Luca Ferrari <fluca1978@gmail.com> writes:
> Therefore my question is about how the statistic collectore decides
> about the number of tuples to be sampled.
It's basically 300 times the largest statistics target:
https://git.postgresql.org/gitweb/?p=postgresql.git;a=blob;f=src/backend/commands/analyze.c;h=0c9591415e4b97dd5c5e693af1860294284a1575;hb=HEAD#l1919
Per that comment, there is good math backing this choice for the task
of making a histogram. It's a little shakier for other sorts of
statistics --- notably, for n_distinct estimation, the error can still
be really bad.
regards, tom lane