Marti Raudsepp <marti(at)juffo(dot)org> writes:
> After a bit of digging, I figured out that it uses the same estimate
> as a semi-join WITHOUT the client_id restriction.
> ...
> For whatever reason, the 1st query completely ignores the fact that
> the client_id clause reduces the result count by a large factor.
Could we see the pg_stats rows for the two join columns?
regards, tom lane