My understanding is that the units registry works on the data type of the first element of the iterable, not the iterable. i.e. my_date_time_nano[0]. So if all your iterables are floats, matplotlib just assumes they are native. I don't think there is a way to get the units machinery to check the class of the container with the current registry, but it is something to consider for the future.

@jklymak That is the current behavior for iterables, yes. For objects not caught by the handful of built-in cases, the converter is used on the object. For example:

class Foo:
    def __init__(self, a, b, c):
        self.a = a
        self.b = b
        self.c = c
        return

class Foo:
    def __init__(self, a, b, c):
        self.a = a
        self.b = b
        self.c = c
        return

class FooConverter(units.ConversionInterface):
    @staticmethod
    def convert(value, unit, axis):
        xs = [value.a, value.b, value.c]
        return xs

units.registry[Foo] = FooConverter

During a call to plt.scatter(my_foo, my_other_foo), FooConverter.convert is called resulting in plotting three points (however this does not work for plt.plot as the value argument to the convert function is an ndarray in that case; I'm not familiar with why).

Perhaps this is an abuse of the units system and so it may not make sense as part of the units module, but it seems reasonable to be able to, given an arbitrary object, register how it would like to be interpreted for plotting.

I think you may want to create a custom dtype. https://stackoverflow.com/questions/2350072/custom-data-types-in-numpy-arrays

More to your immediate concern, matplotlib can handle nanoseconds from an epoch already using datetime64. Right now our default epoch is 1970-01-01, and nanoseconds in datetime64 have +/-292y of usable range. So unless you plan to do nanoseconds for more than 584 years, something like datetime = nano_seconds.astype('datetime64[ns]')should work with the default locators and formatters.

I agree that casting to datetime64 can solve that particular example. It seems that the units module is not intended to solve the problem of making arbitrary things plot-able. So a separate issue would be appropriate for discussing that. Thanks!

It looks the behavior of looking at _is_natively_supported was added in 3.2 via #13738 which makes it so that we do not get to

Perhaps moving the transversal of the mro to the front would make sense? But I am still unsure of whether or not the units module is intended to handle coercing arbitrary objects into arrays-likes for plotting or just coercing unrecognized scalars to numbers.