> The cost estimation code is under active development, and if you check
> the pgsql list archives you will find lively discussions about its
> deficiencies ;-). But I'm having a hard time mustering much concern
> about whether it behaves optimally in the vacuum-but-no-vacuum-analyze
> case.
Here I must slightly disagree, if the impact of vacuum without analyze is so
bad,
then analyze should be the default for vacuum.
Maybe a better default selectivity would be datatype driven, like
int: 1/(2*maxint)
char(n): 1/(n*256)
....
I think this is what my favorite commercial DBMS does.
Because we have an extensible type system I guess this would be overdoing
it,
but in this light a much smaller default selectivity would seem more
appropriate
(like 1/100000).
Andreas