I think that I've almost used a right join before, but not for anything like a legitimate use case. I was attempting to out-think an in-house DB's optimizer that was doing a hash join in the wrong direction. There may have been some rough spots in some of that DB's specialized SQL extensions that played into the problem. I think I ended up using a full outer join plastered with optimizer hints. Like you said, if you're using a right join, probably not unlikely that you're doing something wrong.

I use right_join never in SQL, but often in offline data analysis (e.g. in R). Modern R has an operator called the "pipe" (%>%) which allows left to right evaluation of functions, and so it's fairly common to write code that basically reads like, say, chained method invocations in JavaScript (i.e. object.method1().method2().method3()). The operator works so that the invoking object is automatically passed as the first argument of the function, so func(x, y) is the same as x %>% func(y). You might see where this is going.

Think of left join as a join from x to y, and a right join as a join from y to x, where x is the data we keep all of and y is the data we keep only when there's a match.

Then, in R, I often use right joins when my "y" data requires preprocessing, resulting in lines of code that are like: (y %>% preprocess1() %>% preprocess2() %>% right_join(x)). I could of course write this as "y_preprocessed = y %>% preprocess1() %>% ...; x %>% left_join(y_preprocessed)" but I think the former is actually a little syntactically clearer.

    x %>% left_join(y %>% preprocess1() %>% preprocess2())

A right join is a left join flipped so you never formally need one, like you said.

I believe I used one once, exactly once ever, about a year ago. That was because I was fighting an almighty 5 page CTE charlie fox of a query and had to turn an inner join into an outer join to get data otherwise excluded. It just so happened the data missing was on the right so I made it a right join. Had it not been so awful I'd have flipped it into a left join, my usual style, but I couldn't bear rewriting it.

(background: it was complex but could not be broken down or it lost efficiency. Part of the length was long explicit lists of fields 'select a, b, c, d, ...' which were unavoidable, but it was just plain complex too. The inner join suddenly became an outer join because we found data missing - the DB was a commercial one written by some clueless people with little consistency and absolutely no documentation. I hope I never see that again).

I struggle to think of time I used a right join in the 20-odd years I've used SQL.

The example shows a case where you'd want foreign keys to enforce referential integrity, though FKs aren't fashionable these days. You shouldn't be able to have an entry for owner_id 8 in the Dogs table without a corresponding Owner.

A right join is simply a commuted left join, so barring some bizzare DMBS optimizer oddity, there's no nothing you couldn't trivially replace with left join or vice versa.

I just checked some code, saw 357 left joins and 1 right join.

    SELECT p.password, u.id
    FROM password p
      RIGHT JOIN "user" u ON p.user_id = u.id

Agree on struggling for real scenarios where you'd use a right-join, but your example with

select * from dogs where owner_id not in (select id from owners)

You wouldn't actually recommend something like this would you? The query plans for sub-selects vs. left-joins are very different, with lots of implementations having limits on the "in" clause.

Kind of funny that the way you'd feel if you saw a right join is similar to how I feel about sub-selects!

I'm by no means a SQL expert. But I just used a right join a few days ago (for the first time ever).

I thought it made sense to start the SELECT from one table. From there, I could left join to another table, and from that table I could only right join to get what I needed.

I'd love to know, is there another way around that? I know I could have started from a different table, but I think the query makes more sense starting from where I did.

Thanks for pointing that out. It has been updated.

Generally any equality is fine, but you'd usually do something like on mycustomfunction(a) = mycustomfuntion(b)

And if you do this, it will be incredibly slow because you cant use any indexes in this case.

Would that come with the benefit of not getting nasty surprises when the cantankerous scheduler suddenly decides to mess with a previously efficient solution? Or even protect me from myself in the future accidentally breaking an efficient query through an ostensibly innocent change?

I would appreciate a way to encode execution efficiency parameters in the semantics of the query itself. Specifying which indexes to use, or the runtime complexity of certain operations. This would also help me figure out which indexes I need to create, in the first place. Today I can add a clause in an order which makes using an index impossible, rendering a previously efficient query suddenly extremely expensive. The failure mode is very obscure: it all works, the DB just works overtime. I might not have a large enough DB yet to actually notice this, until one day I see these slow queries and have to backtrack all the way to that one commit that caused it, six months ago.

I have encountered this scenario a few odd times. It is not always obvious to someone changing a query that it is meant to use an index, and it is very hard for someone writing it to specify so in a unit test.

Flexibility for the scheduler to choose efficient strategies cuts both ways. I'd appreciate the ability to leave more explicit annotations about expected runtime characteristics, than mere comments offer.