perf: Add __slots__ to all DTypes

[dangotbanned]

Closes #3185

What type of PR is this? (check all applicable)

• 💾 Refactor
• ✨ Feature
• 🐛 Bug Fix
• 🔧 Optimization
• 📝 Documentation
• ✅ Test
• 🐳 Other

If you have comments or can explain your changes, please do so below

• Implementent on nw.dtypes
• Check if stable.* needs empty __slots__ (It did)

[dangotbanned]

I'm thinking Schema may be a good candidate as well (provided it is allowed on dict subclasses)

[FBruzzesi]

Thanks @dangotbanned - I am quite a big fan of slots so I am definitely onboarded 😂

I left a comment to (hopefully) lower some of the maintenance effort

[FBruzzesi]

Thanks @dangotbanned - I left a non-bloking nitpick comment for the error message but feel free to ignore it if you don't like it much better.

[jonded94]

Sorry for asking here (I'm also not a maintainer of this library, but just sumbled over this issue):

Instead of manually having to explicitly set __slots__ everywhere (which introduces maintainance load and introduces the risk of setting them improperly), why not use typing.NamedTuple (if immutability is not a problem) or dataclass with slots=True (if mutability is required)?
There are already tools in the stdlib that do these kinds of things automatically for you.

Sorry again, just a curious question from an outsider!

[FBruzzesi]

Sorry again, just a curious question from an outsider!

Hey @jonded94 - no need to apologize, actually thank you for questioning these kind of choices. We should definitely not dive into a solution without thinking about pros and cons of different approaches.

which introduces maintainance load and introduces the risk of setting them improperly

I feel you, I tried to propose a metaclass that can lower such burden :) and #3201 is using something towards such direction, so let's see if it makes it in the main branch soon as well.

why not use typing.NamedTuple (if immutability is not a problem) or dataclass with slots=True (if mutability is required)? There are already tools in the stdlib that do these kinds of things automatically for you.

We probably give it for granted in many places, but the main idea is that we want to have an API which is very close to the polars one. That's not always possible, but I would rather not have a discrepancy such as:

from dataclasses import is_dataclass
from polars import DataType
from narwhals.dtypes import DType

is_dataclass(DataType)
# False

is_dataclass(DType)
# True

or similarly:

from collections import namedtuple
import polars as pl

Int64 = namedtuple('Int64', field_names=())

isinstance(pl.Int64(), tuple), isinstance(Int64(), tuple)
# False, True

I hope it makes sense 😇