- Support for Python 3.8 has been dropped, as it is nearing end-of-life.
- Do not reorder dictionaries (schemas, instances) that are printed as part of validation errors.
- Declare support for Py3.13
- Improve
best_match
(and thereby error messages fromjsonschema.validate
) in cases where there are multiple sibling errors from applyinganyOf
/allOf
-- i.e. when multiple elements of a JSON array have errors, we now do prefer showing errors from earlier elements rather than simply showing an error for the full array (#1250). - (Micro-)optimize equality checks when comparing for JSON Schema equality by first checking for object identity, as
==
would.
- Slightly speed up the
contains
keyword by removing some unnecessary validator (re-)creation.
- Fix the behavior of
enum
in the presence of0
or1
to properly considerTrue
andFalse
unequal (#1208). - Special case the error message for
{min,max}{Items,Length,Properties}
when they're checking for emptiness rather than true length.
- Properly consider items (and properties) to be evaluated by
unevaluatedItems
(resp.unevaluatedProperties
) when behind a$dynamicRef
as specified by the 2020 and 2019 specifications. jsonschema.exceptions.ErrorTree.__setitem__
is now deprecated. More broadly, in general users ofjsonschema
should never be mutating objects owned by the library.
- Fix the error message for additional items when used with heterogeneous arrays.
- Don't leak the
additionalItems
keyword into JSON Schema draft 2020-12, where it was replaced byitems
.
- Single label hostnames are now properly considered valid according to the
hostname
format. This is the behavior specified by the relevant RFC (1123). IDN hostname behavior was already correct.
- Importing the
Validator
protocol directly from the package root is deprecated. Import it fromjsonschema.protocols.Validator
instead. - Automatic retrieval of remote references (which is still deprecated) now properly succeeds even if the retrieved resource does not declare which version of JSON Schema it uses. Such resources are assumed to be 2020-12 schemas. This more closely matches the pre-referencing library behavior.
- Set a
jsonschema
specific user agent when automatically retrieving remote references (which is deprecated).
- Declare support for Py3.12
- Improve the hashability of wrapped referencing exceptions when they contain hashable data.
- Properly preserve
applicable_validators
in extended validators. Specifically, validators extending early drafts where siblings of$ref
were ignored will properly ignore siblings in the extended validator.
- Fix an additional regression with the deprecated
jsonschema.RefResolver
and pointer resolution.
- Fix a regression with
jsonschema.RefResolver
based resolution when used in combination with a custom validation dialect (viajsonschema.validators.create
).
This release majorly rehauls the way in which JSON Schema reference resolution is configured. It does so in a way that should be backwards compatible, preserving old behavior whilst emitting deprecation warnings.
jsonschema.RefResolver
is now deprecated in favor of the new referencing library.referencing
will begin in beta, but already is more compliant than the existing$ref
support. This change is a culmination of a meaningful chunk of work to make$ref
resolution more flexible and more correct. Backwards compatibility should be preserved for existing code which usesRefResolver
, though doing so is again now deprecated, and all such use cases should be doable using the new APIs. Please file issues on thereferencing
tracker if there is functionality missing from it, or here on thejsonschema
issue tracker if you have issues with existing code not functioning the same, or with figuring out how to change it to usereferencing
. In particular, this referencing change includes a change concerning automatic retrieval of remote references (retrievinghttp://foo/bar
automatically within a schema). This behavior has always been a potential security risk and counter to the recommendations of the JSON Schema specifications; it has survived this long essentially only for backwards compatibility reasons, and now explicitly produces warnings. Thereferencing
library itself will not automatically retrieve references if you interact directly with it, so the deprecated behavior is only triggered if you fully rely on the default$ref
resolution behavior and also include remote references in your schema, which will still be retrieved during the deprecation period (after which they will become an error).- Support for Python 3.7 has been dropped, as it is nearing end-of-life.
This should not be a "visible" change in the sense that
requires-python
has been updated, so users using 3.7 should still receivev4.17.3
when installing the library. - On draft 2019-09,
unevaluatedItems
now properly does not consider items to be evaluated by anadditionalItems
schema ifitems
is missing from the schema, as the specification says in this case thatadditionalItems
must be completely ignored. - Fix the
date
format checker on Python 3.11 (when format assertion behavior is enabled), where it was too liberal (#1076). - Speed up validation of
unevaluatedProperties
(#1075).
jsonschema.RefResolver
-- see above for details on the replacementjsonschema.RefResolutionError
-- see above for details on the replacement- relying on automatic resolution of remote references -- see above for details on the replacement
- importing
jsonschema.ErrorTree
-- instead import it viajsonschema.exceptions.ErrorTree
- importing
jsonschema.FormatError
-- instead import it viajsonschema.exceptions.FormatError
- Fix instantiating validators with cached refs to boolean schemas rather than objects (#1018).
- Empty strings are not valid relative JSON Pointers (aren't valid under the RJP format).
- Durations without (trailing) units are not valid durations (aren't
valid under the duration format). This involves changing the dependency
used for validating durations (from
isoduration
toisodate
).
- The error message when using
unevaluatedProperties
with a non-trivial schema value (i.e. something other thanfalse
) has been improved (#996).
- The
check_schema
method onjsonschema.protocols.Validator
instances now enables format validation by default when run. This can catch some additional invalid schemas (e.g. containing invalid regular expressions) where the issue is indeed uncovered by validating against the metaschema with format validation enabled as an assertion. - The
jsonschema
CLI (along withjsonschema.cli
the module) are now deprecated. Usecheck-jsonschema
instead, which can be installed viapip install check-jsonschema
and found here.
- Make
ErrorTree
have a more grammatically correctrepr
.
- Improve the base URI behavior when resolving a
$ref
to a resolution URI which is different from the resolved schema's declared$id
. - Accessing
jsonschema.draftN_format_checker
is deprecated. Instead, if you want access to the format checker itself, it is exposed asjsonschema.validators.DraftNValidator.FORMAT_CHECKER
on anyjsonschema.protocols.Validator
.
- A specific API Reference page is now present in the documentation.
$ref
on earlier drafts (specifically draft 7 and 6) has been "fixed" to follow the specified behavior when present alongside a sibling$id
. Specifically the ID is now properly ignored, and references are resolved against whatever resolution scope was previously relevant.
FormatChecker.cls_checks
is deprecated. UseFormatChecker.checks
on an instance ofFormatChecker
instead.unevaluatedItems
has been fixed for draft 2019. It's nonetheless discouraged to use draft 2019 for any schemas, new or old.- Fix a number of minor annotation issues in
protocols.Validator
- Add support for creating validator classes whose metaschema uses a different dialect than its schemas. In other words, they may use draft2020-12 to define which schemas are valid, but the schemas themselves use draft7 (or a custom dialect, etc.) to define which instances are valid. Doing this is likely not something most users, even metaschema authors, may need, but occasionally will be useful for advanced use cases.
- Fix some stray comments in the README.
- Warn at runtime when subclassing validator classes. Doing so was not intended to be public API, though it seems some downstream libraries do so. A future version will make this an error, as it is brittle and better served by composing validator objects instead. Feel free to reach out if there are any cases where changing existing code seems difficult and I can try to provide guidance.
- Make the rendered README in PyPI simpler and fancier. Thanks Hynek (#983)!
jsonschema.validators.validator_for
now properly uses the explicitly provided default validator even if the$schema
URI is not found.
- Fix a second place where subclasses may have added attrs attributes (#982).
- Fix Validator.evolve (and APIs like
iter_errors
which call it) for cases where the validator class has been subclassed. Doing so wasn't intended to be public API, but given it didn't warn or raise an error it's of course understandable. The next release however will make it warn (and a future one will make it error). If you need help migrating usage of inheriting from a validator class feel free to open a discussion and I'll try to give some guidance (#982).
- Add support for referencing schemas with
$ref
across different versions of the specification than the referrer's
- Update some documentation examples to use newer validator releases in their sample code.
- Fix relative
$ref
resolution when the base URI is a URN or other scheme (#544). pkgutil.resolve_name
is now used to retrieve validators provided on the command line. This function is only available on 3.9+, so 3.7 and 3.8 (which are still supported) now rely on the pkgutil_resolve_name backport package. Note however that the CLI itself is due to be deprecated shortly in favor of check-jsonschema.
best_match
no longer traverses intoanyOf
andoneOf
when all of the errors within them seem equally applicable. This should lead to clearer error messages in some cases where no branches were matched.
- Also have
best_match
handle cases where thetype
validator is an array.
- Minor tweak of the PyPI hyperlink names
- Enhance
best_match
to prefer errors from branches of the schema which match the instance's type (#728)
- Fix a number of minor typos in docstrings, mostly private ones (#969)
- Gut the (incomplete) implementation of
recursiveRef
on draft 2019. It needs completing, but for now can lead to recursion errors (e.g. #847).
- Fix
unevaluatedProperties
andunevaluatedItems
for types they should ignore (#949) jsonschema
now uses hatch for its build process. This should be completely transparent to end-users (and only matters to contributors).
- Revert changes to
$dynamicRef
which caused a performance regression in v4.5.0
- Validator classes for each version now maintain references to the correct corresponding format checker (#905)
- Development has moved to a GitHub organization. No functional behavior changes are expected from the change.
- Add
mypy
support (#892) - Add support for Python 3.11
- Properly report deprecation warnings at the right stack level (#899)
- Additional performance improvements for resolving refs (#896)
- Resolving refs has had performance improvements (#893)
- Fix undesired fallback to brute force container uniqueness check on certain input types (#893)
- Implement a PEP544 Protocol for validator classes (#890)
- Pin
importlib.resources
from below (#877)
- Use
importlib.resources
to load schemas (#873) - Ensure all elements of arrays are verified for uniqueness by
uniqueItems
(#866)
- Fix
dependentSchemas
to properly consider non-object instances to be valid (#850)
- Fix
prefixItems
not indicating which item was invalid within the instance path (#862)
- Add Python 3.10 to the list of supported Python versions
- Fix the declaration of minimum supported Python version (#846)
- Partial support for Draft 2020-12 (as well as 2019-09). Thanks to Thomas Schmidt and Harald Nezbeda.
False
and0
are now properly considered non-equal even recursively within a container (#686). As part of this change,uniqueItems
validation may be slower in some cases. Please feel free to report any significant performance regressions, though in some cases they may be difficult to address given the specification requirement.- The CLI has been improved, and in particular now supports a
--output
option (withplain
(default) orpretty
arguments) to control the output format. Future work may add additional machine-parsable output formats. - Code surrounding
DEFAULT_TYPES
and the legacy mechanism for specifying types to validators have been removed, as per the deprecation policy. Validators should use theTypeChecker
object to customize the set of Python types corresponding to JSON Schema types. - Validation errors now have a
json_path
attribute, describing their location in JSON path format - Support for the IP address and domain name formats has been improved
- Support for Python 2 and 3.6 has been dropped, with
python_requires
properly set. multipleOf
could overflow when given sufficiently large numbers. Now, when an overflow occurs,jsonschema
will fall back to using fraction division (#746).jsonschema.__version__
,jsonschema.validators.validators
,jsonschema.validators.meta_schemas
andjsonschema.RefResolver.in_scope
have been deprecated, as has passing a second-argument schema toValidator.iter_errors
andValidator.is_valid
.
- Added a
format_nongpl
setuptools extra, which installs onlyformat
dependencies that are non-GPL (#619).
- Temporarily revert the switch to
js-regex
until #611 and #612 are resolved.
- Regular expressions throughout schemas now respect the ECMA 262 dialect, as recommended by the specification (#609).
- Fixed a bug where
0
andFalse
were considered equal byconst
andenum
(#575).
- Fixed a bug where extending validators did not preserve their notion
of which validator property contains
$id
information.
- Support for Draft 6 and Draft 7
- Draft 7 is now the default
- New
TypeChecker
object for more complex type definitions (and overrides) - Falling back to isodate for the date-time format checker is no longer attempted, in accordance with the specification
- Support for Python 2.6 has been dropped.
- Improve a few error messages for
uniqueItems
(#224) andadditionalProperties
(#317) - Fixed an issue with
ErrorTree
's handling of multiple errors (#288)
- Improved performance on CPython by adding caching around ref resolution (#203)
- Added a CLI (#134)
- Added absolute path and absolute schema path to errors (#120)
- Added
relevance
- Meta-schemas are now loaded via
pkgutil
- Added
by_relevance
andbest_match
(#91) - Fixed
format
to allow adding formats for non-strings (#125) - Fixed the
uri
format to reject URI references (#131)
- Compile the host name regex (#127)
- Allow arbitrary objects to be types (#129)
- Support RFC 3339 datetimes in conformance with the spec
- Fixed error paths for additionalItems + items (#122)
- Fixed wording for min / maxProperties (#117)
- Added
create
andextend
tojsonschema.validators
- Removed
ValidatorMixin
- Fixed array indices ref resolution (#95)
- Fixed unknown scheme defragmenting and handling (#102)
- Better error tracebacks (#83)
- Raise exceptions in
ErrorTree
s for keys not in the instance (#92) - __cause__ (#93)
- More attributes for ValidationError (#86)
- Added
ValidatorMixin.descend
- Fixed bad
RefResolutionError
message (#82)
- Canonicalize URIs (#70)
- Allow attaching exceptions to
format
errors (#77)
- Support for Draft 4
- Support for format
- Longs are ints too!
- Fixed a number of issues with
$ref
support (#66) - Draft4Validator is now the default
ValidationError.path
is now in sequential order- Added
ValidatorMixin
Full support for JSON References
validates
for registering new validatorsDocumentation
Bugfixes
- uniqueItems not so unique (#34)
- Improper any (#47)
- Partial support for (JSON Pointer)
$ref
- Deprecations
Validator
is replaced byDraft3Validator
with a slightly different interfacevalidator(meta_validate=False)
- Bugfixes
- Issue #30 - Wrong behavior for the dependencies property validation
- Fixed a miswritten test
- Bugfixes
- Issue #17 - require path for error objects
- Issue #18 - multiple type validation for non-objects
Preliminary support for programmatic access to error details (Issue #5). There are certainly some corner cases that don't do the right thing yet, but this works mostly.
In order to make this happen (and also to clean things up a bit), a number of deprecations are necessary:
stop_on_error
is deprecated inValidator.__init__
. UseValidator.iter_errors()
instead.number_types
andstring_types
are deprecated there as well. Usetypes={"number" : ..., "string" : ...}
instead.meta_validate
is also deprecated, and instead is now accepted as an argument tovalidate
,iter_errors
andis_valid
.
A bugfix or two
- Default for unknown types and properties is now to not error (consistent with the schema).
- Python 3 support
- Removed dependency on SecureTypes now that the hash bug has been resolved.
- "Numerous bug fixes" -- most notably, a divisibleBy error for floats and a bunch of missing typechecks for irrelevant properties.