Versions follow CalVer with a strict backwards-compatibility policy.
The first number of the version is the year. The second number is incremented with each release, starting at 1 for each year. The third number is when we need to start branches for older releases (only for emergencies).
Put simply, you shouldn’t ever be afraid to upgrade
attrs if you’re only using its public APIs.
Whenever there is a need to break compatibility, it is announced here in the changelog, and raises a
DeprecationWarning for a year (if possible) before it’s finally really broken.
The structure of the
attrs.Attribute class is exempt from this rule.
It will change in the future, but since it should be considered read-only, that shouldn’t matter.
However if you intend to build extensions on top of
attrs you have to anticipate that.
Python 2.7 is not supported anymore.
Dealing with Python 2.7 tooling has become too difficult for a volunteer-run project.
We have supported Python 2 more than 2 years after it was officially discontinued and feel that we have paid our dues. All version up to 21.4.0 from December 2021 remain fully functional, of course. #936
attrs.Attributehas been removed. This does not affect the cmp argument to
attr.sthat can be used as a shortcut to set eq and order at the same time. #939
Instantiation of frozen slotted classes is now faster. #898
eqkey is defined, it is also used before hashing the attribute. #909
attrs.validators.deep_iterable()’s member_validator argument now also accepts a list of validators and wraps them in an
Added missing type stub re-imports for
Added missing stub for
ValueErroris not missing the attribute, expected options, and the value it got anymore. #951
Python 3.11 is now officially supported. #969
@define, converters are now run by default when setting an attribute on an instance – additionally to validators. I.e. the new default is
This is unfortunately a breaking change, but it was an oversight, impossible to raise a
DeprecationWarningabout, and it’s better to fix it now while the APIs are very fresh with few users. #835, #886
import attrshas finally landed! As of this release, you can finally import
attrsusing its proper name.
Not all names from the
attrnamespace have been transferred; most notably
attr.ibare missing. See
attrs.fieldif you haven’t seen our next-generation APIs yet. A more elaborate explanation can be found On The Core API Names
This feature is at least for one release provisional. We don’t plan on changing anything, but such a big change is unlikely to go perfectly on the first strike.
The API docs have been mostly updated, but it will be an ongoing effort to change everything to the new APIs. Please note that we have not moved – or even removed – anything from
Please do report any bugs or documentation inconsistencies! #887
__match_args__are now generated to support Python 3.10’s Structural Pattern Matching. This can be controlled by the
match_argsargument to the class decorators on Python 3.10 and later. On older versions, it is never added and the argument is ignored. #815
If the class-level on_setattr is set to
@mutable) but no field defines a validator, pretend that it’s not set. #817
__repr__is significantly faster on Pythons with f-strings. #819
Attributes transformed via
field_transformerare wrapped with
Generated source code is now cached more efficiently for identical classes. #828
Added new validators:
gt(val)(> val), and
Added new context manager
attrs.validators.(set|get)_disabled(). They deprecate
attrs.(set|get)_run_validators(). All functions are interoperable and modify the same internal state. They are not – and never were – thread-safe, though. #859
attrs.validators.matches_re()now accepts pre-compiled regular expressions in addition to pattern strings. #877
We had to revert the recursive feature for
attr.evolve()because it broke some use-cases – sorry! #806
Python 3.4 is now blocked using packaging metadata because
attrscan’t be imported on it anymore. To ensure that 3.4 users can keep installing
attrseasily, we will yank 21.1.0 from PyPI. This has no consequences if you pin
attrsto 21.1.0. #807
The long-awaited, much-talked-about, little-delivered
import attrsis finally upon us!
Since the NG APIs have now been proclaimed stable, the next release of
attrswill allow you to actually
import attrs. We’re taking this opportunity to replace some defaults in our APIs that made sense in 2015, but don’t in 2021.
So please, if you have any pet peeves about defaults in
attrs’s APIs, now is the time to air your grievances in #487! We’re not gonna get such a chance for a second time, without breaking our backward-compatibility guarantees, or long deprecation cycles. Therefore, speak now or forever hold you peace! #487
The cmp argument to
attr.ib()has been undeprecated It will continue to be supported as syntactic sugar to set eq and order in one go.
I’m terribly sorry for the hassle around this argument! The reason we’re bringing it back is it’s usefulness regarding customization of equality/ordering.
cmpattribute and argument on
attr.Attributeremains deprecated and will be removed later this year. #773
The instant favorite next-generation APIs are not provisional anymore!
They are also officially supported by Mypy as of their 0.800 release.
If an attribute defines a converter, the type of its parameter is used as type annotation for its corresponding
attr.converters.pipeis used, the first one’s is used. #710
Fixed the creation of an extra slot for an
attr.ibwhen the parent class already has a slot with the same name. #718
__attrs__init__()will now be injected if
init=False, or if
auto_detect=Trueand a user-defined
This enables users to do “pre-init” work in their
__init__()can then delegate constructor argument processing to
self.__attrs_init__(*args, **kwargs). #731
It’s now possible to use
super()inside of properties of slotted classes. #747
Allow for a
__attrs_pre_init__()method that – if defined – will get called at the beginning of the
attr.Attribute.evolve()to type stubs. #752
attrs.evolve()now works recursively with nested
Python 3.10 is now officially supported. #763
attr.resolve_types()now takes an optional attrib argument to work inside a
To make it easier to customize attribute comparison (#435), we have added the
Added provisional support for static typing in
pyrightvia the dataclass_transforms specification. Both the
attrsimplementation may change in future versions of both projects.
This release does not change anything about them and they are already used widely in production though.
If you wish to use them together with mypy, you can simply drop this plugin into your project.
Feel free to provide feedback to them in the linked issue #668.
We will release the
attrsnamespace once we have the feeling that the APIs have properly settled. #668
attr.s()now has a field_transformer hook that is called for all
Attributes and returns a (modified or updated) list of
attr.asdict()has a value_serializer hook that can change the way values are converted. Both hooks are meant to help with data (de-)serialization workflows. #653
kw_only=Truenow works on Python 2. #700
sets with regards to the retain_collection_types argument. #704
The type stubs for
attr.make_class()are not missing the collect_by_mro argument anymore. #711
This release fixes a bunch of bugs and ergonomics but they remain mostly unchanged.
If you wish to use them together with mypy, you can simply drop this plugin into your project.
Feel free to provide feedback to them in the linked issue #668.
We will release the
attrsnamespace once we have the feeling that the APIs have properly settled. #668
attr.define()et al now correct detect
attr.define()et al’s hybrid behavior now also works correctly when arguments are passed. #675
It’s possible to define custom
__setattr__methods on slotted classes again. #681
In 20.1.0 we introduced the
inheritedattribute on the
attr.Attributeclass to differentiate attributes that have been inherited and those that have been defined directly on the class.
It has shown to be problematic to involve that attribute when comparing instances of
attr.Attributethough, because when sub-classing, attributes from base classes are suddenly not equal to themselves in a super class.
inheritedattribute will now be ignored when hashing and comparing instances of
zope.interfaceis now a “soft dependency” when running the test suite; if
zope.interfaceis not installed when running the test suite, the interface-related tests will be automatically skipped. #685
The ergonomics of creating frozen classes using
@define(frozen=True)and sub-classing frozen classes has been improved: you don’t have to set
Python 3.4 is not supported anymore. It has been unsupported by the Python core team for a while now, its PyPI downloads are negligible, and our CI provider removed it as a supported option.
It’s very unlikely that
attrswill break under 3.4 anytime soon, which is why we do not block its installation on Python 3.4. But we don’t test it anymore and will block it once someone reports breakage. #608
Less of a deprecation and more of a heads up: the next release of
attrswill introduce an
attrsnamespace. That means that you’ll finally be able to run
import attrswith new functions that aren’t cute abbreviations and that will carry better defaults.
This should not break any of your code, because project-local packages have priority before installed ones. If this is a problem for you for some reason, please report it to our bug tracker and we’ll figure something out.
attrnamespace isn’t going anywhere and its defaults are not changing – this is a purely additive measure. Please check out the linked issue for more details.
attr.resolve_types(). It ensures that all forward-references and types in string form are resolved into concrete types.
@attr.s(collect_by_mro=False)argument that if set to
Truefixes the collection of attributes from base classes.
It’s only necessary for certain cases of multiple-inheritance but is kept off for now for backward-compatibility reasons. It will be turned on by default in the future.
On Python 3, all generated methods now have a docstring explaining that they have been created by
It is now possible to prevent
attrsfrom auto-generating the
__getstate__methods that are required for pickling of slotted classes.
@attr.s(auto_detect=True)and implement them yourself: if
attrsfinds either of the two methods directly on the decorated class, it assumes implicitly
getstate_setstate=False(and implements neither).
ValueError: Cell is emptybug that could happen in some rare edge cases. #590
attrscan now automatically detect your own implementations and infer
hash=Falseif you set
attrswill ignore inherited methods. If the argument implies more than one method (e.g.
__ne__), it’s enough for one of them to exist and
attrswill create neither.
This feature requires Python 3. #607
attr.converters.pipe(). The feature allows combining multiple conversion callbacks into one by piping the value through all of them, and retuning the last result.
As part of this feature, we had to relax the type information for converter callables. #618
Fixed serialization behavior of non-slots classes with
cache_hash=True. The hash cache will be cleared on operations which make “deep copies” of instances of classes with hash caching, though the cache will not be cleared with shallow copies like those made by
copy.deepcopy()or serialization and deserialization with
picklewould result in an un-initialized object.
It is now possible to specify hooks that are called whenever an attribute is set after a class has been instantiated.
You can pass
@attr.s()to set the default for all attributes on a class, and to
@attr.ib()to overwrite it for individual attributes.
Provisional APIs called
attr.frozen()have been added.
They are only available on Python 3.6 and later, and call
attr.s()with different default values.
If nothing comes up, they will become the official way for creating classes in 20.2.0 (see above).
Please note that it may take some time until mypy – and other tools that have dedicated support for
attrs– recognize these new APIs. Please do not open issues on our bug tracker, there is nothing we can do about it. #666
We have also provisionally added
attr.ib(). It also requires at least Python 3.6 and is keyword-only. Other than that, it only dropped a few arguments, but changed no defaults.
attr.ib()is not going anywhere. #669
auto_attribsusage when default values cannot be compared directly with
==, such as
__ge__do not consider subclasses comparable anymore.
This has been deprecated since 18.2.0 and was raising a
DeprecationWarningfor over a year. #570
attr.ib()is now deprecated.
eqto add equality methods (
orderto add ordering methods (
__ge__) instead – just like with dataclasses.
Both are effectively
Trueby default but it’s enough to set
eq=Falseto disable both at once. Passing
eq=False, order=Trueexplicitly will raise a
Since this is arguably a deeper backward-compatibility break, it will have an extended deprecation period until 2021-06-01. After that day, the
cmpargument will be removed.
attr.Attributealso isn’t orderable anymore. #574
Slotted classes now use a pure Python mechanism to rewrite the
__class__cell when rebuilding the class, so
super()works even on environments where
ctypesis not installed. #522
attr.validators.deep_mapping()type stubs. #533
attr.validators.is_callable()validator now raises an exception
attr.exceptions.NotCallableError, a subclass of
TypeError, informing the received value. #536
attr.validators.matches_re()that checks string attributes whether they match a regular expression. #552
Keyword-only attributes (
kw_only=True) and attributes that are excluded from the
init=False) now can appear before mandatory attributes. #559
The fake filename for generated methods is now more stable. It won’t change when you restart the process. #560
The value passed to
@attr.ib(repr=…)can now be either a boolean (as before) or a callable. That callable must return a string and is then used for formatting the attribute by the generated
attr.__version_info__that can be used to reliably check the version of
attrsand write forward- and backward-compatible code. Please check out the section on deprecated APIs on how to use it. #580
Fixed a bug where deserialized objects with
cache_hash=Truecould have incorrect hash code values. This change breaks classes with
cache_hash=Truewhen a custom
__setstate__is present. An exception will be thrown when applying the
attrsannotation to such a class. This limitation is tracked in issue #494. #482
is_callable: validates that a value is callable
deep_iterable: Allows recursion down into an iterable, applying another validator to every member in the iterable as well as applying an optional validator to the iterable itself.
deep_mapping: Allows recursion down into the items in a mapping object, applying a key validator and a value validator to the key and value in every item. Also applies an optional validator to the mapping object itself.
You can find them in the
Fixed stub files to prevent errors raised by mypy’s
disallow_any_generics = Trueoption. #443
init=Falsenow can follow after
attrsnow has first class support for defining exception classes.
If you define a class using
@attr.s(auto_exc=True)and subclass an exception, the class will behave like a well-behaved exception class including an appropriate
__str__method, and all attributes additionally available in an
Clarified documentation for hashing to warn that hashable objects should be deeply immutable (in their usage, even if this is not enforced). #503
Comparing subclasses using
>=is now deprecated. The docs always claimed that instances are only compared if the types are identical, so this is a first step to conform to the docs.
Equality operators (
!=) were always strict in this regard. #394
Added kw_only arguments to
attr.s, and a corresponding kw_only attribute to
attr.Attribute. This change makes it possible to have a generated
__init__with keyword-only arguments on Python 3, relaxing the required ordering of default and non-default valued attributes. #281, #411
attr.validators.in_()now raises a
ValueErrorwith a useful message even if the options are a string and the value is not a string. #383
attr.asdict()now properly handles deeply nested lists and dictionaries. #395
Fixed a reference leak where the original class would remain live after being replaced when
slots=Trueis set. #407
Slotted classes can now be made weakly referenceable by passing
Added cache_hash option to
@attr.swhich causes the hash code to be computed once and stored on the object. #426
Attributes can be named
It is now possible to override a base class’ class variable using only class annotations. #431
x=X(); x.cycle = x; repr(x)will no longer raise a
RecursionError, and will instead show as
attr.ib(factory=f)is now syntactic sugar for the common case of
attr.field_dict()to return an ordered dictionary of
attrsattributes for a class, whose keys are the attribute names.
The order of attributes that are passed into
attr.make_class()or the these argument of
@attr.s()is now retained if the dictionary is ordered (i.e.
dicton Python 3.6 and later,
Before, the order was always determined by the order in which the attributes have been defined which may not be desirable when creating classes programmatically.
In slotted classes,
__setstate__now ignore the
Setting the cell type is now completely best effort. This fixes
We cannot make any guarantees regarding Jython though, because our test suite cannot run due to dependency incompatibilities.
attr.sis passed a these argument, it will no longer attempt to remove attributes with the same name from the class body.
The hash of
attr.NOTHINGis now vegan and faster on 32bit Python builds.
The overhead of instantiating frozen dict classes is virtually eliminated. #336
__init__methods now have an
__annotations__attribute derived from the types of the fields.
We have restructured the documentation a bit to account for
attrs’ growth in scope. Instead of putting everything into the examples page, we have started to extract narrative chapters.
Expect more to come!
The traversal of MROs when using multiple inheritance was backward: If you defined a class
attrswould have collected the attributes from
Bbefore those of
This is now fixed and means that in classes that employ multiple inheritance, the output of
__repr__and the order of positional arguments in
__init__changes. Because of the nature of this bug, a proper deprecation cycle was unfortunately impossible.
Generally speaking, it’s advisable to prefer
kwargs-based initialization anyways – especially if you employ multiple inheritance and diamond-shaped hierarchies.
attrsno longer produces an
AttributeErrorwhen the instance is missing some of the specified attributes (either through deleting or after using
init=Falseon some attributes).
This can break code that relied on
AttributeErrorto check if any
attrs-specified members were unset.
If you were using this, you can implement a custom method for checking this:
def has_unset_members(self): for field in attr.fields(type(self)): try: getattr(self, field.name) except AttributeError: return True return False
attr.ib(convert=callable)option is now deprecated in favor of
This is done to achieve consistency with other noun-based arguments like validator.
convert will keep working until at least January 2019 while raising a
__hash__methods now hash the class type along with the attribute values. Until now the hashes of two classes with the same values were identical which was a bug.
The generated method is also much faster now.
attr.ib’s metadata argument now defaults to a unique empty
dictinstance instead of sharing a common empty
dictfor all. The singleton empty
dictis still enforced.
ctypesis optional now however if it’s missing, a bare
super()will not work in slotted classes. This should only happen in special environments like Google App Engine.
The attribute redefinition feature introduced in 17.3.0 now takes into account if an attribute is redefined via multiple inheritance. In that case, the definition that is closer to the base of the class hierarchy wins.
auto_attribs=Truecan be empty now.
Equality tests are much faster now.
All generated methods now have correct
__name__, and (on Python 3)
Attributes are no longer defined on the class body.
This means that if you define a class
Cwith an attribute
x, the class will not have an attribute
xfor introspection. Instead of
attr.fields(C).xor look at
C.__attrs_attrs__. The old behavior has been deprecated since version 16.1. (#253)
Added type argument to
This change paves the way for automatic type checking and serialization (though as of this release
attrsdoes not make use of it). In Python 3.6 or higher, the value of
attr.Attribute.typecan alternately be set using variable type annotations (see PEP 526). (#151, #214, #215, #239)
The combination of
slots=Truenow works on Python 2. (#198)
attr.Factoryis hashable again. (#204)
Subclasses now can overwrite attribute definitions of their base classes.
Added new option auto_attribs to
@attr.sthat allows to collect annotated fields without setting them to
Setting a field to an
attr.ib()is still possible to supply options like validators. Setting it to any other value is treated like it was passed as
attr.ib(default=value)– passing an instance of
attr.Factoryalso works as expected. (#262, #277)
Instances of classes created using
attr.make_class()can now be pickled. (#282)
Validators are hashable again. Note that validators may become frozen in the future, pending availability of no-overhead frozen classes. #192
To encourage more participation, the project has also been moved into a dedicated GitHub organization and everyone is most welcome to join!
attrs also has a logo now!
attrswill set the
Noneby default now. The way hashes were handled before was in conflict with Python’s specification. This may break some software although this breakage is most likely just surfacing of latent bugs. You can always make
@attr.s(hash=True). See #136 for the rationale of this change.
Please do not upgrade blindly and do test your software! Especially if you use instances as dict keys or put them into sets!
attr.ib’s hash argument is
Noneby default too and mirrors the cmp argument as it should.
attr.assoc()is now deprecated in favor of
attr.evolve()and will stop working in 2018.
attr.evolve()that, given an instance of an
attrsclass and field changes as keyword arguments, will instantiate a copy of the given instance with the changes applied.
assoc(), which is now deprecated.
evolve()is significantly faster than
assoc(), and requires the class have an initializer that can take the field values as keyword arguments (like
attrsitself can generate). #116 #124 #135
FrozenInstanceErroris now raised when trying to delete an attribute from a frozen class. #118
Frozen-ness of classes is now inherited. #128
__attrs_post_init__()is now run if validation is disabled. #130
attr.validators.in_(options)that, given the allowed
options, checks whether the attribute value is in it. This can be used to check constants, enums, mappings, etc. #181
attr.validators.and_()that composes multiple validators into one. #161
For convenience, the validator argument of
@attr.snow can take a list of validators that are wrapped using
attr.validators.optional()now can take a list of validators too. #161
attr.Factory()now has a takes_self argument that makes the initializer to pass the partially initialized instance into the factory. In other words you can define attribute defaults based on other attributes. #165 #189
Default factories can now also be defined inline using decorators. They are always passed the partially initialized instance. #165
attr.make_class()now accepts the keyword argument
baseswhich allows for subclassing. #152
Metaclasses are now preserved with
Attributes now can have user-defined metadata which greatly improves
attrs’s extensibility. #96
Allow for a
__attrs_post_init__()method that – if defined – will get called at the end of the
@attr.s(str=True)that will optionally create a
__str__()method that is identical to
__repr__(). This is mainly useful with
Exceptions and other classes that rely on a useful
__str__()implementation but overwrite the default one through a poor own one. Default Python class behavior is to use
If you tried using
Exceptions and were puzzled by the tracebacks: this option is for you.
__name__is no longer overwritten with
attr.astuple()that – similarly to
attr.asdict()– returns the instance as a tuple. #77
Converters now work with frozen classes. #76
attrsclasses with converters is now significantly faster. #80
Pickling now works with slotted classes. #81
attr.assoc()now works with slotted classes. #84
The tuple returned by
attr.fields()now also allows to access the
Attributeinstances by name. Yes, we’ve subclassed
tupleso you don’t have to! Therefore
attr.fields(C).xis equivalent to the deprecated
C.xand works with slotted classes. #88
All instances where function arguments were called
clhave been changed to the more Pythonic
cls. Since it was always the first argument, it’s doubtful anyone ever called those function with in the keyword form. If so, sorry for any breakage but there’s no practical deprecation path to solve this ugly wart.
Attributeinstances on class objects is now deprecated and will stop working in 2017. If you need introspection please use the
__attrs_attrs__attribute or the
attr.fields()function that carry them too. In the future, the attributes that are defined on the class body and are usually overwritten in your
__init__method are simply removed after
@attr.shas been applied.
This will remove the confusing error message if you write your own
__init__and forget to initialize some attribute. Instead you will get a straightforward
AttributeError. In other words: decorated classes will work more like plain Python classes which was always
The serious-business aliases
attr.attrhave been deprecated in favor of
attr.attribwhich are much more consistent and frankly obvious in hindsight. They will be purged from documentation immediately but there are no plans to actually remove them.
dict_factoryarguments is now propagated on recursion. #45
attr.attribas a more consistent aliases for
Add frozen option to
attr.sthat will make instances best-effort immutable. #60
retain_collection_typesas an argument. If
True, it does not convert attributes of type
Python 3.3 and 2.6 are no longer supported. They may work by chance but any effort to keep them working has ceased.
The last Python 2.6 release was on October 29, 2013 and is no longer supported by the CPython core team. Major Python packages like Django and Twisted dropped Python 2.6 a while ago already.
Python 3.3 never had a significant user base and wasn’t part of any distribution’s LTS release.
Allow the case of initializing attributes that are set to
init=False. This allows for clean initializer parameter lists while being able to initialize attributes to default values. #32
attr.asdict()can now produce arbitrary mappings instead of Python
dicts when provided with a
Multiple performance improvements.