[PERFORMANCE] object indexes repeatedly attempt to import missing cftime (regression) #11559

What is your issue?

Operations on arrays with indexes of object dtype are significantly slower if cftime library (not a minimal dependency) is not installed.

This is due to the fact that _maybe_cast_to_cftimeindex eagerly tries to import cftime every time it is called. Failed imports are not cached, so this is re-attempted many times over the course of an operation, leading to significant slowdown. If cftime is installed (e.g., by including the io extra dependency group), the issue is not encountered.

I dug a little bit: issue was similarly reported in #5929 and then resolved by #5640, but regressed back in #9561.

This one instance can be resolved by patching the _maybe_cast_to_cftimeindex helper to use utils.module_available, which is lru-cached. Something like

def _maybe_cast_to_cftimeindex(index: pd.Index) -> pd.Index:
    from xarray.coding.cftimeindex import CFTimeIndex

    if (
        len(index) > 0 
        and index.dtype == "O"
        and utils.module_available("cftime") # this is new
        and not isinstance(index, CFTimeIndex)
    ):
        try:
            return CFTimeIndex(index)
        except (ImportError, TypeError):
            return index
    else:
        return index

I propose this in #11558.
Similarly happy with any other resolution, if you don't like this one.