I will reassert my comment from the semilog version of this (#27320):

Yes, log/semilog/etc axes with datetimes do not make much sense, as they are based on the 1970-01-01 zero mark for computing the scale, and so you are rarely working on multiple orders of magnitude where log scaling makes sense (and even when you are, the zero mark is completely arbitrary and unlikely to be related to your data).

Given that the implementation of semilog[xy] is really just some kwarg management over set_[xy]scale plus plot, the latter of which is already tested with datetimes and the former of which is tested pretty thoroughly independently, I think my vote would be to just omit the tests from the datetime file.

If we were to keep it (which, again, I'd argue we likely don't), just to prove that nothing about units prevents it from being applied, this is actually the case where using dates closer to 1970 (and spanning several decades, most likely) is likely beneficial, as it would make the scale transform visually apparent.

mpl.rcParams["date.converter"] = 'concise'
fig, (ax1, ax2, ax3) = plt.subplots(3, 1, layout='constrained')
x = np.arange(np.datetime64('2015-12-01'), np.datetime64('2023-12-06'))
y = 10 ** np.linspace(0, 4, len(x))
ax1.loglog(x, y)
ax2.loglog(y, x)
ax3.loglog(x, x)

As @ksunden points out, this is hardly great for larger time spans, with minor ticks being mislabeled.

OTOH, it doesn't spit out an error message, so maybe thats OK for the purposes of these smoke tests?

I updated the PR to have wider date range input