Fixed

• Wrong results in type: integer, const, enum, uniqueItems, multipleOf, minimum, maximum, exclusiveMinimum and exclusiveMaximum for numbers with exponents past one million, such as 1e-2000000.

Performance

• Validation of numbers with many digits or large exponents in const, enum, uniqueItems, type: integer, multipleOf, minimum, maximum, exclusiveMinimum and exclusiveMaximum taking time growing faster than the square of the digit count, instead of at most the square.
• Validation of numbers with more than ten thousand digits, written or implied by their exponent, taking time linear in their written length instead of quadratic in their digit count.

• Panic on numbers past f64 in const, enum, minimum, maximum, exclusiveMinimum, exclusiveMaximum and type: integer when another crate in the build enables serde_json/arbitrary_precision while the arbitrary-precision feature is off. #1790
• #[jsonschema::validator] panicking during expansion, or generating code that panics, on schema numbers past f64 when another crate in the build enables serde_json/arbitrary_precision while the arbitrary-precision feature is off. Such numbers are compile errors, as without that serde_json feature. #1790
• const, enum and uniqueItems treating objects with the same members in a different key order as unequal when another crate in the build enables serde_json/preserve_order. #1791
• Wrong results in type: integer, const, enum, uniqueItems, multipleOf, minimum, maximum, exclusiveMinimum and exclusiveMaximum for numbers with exponents past one million, such as 1e-2000000, with the arbitrary-precision feature.
• #[jsonschema::validator] rejecting multipleOf with a positive divisor below the smallest f64, such as 1e-400, with the arbitrary-precision feature.

Performance

• Validation of numbers with many digits or large exponents in const, enum, uniqueItems, type: integer, multipleOf, minimum, maximum, exclusiveMinimum and exclusiveMaximum with the arbitrary-precision feature taking time growing faster than the square of the digit count, instead of at most the square.
• Validation of numbers with more than ten thousand digits, written or implied by their exponent, with the arbitrary-precision feature taking time linear in their written length instead of quadratic in their digit count.

Fixed

• Wrong results in type: integer, const, enum, uniqueItems, multipleOf, minimum, maximum, exclusiveMinimum and exclusiveMaximum for numbers with exponents past one million, such as 1e-2000000.

Performance

• Validation of numbers with many digits or large exponents in const, enum, uniqueItems, type: integer, multipleOf, minimum, maximum, exclusiveMinimum and exclusiveMaximum taking time growing faster than the square of the digit count, instead of at most the square.
• Validation of numbers with more than ten thousand digits, written or implied by their exponent, taking time linear in their written length instead of quadratic in their digit count.

Changed

• ValidationErrorKind::Enum and ValidationErrorKind::Constant holding their values as Arc<Value> shared with the validator, instead of a copy per error. #1777

Performance

• Validation with unfinished regex matches running each such match once instead of once per run, and comparing errors across runs in linear time instead of quadratic. #1784
• items and additionalProperties looking up a single-keyword subschema once per array or object instead of once per value.
• is_valid and validate reading fewer cache lines per subschema, by keeping error locations apart from validators.
• is_valid on recursive schemas caching only $ref results that took many steps to compute, instead of every one.
• Error evaluation paths taking memory linear in the $ref nesting depth, not quadratic. #1778

Performance

• Validation with unfinished regex matches running each such match once instead of once per run, and comparing errors across runs in linear time instead of quadratic. #1784
• items and additionalProperties looking up a single-keyword subschema once per array or object instead of once per value.
• is_valid and validate reading fewer cache lines per subschema, by keeping error locations apart from validators.
• is_valid on recursive schemas caching only $ref results that took many steps to compute, instead of every one.
• Error evaluation paths taking memory linear in the $ref nesting depth, not quadratic. #1778

Added

• Support for Python 3.15.

Changed

• Update pyo3 to 0.30.
• canonicalize and canonical.find_unsatisfiable take schema positionally only, like bundle and dereference.

Fixed

• Type stubs rejecting valid usage: dict | bool schemas, ReferencingError in except, match on error kinds.
• import jsonschema_rs.canonical.json and import jsonschema_rs.canonical.schema raising ModuleNotFoundError.
• inspect.signature of validator and Evaluation methods missing self.
• ValidationErrorKind reporting builtins as its module.
• validator_for returning instances that fail isinstance checks against Validator and draft classes.

Performance

• Validation with unfinished regex matches running each such match once instead of once per run, and comparing errors across runs in linear time instead of quadratic. #1784
• items and additionalProperties looking up a single-keyword subschema once per array or object instead of once per value.
• is_valid and validate reading fewer cache lines per subschema, by keeping error locations apart from validators.
• is_valid on recursive schemas caching only $ref results that took many steps to compute, instead of every one.
• Error evaluation paths taking memory linear in the $ref nesting depth, not quadratic. #1778

Added

• CLI: - as a path reads the schema or an instance from stdin.
• CLI: -V as the version flag; -v stays as an alias.

Changed

• CLI: Exit status 3 for errors that prevent a verdict and 2 for usage errors, so 1 always means an invalid instance or schema.
• CLI: Errors print to stderr instead of stdout.
• CLI: --version prints jsonschema-cli <VERSION> instead of Version: <VERSION>.

Fixed

• CLI: Panic when stdout is closed early, as in jsonschema-cli validate ... | head.
• CLI: Missing file path in the error for an unreadable instance in text output.
• CLI: jsonschema-cli bundle, jsonschema-cli dereference and jsonschema-cli canonicalize truncating the --output file when the write fails partway.
• CLI: jsonschema-cli validate --help describing --offline in terms of --resource, which validate does not accept.
• CLI: Error and warning messages suggesting jsonschema commands instead of the installed jsonschema-cli binary.
• Canonicalizing several subschemas of one prepared document, where a $ref in a later output could point at another subschema.
• canonicalize leaving unevaluatedProperties or unevaluatedItems beside anyOf or oneOf raw.
• canonicalize giving a different form for a oneOf depending on the order of its branches.
• canonicalize output changing when canonicalized again, for a not over a $ref it cannot fully negate.
• canonicalize output changing when canonicalized again, for a oneOf whose branches reduce to the same $ref.
• canonicalize leaving an anyOf over many overlapping objects raw.
• canonicalize leaving dependencies or dependentRequired raw when their keys form a long chain or cycle.
• canonicalize treating -0 as a value distinct from 0 under arbitrary-precision.
• Regex matches the engine cannot finish, such as on reaching the backtrack limit, counted as non-matches, letting not or patternProperties accept invalid instances. #1715
• canonicalize treating regex matches the engine cannot finish as non-matches.

Performance

• canonicalize giving up sooner on an allOf over unions too wide to model.

Fixed

• Canonicalizing several subschemas of one prepared document, where a $ref in a later output could point at another subschema.
• canonicalize leaving unevaluatedProperties or unevaluatedItems beside anyOf or oneOf raw.
• canonicalize giving a different form for a oneOf depending on the order of its branches.
• canonicalize output changing when canonicalized again, for a not over a $ref it cannot fully negate.
• canonicalize output changing when canonicalized again, for a oneOf whose branches reduce to the same $ref.
• canonicalize leaving an anyOf over many overlapping objects raw.
• canonicalize leaving dependencies or dependentRequired raw when their keys form a long chain or cycle.
• Regex matches the engine cannot finish, such as on reaching the backtrack limit, counted as non-matches, letting not or patternProperties accept invalid instances. #1715
• canonicalize treating regex matches the engine cannot finish as non-matches.

Performance

• canonicalize giving up sooner on an allOf over unions too wide to model.

Changed

• macOS arm64 wheel replaces the universal2 wheel.

Fixed

• Canonicalizing several subschemas of one prepared document, where a $ref in a later output could point at another subschema.
• canonicalize leaving unevaluatedProperties or unevaluatedItems beside anyOf or oneOf raw.
• canonicalize giving a different form for a oneOf depending on the order of its branches.
• canonicalize output changing when canonicalized again, for a not over a $ref it cannot fully negate.
• canonicalize output changing when canonicalized again, for a oneOf whose branches reduce to the same $ref.
• canonicalize leaving an anyOf over many overlapping objects raw.
• canonicalize leaving dependencies or dependentRequired raw when their keys form a long chain or cycle.
• Regex matches the engine cannot finish, such as on reaching the backtrack limit, counted as non-matches, letting not or patternProperties accept invalid instances. #1715
• canonicalize treating regex matches the engine cannot finish as non-matches.

Performance

• canonicalize giving up sooner on an allOf over unions too wide to model.

Fixed

• \s in pattern missing most Unicode space separators and U+2028.
• ECMA regex character class literals being interpreted as Rust set operators.