Background

PEP 472 was opened in 2014. The PEP detailed various use cases and was created by extracting implementation strategies from a broad discussion on the python-ideas mailing list, although no clear consensus was reached on which strategy should be used. Many corner cases have been examined more closely and felt awkward, backward incompatible or both.

The PEP was eventually rejected in 2019 1 mostly due to lack of interest for the feature despite its 5 years of existence.

However, with the introduction of type hints in PEP 484 2 the square bracket notation has been used consistently to enrich the typing annotations, e.g. to specify a list of integers as Sequence[int]. Additionally, there has been an expanded growth of packages for data analysis such as pandas and xarray, which use names to describe columns in a table (pandas) or axis in an nd-array (xarray). These packages allow users to access specific data by names, but cannot currently use index notation ([]) for this functionality.

As a result, a renewed interest in a more flexible syntax that would allow for named information has been expressed occasionally in many different threads on python-ideas, recently by Caleb Donovick 3 in 2019 and Andras Tantos 4 in 2020. These requests prompted a strong activity on the python-ideas mailing list, where the various options have been re-discussed and a general consensus on an implementation strategy has now been reached.

Use cases

The following practical use cases present different cases where a keyword specification would improve notation and provide additional value:

1. To provide a more communicative meaning to the index, preventing e.g. accidental inversion of indexes:

   >>> grid_position[x=3, y=5, z=8]
   >>> rain_amount[time=0:12, location=location]
   >>> matrix[row=20, col=40]
   

2. To enrich the typing notation with keywords, especially during the use of generics:

   def function(value: MyType[T=int]):
   

3. In some domain, such as computational physics and chemistry, the use of a notation such as Basis[Z=5] is a Domain Specific Language notation to represent a level of accuracy:

   >>> low_accuracy_energy = computeEnergy(molecule, BasisSet[Z=3])
   

4. Pandas currently uses a notation such as:

   >>> df[df['x'] == 1]
   

   which could be replaced with df[x=1].

5. xarray has named dimensions. Currently these are handled with functions .isel:

   >>> data.isel(row=10)  # Returns the tenth row
   

   which could also be replaced with data[row=10]. A more complex example:

   >>> # old syntax
   >>> da.isel(space=0, time=slice(None, 2))[...] = spam
   >>> # new syntax
   >>> da[space=0, time=:2] = spam
   

   Another example:

   >>> # old syntax
   >>> ds["empty"].loc[dict(lon=5, lat=6)] = 10
   >>> # new syntax
   >>> ds["empty"][lon=5, lat=6] = 10
   
   >>> # old syntax
   >>> ds["empty"].loc[dict(lon=slice(1, 5), lat=slice(3, None))] = 10
   >>> # new syntax
   >>> ds["empty"][lon=1:5, lat=6:] = 10
   

6. Functions/methods whose argument is another function (plus its arguments) need some way to determine which arguments are destined for the target function, and which are used to configure how they run the target. This is simple (if non-extensible) for positional parameters, but we need some way to distinguish these for keywords. 6

   An indexed notation would afford a Pythonic way to pass keyword arguments to these functions without cluttering the caller’s code.

   >>> # Let's start this example with basic syntax without keywords.
   >>> # the positional values are arguments to `func` while
   >>> # `name=` is processed by `trio.run`.
   >>> trio.run(func, value1, value2, name="func")
   >>> # `trio.run` ends up calling `func(value1, value2)`.
   
   >>> # If we want/need to pass value2 by keyword (keyword-only argument,
   >>> # additional arguments that won't break backwards compatibility ...),
   >>> # currently we need to resort to functools.partial:
   >>> trio.run(functools.partial(func, param2=value2), value1, name="func")
   >>> trio.run(functools.partial(func, value1, param2=value2), name="func")
   
   >>> # One possible workaround is to convert `trio.run` to an object
   >>> # with a `__call__` method, and use an "option" helper,
   >>> trio.run.option(name="func")(func, value1, param2=value2)
   >>> # However, foo(bar)(baz) is uncommon and thus disruptive to the reader.
   >>> # Also, you need to remember the name of the `option` method.
   
   >>> # This PEP allows us to replace `option` with `__getitem__`.
   >>> # The call is now shorter, more mnemonic, and looks+works like typing
   >>> trio.run[name="func"](func, value1, param2=value2)
   

7. Availability of star arguments would benefit PEP-646 Variadic Generics 7, especially in the forms a[*x] and a[*x, *y, p, q, *z]. The PEP details exactly this notation in its “Unpacking: Star Operator” section.

It is important to note that how the notation is interpreted is up to the implementation. This PEP only defines and dictates the behavior of Python regarding passed keyword arguments, not how these arguments should be interpreted and used by the implementing class.

Current status of indexing operation

Before detailing the new syntax and semantics to the indexing notation, it is relevant to analyse how the indexing notation works today, in which contexts, and how it is different from a function call.

Subscripting obj[x] is, effectively, an alternate and specialised form of function call syntax with a number of differences and restrictions compared to obj(x). The current Python syntax focuses exclusively on position to express the index, and also contains syntactic sugar to refer to non-punctiform selection (slices). Some common examples:

>>> a[3]       # returns the fourth element of 'a'
>>> a[1:10:2]  # slice notation (extract a non-trivial data subset)
>>> a[3, 2]    # multiple indexes (for multidimensional arrays)

This translates into a __(get|set|del)item__ dunder call which is passed a single parameter containing the index (for __getitem__ and __delitem__) or two parameters containing index and value (for __setitem__).

The behavior of the indexing call is fundamentally different from a function call in various aspects:

The first difference is in meaning to the reader. A function call says “arbitrary function call potentially with side-effects”. An indexing operation says “lookup”, typically to point at a subset or specific sub-aspect of an entity (as in the case of typing notation). This fundamental difference means that, while we cannot prevent abuse, implementors should be aware that the introduction of keyword arguments to alter the behavior of the lookup may violate this intrinsic meaning.

The second difference of the indexing notation compared to a function is that indexing can be used for both getting and setting operations. In Python, a function cannot be on the left hand side of an assignment. In other words, both of these are valid:

>>> x = a[1, 2]
>>> a[1, 2] = 5

but only the first one of these is valid:

>>> x = f(1, 2)
>>> f(1, 2) = 5  # invalid

This asymmetry is important, and makes one understand that there is a natural imbalance between the two forms. It is therefore not a given that the two should behave transparently and symmetrically.

The third difference is that functions have names assigned to their arguments, unless the passed parameters are captured with *args, in which case they end up as entries in the args tuple. In other words, functions already have anonymous argument semantic, exactly like the indexing operation. However, __(get|set|del)item__ is not always receiving a tuple as the index argument (to be uniform in behavior with *args). In fact, given a trivial class:

class X:
    def __getitem__(self, index):
        print(index)

The index operation basically forwards the content of the square brackets “as is” in the index argument:

>>> x=X()
>>> x[0]
0
>>> x[0, 1]
(0, 1)
>>> x[(0, 1)]
(0, 1)
>>>
>>> x[()]
()
>>> x[{1, 2, 3}]
{1, 2, 3}
>>> x["hello"]
hello
>>> x["hello", "hi"]
('hello', 'hi')

The fourth difference is that the indexing operation knows how to convert colon notations to slices, thanks to support from the parser. This is valid:

a[1:3]

this one isn’t:

f(1:3)

The fifth difference is that there’s no zero-argument form. This is valid:

f()

this one isn’t:

a[]

Indexing behavior in standard classes (dict, list, etc.)

None of what is proposed in this PEP will change the behavior of the current core classes that use indexing. Adding keywords to the index operation for custom classes is not the same as modifying e.g. the standard dict type to handle keyword arguments. In fact, dict (as well as list and other stdlib classes with indexing semantics) will remain the same and will continue not to accept keyword arguments. In other words, if d is a dict, the statement d[1, a=2] will raise TypeError, as their implementation will not support the use of keyword arguments. The same holds for all other classes (list, dict, etc.)

Corner case and Gotchas

With the introduction of the new notation, a few corner cases need to be analysed.

1. Technically, if a class defines their getter like this:

   def __getitem__(self, index):
   

   then the caller could call that using keyword syntax, like these two cases:

   obj[3, index=4]
   obj[index=1]
   

   The resulting behavior would be an error automatically, since it would be like attempting to call the method with two values for the index argument, and a TypeError will be raised. In the first case, the index would be 3, in the second case, it would be the empty tuple ().

   Note that this behavior applies for all currently existing classes that rely on indexing, meaning that there is no way for the new behavior to introduce backward compatibility issues on this respect.

   Classes that wish to stress this behavior explicitly can define their parameters as positional-only:

   def __getitem__(self, index, /):
   

2. a similar case occurs with setter notation:

   # Given type(obj).__setitem__(obj, index, value):
   obj[1, value=3] = 5
   

   This poses no issue because the value is passed automatically, and the Python interpreter will raise TypeError: got multiple values for keyword argument 'value'

3. If the subscript dunders are declared to use positional-or-keyword parameters, there may be some surprising cases when arguments are passed to the method. Given the signature:

   def __getitem__(self, index, direction='north')
   

   if the caller uses this:

   obj[0, 'south']
   

   they will probably be surprised by the method call:

   # expected type(obj).__getitem__(obj, 0, direction='south')
   # but actually get:
   type(obj).__getitem__(obj, (0, 'south'), direction='north')
   

   Solution: best practice suggests that keyword subscripts should be flagged as keyword-only when possible:

   def __getitem__(self, index, *, direction='north')
   

   The interpreter need not enforce this rule, as there could be scenarios where this is the desired behaviour. But linters may choose to warn about subscript methods which don’t use the keyword-only flag.

4. As we saw, a single value followed by a keyword argument will not be changed into a tuple, i.e.: d[1, a=3] is treated as __getitem__(d, 1, a=3), NOT __getitem__(d, (1,), a=3). It would be extremely confusing if adding keyword arguments were to change the type of the passed index. In other words, adding a keyword to a single-valued subscript will not change it into a tuple. For those cases where an actual tuple needs to be passed, a proper syntax will have to be used:

   obj[(1,), a=3]
   # calls type(obj).__getitem__(obj, (1,), a=3)
   

   In this case, the call is passing a single element (which is passed as is, as from rule above), only that the single element happens to be a tuple.

   Note that this behavior just reveals the truth that the obj[1,] notation is shorthand for obj[(1,)] (and also obj[1] is shorthand for obj[(1)], with the expected behavior). When keywords are present, the rule that you can omit this outermost pair of parentheses is no longer true:

   obj[1]
   # calls type(obj).__getitem__(obj, 1)
   
   obj[1, a=3]
   # calls type(obj).__getitem__(obj, 1, a=3)
   
   obj[1,]
   # calls type(obj).__getitem__(obj, (1,))
   
   obj[(1,), a=3]
   # calls type(obj).__getitem__(obj, (1,), a=3)
   

   This is particularly relevant in the case where two entries are passed:

   obj[1, 2]
   # calls type(obj).__getitem__(obj, (1, 2))
   
   obj[(1, 2)]
   # same as above
   
   obj[1, 2, a=3]
   # calls type(obj).__getitem__(obj, (1, 2), a=3)
   
   obj[(1, 2), a=3]
   # calls type(obj).__getitem__(obj, (1, 2), a=3)
   

   And particularly when the tuple is extracted as a variable:

   t = (1, 2)
   obj[t]
   # calls type(obj).__getitem__(obj, (1, 2))
   
   obj[t, a=3]
   # calls type(obj).__getitem__(obj, (1, 2), a=3)
   

   Why? because in the case obj[1, 2, a=3] we are passing two elements (which are then packed as a tuple and passed as the index). In the case obj[(1, 2), a=3] we are passing a single element (which is passed as is) which happens to be a tuple. The final result is that they are the same.

Previous PEP 472 solutions

PEP 472 presents a good amount of ideas that are now all to be considered Rejected. A personal email from D’Aprano to the author specifically said:

I have now carefully read through PEP 472 in full, and I am afraid I cannot support any of the strategies currently in the PEP.

We agree that those options are inferior to the currently presented, for one reason or another.

To keep this document compact, we will not present here the objections for all options presented in PEP 472. Suffice to say that they were discussed, and each proposed alternative had one or few dealbreakers.

Adding new dunders

It was proposed to introduce new dunders __(get|set|del)item_ex__ that are invoked over the __(get|set|del)item__ triad, if they are present.

The rationale around this choice is to make the intuition around how to add kwd arg support to square brackets more obvious and in line with the function behavior. Given:

def __getitem_ex__(self, x, y): ...

These all just work and produce the same result effortlessly:

obj[1, 2]
obj[1, y=2]
obj[y=2, x=1]

In other words, this solution would unify the behavior of __getitem__ to the traditional function signature, but since we can’t change __getitem__ and break backward compatibility, we would have an extended version that is used preferentially.

The problems with this approach were found to be:

• It will slow down subscripting. For every subscript access, this new dunder attribute gets investigated on the class, and if it is not present then the default key translation function is executed. Different ideas were proposed to handle this, from wrapping the method only at class instantiation time, to add a bit flag to signal the availability of these methods. Regardess of the solution, the new dunder would be effective only if added at class creation time, not if it’s added later. This would be unusual and would disallow (and behave unexpectedly) monkeypatching of the methods for whatever reason it might be needed.
• It adds complexity to the mechanism.
• Will require a long and painful transition period during which time libraries will have to somehow support both calling conventions, because most likely, the extended methods will delegate to the traditional ones when the right conditions are matched in the arguments, or some classes will support the traditional dunder and others the extended dunder. While this will not affect calling code, it will affect development.
• it would potentially lead to mixed situations where the extended version is defined for the getter, but not for the setter.
• In the __setitem_ex__ signature, value would have to be made the first element, because the index is of arbitrary length depending on the specified indexes. This would look awkward because the visual notation does not match the signature:

  obj[1, 2] = 3
  # calls type(obj).__setitem_ex__(obj, 3, 1, 2)
  

• the solution relies on the assumption that all keyword indices necessarily map into positional indices, or that they must have a name. This assumption may be false: xarray, which is the primary Python package for numpy arrays with labelled dimensions, supports indexing by additional dimensions (so called “non-dimension coordinates”) that don’t correspond directly to the dimensions of the underlying numpy array, and those have no position to match up to. In other words, anonymous indexes are a plausible use case that this solution would remove, although it could be argued that using *args would solve that issue.

Adding an adapter function

Similar to the above, in the sense that a pre-function would be called to convert the “new style” indexing into “old style indexing” that is then passed. Has problems similar to the above.

create a new “kwslice” object

This proposal has already been explored in “New arguments contents” P4 in PEP 472:

obj[a, b:c, x=1]
# calls type(obj).__getitem__(obj, a, slice(b, c), key(x=1))

This solution requires everyone who needs keyword arguments to parse the tuple and/or key object by hand to extract them. This is painful and opens up to the get/set/del function to always accept arbitrary keyword arguments, whether they make sense or not. We want the developer to be able to specify which arguments make sense and which ones do not.

Using a single bit to change the behavior

A special class dunder flag:

__keyfn__ = True

would change the signature of the __get|set|delitem__ to a “function like” dispatch, meaning that this:

>>> d[1, 2, z=3]

would result in a call to:

>>> type(obj).__getitem__(obj, 1, 2, z=3)
# instead of type(obj).__getitem__(obj, (1, 2), z=3)

This option has been rejected because it feels odd that a signature of a method depends on a specific value of another dunder. It would be confusing for both static type checkers and for humans: a static type checker would have to hard-code a special case for this, because there really is nothing else in Python where the signature of a dunder depends on the value of another dunder. A human that has to implement a __getitem__ dunder would have to look if in the class (or in any of its subclasses) for a __keyfn__ before the dunder can be written. Moreover, adding a base classes that have the __keyfn__ flag set would break the signature of the current methods. This would be even more problematic if the flag is changed at runtime, or if the flag is generated by calling a function that returns randomly True or something else.

Allowing for empty index notation obj[]

The current proposal prevents obj[] from being valid notation. However a commenter stated

We have Tuple[int, int] as a tuple of two integers. And we have Tuple[int] as a tuple of one integer. And occasionally we need to spell a tuple of no values, since that’s the type of (). But we currently are forced to write that as Tuple[()]. If we allowed Tuple[] that odd edge case would be removed.

So I probably would be okay with allowing obj[] syntactically, as long as the dict type could be made to reject it.

This proposal already established that, in case no positional index is given, the passed value must be the empty tuple. Allowing for the empty index notation would make the dictionary type accept it automatically, to insert or refer to the value with the empty tuple as key. Moreover, a typing notation such as Tuple[] can easily be written as Tuple without the indexing notation.

However, subsequent discussion with Brandt Bucher during implementation has revealed that the case obj[] would fit a natural evolution for variadic generics, giving more strength to the above comment. In the end, after a discussion between D’Aprano, Bucher and the author, we decided to leave the obj[] notation as a syntax error for now, and possibly extend the notation with an additional PEP to hold the equivalence obj[] as obj[()].

Sentinel value for no given positional index

The topic of which value to pass as the index in the case of:

obj[k=3]

has been considerably debated.

One apparently rational choice would be to pass no value at all, by making use of the keyword only argument feature, but unfortunately will not work well with the __setitem__ dunder, as a positional element for the value is always passed, and we can’t “skip over” the index one unless we introduce a very weird behavior where the first argument refers to the index when specified, and to the value when index is not specified. This is extremely deceiving and error prone.

The above consideration makes it impossible to have a keyword only dunder, and opens up the question of what entity to pass for the index position when no index is passed:

obj[k=3] = 5
# would call type(obj).__setitem__(obj, ???, 5, k=3)

A proposed hack would be to let the user specify which entity to use when an index is not specified, by specifying a default for the index, but this forces necessarily to also specify a (never going to be used, as a value is always passed by design) default for the value, as we can’t have non-default arguments after defaulted one:

def __setitem__(self, index=SENTINEL, value=NEVERUSED, *, k)

which seems ugly, redundant and confusing. We must therefore accept that some form of sentinel index must be passed by the Python implementation when the obj[k=3] notation is used. This also means that default arguments to those parameters are simply never going to be used (but it’s already the case with the current implementation, so no change there).

Additionally, some classes may want to use **kwargs, instead of a keyword-only argument, meaning that having a definition like:

def __setitem__(self, index, value, **kwargs):

and a user that wants to pass a keyword value:

x[value=1] = 0

expecting a call like:

type(obj).__setitem__(obj, SENTINEL, 0, **{"value": 1})

will instead accidentally be caught by the named value, producing a duplicate value error. The user should not be worried about the actual local names of those two arguments if they are, for all practical purposes, positional only. Unfortunately, using positional-only values will ensure this does not happen but it will still not solve the need to pass both index and value even when the index is not provided. The point is that the user should not be prevented to use keyword arguments to refer to a column index, value (or self) just because the class implementor happens to use those names in the parameter list.

Moreover, we also require the three dunders to behave in the same way: it would be extremely inconvenient if only __setitem__ were to receive this sentinel, and __get|delitem__ would not because they can get away with a signature that allows for no index specification, thus allowing for a user-specified default index.

Whatever the choice of the sentinel, it will make the following cases degenerate and thus impossible to differentiate in the dunder:

obj[k=3]
obj[SENTINEL, k=3]

The question now shifts to which entity should represent the sentinel: the options were:

1. Empty tuple
2. None
3. NotImplemented
4. a new sentinel object (e.g. NoIndex)

For option 1, the call will become:

type(obj).__getitem__(obj, (), k=3)

therefore making obj[k=3] and obj[(), k=3] degenerate and indistinguishable.

This option sounds appealing because:

1. The numpy community was inquired 8, and the general consensus of the responses was that the empty tuple felt appropriate.
2. It shows a parallel with the behavior of *args in a function, when no positional arguments are given:

   >>> def foo(*args, **kwargs):
   ...     print(args, kwargs)
   ...
   >>> foo(k=3)
   () {'k': 3}
   

   Although we do accept the following asymmetry in behavior compared to functions when a single value is passed, but that ship has sailed:

   >>> foo(5, k=3)
   (5,) {'k': 3}   # for indexing, a plain 5, not a 1-tuple is passed
   

For option 2, using None, it was objected that NumPy uses it to indicate inserting a new axis/dimensions (there’s a np.newaxis alias as well):

arr = np.array(5)
arr.ndim == 0
arr[None].ndim == arr[None,].ndim == 1

While this is not an insurmountable issue, it certainly will ripple onto numpy.

The only issues with both the above is that both the empty tuple and None are potential legitimate indexes, and there might be value in being able to differentiate the two degenerate cases.

So, an alternative strategy (option 3) would be to use an existing entity that is unlikely to be used as a valid index. One option could be the current built-in constant NotImplemented, which is currently returned by operators methods to report that they do not implement a particular operation, and a different strategy should be attempted (e.g. to ask the other object). Unfortunately, its name and traditional use calls back to a feature that is not available, rather than the fact that something was not passed by the user.

This leaves us with option 4: a new built-in constant. This constant must be unhashable (so it’s never going to be a valid key) and have a clear name that makes it obvious its context: NoIndex. This would solve all the above issues, but the question is: is it worth it?

From a quick inquire, it seems that most people on python-ideas seem to believe it’s not crucial, and the empty tuple is an acceptable option. Hence the resulting series will be:

obj[k=3]
# type(obj).__getitem__(obj, (), k=3). Empty tuple

obj[1, k=3]
# type(obj).__getitem__(obj, 1, k=3). Integer

obj[1, 2, k=3]
# type(obj).__getitem__(obj, (1, 2), k=3). Tuple

and the following two notation will be degenerate:

obj[(), k=3]
# type(obj).__getitem__(obj, (), k=3)

obj[k=3]
# type(obj).__getitem__(obj, (), k=3)