Fix divisions when joining two single-partition dataframes

[bluecoconut]

There was an error where if npartition=1 for both left and right on a merge, then the check inside of single_partition_join would short to using the right divisions, even if it was supposed to be left table's divisions that are used.

• Tests added / passed
• Passes flake8 dask

Note: the test I added first validated the bug (on previous code), and then was resolved with the fix in single_partition_join

This answers the bug identified in #4617

[mrocklin]

Are there cases where neither of these two branches is taken?

[bluecoconut]

These are the conditions required to enter single_partition_join method.

dask.dataframe.multi.py#370-371

[bluecoconut]

    # Single partition on one side
    elif (left.npartitions == 1 and how in ('inner', 'right') or
          right.npartitions == 1 and how in ('inner', 'left')):
        return single_partition_join(left, right, how=how, right_on=right_on,
                                     left_on=left_on, left_index=left_index,
                                     right_index=right_index,
                                     suffixes=suffixes, indicator=indicator)

[bluecoconut]

Could add (if desired for future developer ease) an else: raise NotImplementedError... Not sure what best practices in this code base are. At the moment, the method is only called inside of that parent elif statement.

[mrocklin]

Ah, ok, sounds like it's safe then. An else: raise block seems reasonable to me. I'm happy to defer to your preferences here.

[bluecoconut]

Added the raise error~

[mrocklin]

Thanks for taking this on @bluecoconut ! Sorry for the slow response on your issue. A combination of conferences and newborns I think is saturating most maintainer cycles these days :)

[mrocklin]

Thanks @bluecoconut !