> The whole thing works perfectly after a VACUUM ANALYZE on the
> table.
>
> IMHO this is somewhat non-optimal. In the absence of information
> to the contrary, PostgreSQL should default to using an index if
> it might be appropriate, not ignore it.
There was lots of discussion about this issue, and I was one of those
who are 100% with you. The result was that Tom Lane tried to provide
defaults, that do use the indexes when stats are missing. This does work
quite well, only it seems to fail in this particular case.
Imho we should look at exactly this query and try to find why it ignores
the index, because it should not.
Andreas