Distinguish between datetime.datetime objects with a timezone and those without

[wojm]

datetime.datetime objects created with timezone information cannot be compared to datetime objects without timezone information. This error comes up

TypeError: can't compare offset-naive and offset-aware datetimes

I believe it is possible for the type system to handle distinguishing between these two objects.

I have a proof of concept at my Company that looks something like this. The idea is to distinguish between these two object types and forbid their corresponding comparisons.

class DatetimeWithTimezone(datetime_lib.datetime):
  """Datetime with timezone."""

  @overload
  def replace(
      self,
      year: SupportsIndex = ...,
      month: SupportsIndex = ...,
      day: SupportsIndex = ...,
      hour: SupportsIndex = ...,
      minute: SupportsIndex = ...,
      second: SupportsIndex = ...,
      microsecond: SupportsIndex = ...,
      tzinfo: _TzInfo = ...,
      *,
      fold: int = ...,
  ) -> Self: ...
  @overload
  def replace(
      self,
      year: SupportsIndex = ...,
      month: SupportsIndex = ...,
      day: SupportsIndex = ...,
      hour: SupportsIndex = ...,
      minute: SupportsIndex = ...,
      second: SupportsIndex = ...,
      microsecond: SupportsIndex = ...,
      tzinfo: None = ...,
      *,
      fold: int = ...,
  ) -> DatetimeWithoutTimezone: ...
  @overload
  def astimezone(self, tz: _TzInfo) -> Self: ...
  @overload
  def astimezone(self, tz: None) -> DatetimeWithoutTimezone: ...
  def utcoffset(self) -> timedelta: ...
  def tzname(self) -> str: ...
  def dst(self) -> timedelta: ...
  def __le__(self, value: DatetimeWithTimezone, /) -> bool: ...  # type: ignore[override]
  def __lt__(self, value: DatetimeWithTimezone, /) -> bool: ...  # type: ignore[override]
  def __ge__(self, value: DatetimeWithTimezone, /) -> bool: ...  # type: ignore[override]
  def __gt__(self, value: DatetimeWithTimezone, /) -> bool: ...  # type: ignore[override]
  @overload  # type: ignore[override]
  def __sub__(self, value: Self, /) -> timedelta: ...
  @overload
  def __sub__(self, value: timedelta, /) -> Self: ...

class DatetimeWithoutTimezone(datetime_lib.datetime):
  """Datetime without timezone."""

  @overload
  def replace(
      self,
      year: SupportsIndex = ...,
      month: SupportsIndex = ...,
      day: SupportsIndex = ...,
      hour: SupportsIndex = ...,
      minute: SupportsIndex = ...,
      second: SupportsIndex = ...,
      microsecond: SupportsIndex = ...,
      tzinfo: _TzInfo = ...,
      *,
      fold: int = ...,
  ) -> DatetimeWithTimezone: ...
  @overload
  def replace(
      self,
      year: SupportsIndex = ...,
      month: SupportsIndex = ...,
      day: SupportsIndex = ...,
      hour: SupportsIndex = ...,
      minute: SupportsIndex = ...,
      second: SupportsIndex = ...,
      microsecond: SupportsIndex = ...,
      tzinfo: None = ...,
      *,
      fold: int = ...,
  ) -> Self: ...
  @overload
  def astimezone(self, tz: _TzInfo) -> DatetimeWithoutTimezone: ...
  @overload
  def astimezone(self, tz: None) -> Self: ...
  def utcoffset(self) -> None: ...
  def tzname(self) -> None: ...
  def dst(self) -> None: ...
  def __le__(self, value: DatetimeWithoutTimezone, /) -> bool: ...  # type: ignore[override]
  def __lt__(self, value: DatetimeWithoutTimezone, /) -> bool: ...  # type: ignore[override]
  def __ge__(self, value: DatetimeWithoutTimezone, /) -> bool: ...  # type: ignore[override]
  def __gt__(self, value: DatetimeWithoutTimezone, /) -> bool: ...  # type: ignore[override]
  @overload  # type: ignore[override]
  def __sub__(self, value: Self, /) -> timedelta: ...
  @overload
  def __sub__(self, value: timedelta, /) -> Self: ...

class datetime(datetime_lib.datetime):
  @overload
  def __new__(
      cls,
      year: SupportsIndex,
      month: SupportsIndex,
      day: SupportsIndex,
      hour: SupportsIndex = ...,
      minute: SupportsIndex = ...,
      second: SupportsIndex = ...,
      microsecond: SupportsIndex = ...,
      tzinfo: _TzInfo = ...,
      *,
      fold: int = ...,
  ) -> DatetimeWithTimezone: ...
  @overload
  def __new__(
      cls,
      year: SupportsIndex,
      month: SupportsIndex,
      day: SupportsIndex,
      hour: SupportsIndex = ...,
      minute: SupportsIndex = ...,
      second: SupportsIndex = ...,
      microsecond: SupportsIndex = ...,
      tzinfo: None = ...,
      *,
      fold: int = ...,
  ) -> DatetimeWithoutTimezone: ...
  # On <3.12, the name of the first parameter in the pure-Python implementation
  # didn't match the name in the C implementation,
  # meaning it is only *safe* to pass it as a keyword argument on 3.12+
  # Assume are on 3.12+
  @overload
  @classmethod
  def fromtimestamp(
      cls, timestamp: float, tz: None = ...
  ) -> DatetimeWithoutTimezone: ...
  @overload
  @classmethod
  def fromtimestamp(
      cls, timestamp: float, tz: _TzInfo
  ) -> DatetimeWithTimezone: ...
  @overload
  @classmethod
  def now(cls, tz: _TzInfo) -> DatetimeWithTimezone: ...
  @overload
  @classmethod
  def now(cls, tz: None = ...) -> DatetimeWithoutTimezone: ...
  @overload
  @classmethod
  def combine(
      cls, date: _Date, time: _Time, tzinfo: None = ...
  ) -> DatetimeWithoutTimezone: ...
  @overload
  @classmethod
  def combine(
      cls, date: _Date, time: _Time, tzinfo: _TzInfo
  ) -> DatetimeWithTimezone: ...
  @classmethod
  def strptime(
      cls, date_string: str, format: str, /
  ) -> DatetimeWithTimezone | DatetimeWithoutTimezone: ...

I have some questions before I'm certain this could be possible for everyone

1. Is it possible to make this backwards compatible?
2. Is this something others are interested in?
3. Should this actually be solved in the datetime library directly?

[srittau]

It would be great if we could somehow distinguish between tz-aware and naive datetimes using type checking and this is a clever idea. And yes, ideally this should be solved in the datetime library (like the nonsensical inheritance of datetime from date), but I can't see that happening realistically.

That said, I don't think this solution would really work for typeshed. Sticking close to the implementation is fairly important to use for various reasons, and introducing "fake" classes like this can easily add unforeseen problems.

Another approach I tried (and which I should look at again) is making datetime generic over tzinfo, see python/typeshed#11844.