Richard Huxton <dev(at)archonet(dot)com> writes:
> 1. PG has changed the way it reports row width (I don't remember any such
> change).
My recollection is that up till 7.2, the estimation of widths for
variable-width columns was completely bogus. Since 7.2 it's driven by
an actual average width for the column as measured by ANALYZE. So if
the query is selecting some fairly wide variable-width columns then it's
entirely likely for the width estimate to take a big jump.
Given that we haven't seen any EXPLAIN ANALYZE output it's hard to say
anything about what the *real* problem is ...
regards, tom lane