This is the summary exercise for the topic: the exact same data, four join types, four different answers.

The data: on the left, ids 1 and 2; on the right, 2 and 3. Only id 2 is shared.

inner - one row. Only the match. id=1 and id=3 both disappeared.
left - two rows. All of the left kept, id=1 got None in v.
right - two rows. All of the right kept, id=3 got None in n.
outer - three rows. No row is lost: 1 missing on the right, 2 fully matched, 3 missing on the left.

outer is the right tool when you're comparing two sources and want to know what's in each of them and what's missing from the other - for example comparing a customer list in the billing system to the list in the CRM. The rows with the gaps are the answer.

And finally, notice the error: how="banana" raises a ValueError that lists the four valid values. That's intentional - a loud error is better than a silently wrong result. The exact same behavior in real pandas.

Note: a.merge(b, how="right") is equivalent to b.merge(a, how="left"), up to column order. In practice most analysts only ever write left and outer, and arrange their tables accordingly.

Exercises (click a number):