Is there an alternative where the user does this in post-processing? I.e. can we make this Not Our Problem?

The alternative is to use the ax object created by matplotlib and then change the legend location but the problem arises when we call secondary_y. Pandas internally calls matplotlib's twinx and attaches it to ax as:

• ax, ax.right_ax: When there are few columns on secondary_y and not all. In this case you need to get the legend handles for both ax and ax.right_ax, append them and then change their location. Something like the following:

h1, l1 = ax.get_legend_handles_labels()
h2, l2 = ax.right_ax.get_legend_handles_labels()
ax.legend(h1 + h2, l1 + l2, loc='upper center')

• ax, ax.left_ax: When all the columns are on secondary_y. In order to set the legend location correctly in this case, we will have to follow following steps:

While plotting turn off the legend by (legend=False) and then set ax.legend(loc='upper..')
However if we don't turn off the legend then the legend will be overwritten instead of changing the location

leg = ax.left_ax.get_legend()
leg.set_loc('upper center')

So just giving the legend location legend_loc as the parameter initially solves all these problems in a non-hacky way.