=?ISO-8859-1?Q?Andr=E9_Volpato?= <andre.volpato@ecomtecnologia.com.br> writes:
> Besides the (expected) weak guess on rows for both servers on seq scan
> on jtest, there is something nasty with [2] that prevents the planner to
> use the index.
There isn't anything "preventing" either version from choosing any of
the three plans, as you can easily prove for yourself by experimenting
with enable_nestloop/enable_mergejoin/enable_hashjoin. The cost
estimates seem close enough that random variations in ANALYZE stats
would change which one looks cheapest.
regards, tom lane