them yourself. This one serves schemas from a dict:

import jsonschema_rs

def retrieve(uri: str):
    schemas = {
        "https://example.com/person.json": {
            "type": "object",
            "properties": {
                "name": {"type": "string"},
                "age": {"type": "integer"}
            },
            "required": ["name", "age"]
        }
    }
    if uri not in schemas:
        raise KeyError(f"Schema not found: {uri}")
    return schemas[uri]

schema = {
    "$ref": "https://example.com/person.json"
}

validator = jsonschema_rs.validator_for(schema, retriever=retrieve)

# This is valid
validator.is_valid({
    "name": "Alice",
    "age": 30
})

# This is invalid (missing "age")
validator.is_valid({
    "name": "Bob"
})  # False

For schemas from untrusted sources, pass offline=True. The validator then refuses to fetch any $ref target, so a schema cannot reach your network or file system. Schemas held in a Registry still resolve:

try:
    jsonschema_rs.validator_for(
        {"$ref": "https://example.com/other.json"}, offline=True
    )
except jsonschema_rs.ValidationError as exc:
    assert "Retrieval is disabled" in str(exc)

bundle, dereference and canonicalize take offline too.

Schema Registry

A Registry holds schemas by URI, so a validator resolves $refs to them without fetching anything:

import jsonschema_rs

# Create a registry with schemas
registry = jsonschema_rs.Registry([
    ("https://example.com/address.json", {
        "type": "object",
        "properties": {
            "street": {"type": "string"},
            "city": {"type": "string"}
        }
    }),
    ("https://example.com/person.json", {
        "type": "object",
        "properties": {
            "name": {"type": "string"},
            "address": {"$ref": "https://example.com/address.json"}
        }
    })
])

# Use the registry with any validator
validator = jsonschema_rs.validator_for(
    {"$ref": "https://example.com/person.json"},
    registry=registry
)

# Validate instances
assert validator.is_valid({
    "name": "John",
    "address": {"street": "Main St", "city": "Boston"}
})

Registry also takes a default draft and a retriever for URIs it does not hold:

import jsonschema_rs

registry = jsonschema_rs.Registry(
    resources=[
        (
            "https://example.com/address.json",
            {}
        )
    ],  # Your schemas
    draft=jsonschema_rs.Draft202012,  # Optional
    retriever=lambda uri: {}  # Optional
)

Error Handling

A ValidationError carries the message, both locations and a kind with keyword-specific details:

import jsonschema_rs

schema = {"type": "string", "maxLength": 5}

try:
    jsonschema_rs.validate(schema, "too long")
except jsonschema_rs.ValidationError as error:
    # Basic error information
    print(error.message)       # '"too long" is longer than 5 characters'
    print(error.instance_path) # Location in the instance that failed
    print(error.schema_path)   # Location in the schema that failed

    # Detailed error information via `kind`
    if isinstance(error.kind, jsonschema_rs.ValidationErrorKind.MaxLength):
        assert error.kind.limit == 5
        print(f"Exceeded maximum length of {error.kind.limit}")

The type stubs list every error kind and its attributes.

Error Kind Properties

kind also has generic accessors:

for error in jsonschema_rs.iter_errors({"minimum": 5}, 3):
    print(error.kind.name)      # "minimum"
    print(error.kind.value)     # 5
    print(error.kind.as_dict()) # {"limit": 5}

Each kind is a class you can match on:

for error in jsonschema_rs.iter_errors({"minimum": 5}, 3):
    match error.kind:
        case jsonschema_rs.ValidationErrorKind.Minimum(limit=limit):
            print(f"Value below {limit}")
        case jsonschema_rs.ValidationErrorKind.Type(types=types):
            print(f"Expected one of {types}")

Error Message Masking

Pass mask to replace instance values in error messages with a placeholder:

import jsonschema_rs

schema = {
    "type": "object",
    "properties": {
        "password": {"type": "string", "minLength": 8},
        "api_key": {"type": "string", "pattern": "^[A-Z0-9]{32}$"}
    }
}

# Replace instance values with "[REDACTED]"
validator = jsonschema_rs.validator_for(schema, mask="[REDACTED]")

try:
    validator.validate({
        "password": "123",
        "api_key": "secret_key_123"
    })
except jsonschema_rs.ValidationError as exc:
    assert str(exc) == '''[REDACTED] is shorter than 8 characters

Failed validating "minLength" in schema["properties"]["password"]

On instance["password"]:
    [REDACTED]'''

Performance

Compared with other Python validators:

• 215-15,699x faster than jsonschema for complex schemas and large instances
• 12-2,623x faster than fastjsonschema on CPython

Full results and methodology are in BENCHMARKS.md.

Compile-Time Validators

If you ship your own extension module and know the schema at build time, the Rust crate's #[jsonschema::validator(..., backend = Pyo3)] macro compiles it into a validator that reads Python objects directly, so nothing is parsed or compiled when your module is imported, and validation runs up to 3.0x faster than with a validator built at run time. See the macro documentation.

This is not available from the jsonschema-rs package on PyPI, which takes its schemas at run time. A complete extension with its build and test commands lives in examples/pyo3-extension.

Python support

jsonschema-rs supports CPython 3.10 through 3.14 and PyPy 3.10+.

Pre-built wheels are available for:

• Linux: x86_64, i686, aarch64 (glibc and musl)
• macOS: x86_64, aarch64, universal2
• Windows: x64, x86, arm64

Troubleshooting

If a source build on Linux fails with linking errors (e.g., undefined symbols from ring), use the mold linker:

RUSTFLAGS="-C link-arg=-fuse-ld=mold" \
  pip install jsonschema-rs --no-binary :all:

Acknowledgements

The API design draws on the Python jsonschema package. Thanks to its maintainers and contributors.

Support

Ask questions and suggest improvements in GitHub Discussions.

Sponsorship

If you find jsonschema-rs useful, please consider sponsoring its development.

Contributing

Ways to help:

• Share your use cases
• Implement missing keywords
• Fix failing test cases from the JSON Schema test suite

See CONTRIBUTING.md for more details.

License

Licensed under MIT License.