On Fri, Oct 4, 2019 at 3:59 PM Caleb Donovick <donovick@cs.stanford.edu> wrote:
While there is no restriction on passing dicts to getitem.  Doing so tends to be a bit ugly.  I have two main use cases in mind for this syntax.

The first  and perhaps the most obvious, is doing relational queries.
where_x_1 = db[x=1]

is more beautiful than
where_x_1 = db[dict(x=1)]  
where_x_1 = db[{'x': 1}]
# or by abusing slices
where_x_1 = db['x':1]
# or in the style of Pandas
where_x_1 = db[db['x'] == 1]

Beyond relational queries my own personal use case is a shorthand for dataclasses / protocols.
foo: ProtoRecord[x=int, y=int] = DataRecord[x=int, y=int](0, 1)
where `DataRecord[field0=T0, ..., fieldk=Tk]` generates 
class Record:

      field0: T0
      fieldk: Tk 
and `ProtoRecord[field0=T0, ..., fieldk=Tk]` generates a similar protocol.

Allowing key value pairs in geitem need not change the interface of getitem.   All the key value pairs could be collected as a dict and passed to getitem as the index. Similar to how the all the positional arguments are gather into a single tuple.
class Foo:
  def __getitem__(self, idx):

f = Foo()
f[x=1, y=2] # {'x': 1, 'y': 2}
This would make any legacy code using normal dicts as keys (I don't know how prevalent that is) automatically work with  the new syntax. 

There doesn't necessarily need to be support for mixing of tuple based indexing and keyword indexing. i.e.
obj[0, x=1] # SyntaxError

I don't really know anything about parsers but I think the grammar could be extended without issue with the following rule:
subscriptlist: ... | kwargsubscript (','  kwargsubscript )* [',']
kwargsubscript: NAME '=' test  
if `NAME '=' test` would result in ambiguity similar to argument it could be `test '=' test` with a block in ast.c

   -  Caleb Donovick

I thought about this in terms of labelled arrays like pandas and xarray, but I haven't talked to them about whether they would actually want to use such a feature so I hesitated to post it here.

One possible approach I thought of for backwards-compatibility would be to be to create a new "kwslice" object, that would be API-incompatible and not derived from "slice".  Anything that isn't explicitly programmed to deal with it would simply choke on the unknown object.  The downside is the error message would be somewhat cryptic.