Should date::format throw a date::format_error to be more consistent with C++20?

[ecorm]

date::format currently throws std::ios_base::failure upon formatting errors. The C++20 spec mandates that a std::format_error be thrown for chrono formatting errors (reference).

Should date::format instead throw a (proposed) date::format_error, to be more consistent with the standard? This would allow users to catch formatting errors separately from other stream errors:

    try
    {
        std::cout << date::format("%F", std::chrono::seconds(42));
    }
    catch (const std::ios_base::failure& e)
    {
        std::cerr << "Stream failure: " << e.what() << "\n" << std::flush;
    }
    catch (const date::format_error& e)
    {
        std::cerr << "Formatting error: " << e.what() << "\n";
    }

If my proposed date::format_error derives from std::ios_base::failure, then existing applications catching an std::ios_base::failure should still work. Since std::ios_base::failure derives from std::runtime_error, date::format_error would also be consistent with std::format_error deriving from std::runtime_error (i.e. catching std::runtime_error would catch both std::format_error and date::format_error).

The lack of date::format_error is not a showstopper for me, but I thought it may be something "nice to have" in order to be more consistent with C++20.

[ecorm]

My proposed date::format_error could also include more helpful what messages than the ones currently generated by setting the failbit on the internal ostringstream.

[HowardHinnant]

Ideally date should build on https://github.com/fmtlib/fmt . That being said, I don't think a half-way solution would be valuable.

Pros to adopting fmt:

• Much closer compatibility with C++20.
• Superior formatting facilities.

Cons to adopting fmt:

• More work that I haven't made time to do.
• Instability for existing clients.

With any half-way attempt I'm afraid that the cons would hit without really achieving much of the pros.

[ecorm]

Or perhaps it could be the other way around with fmt adding support for time_point via the help of your library (to do the date calculations).