Vlad Romascanu <vromascanu@accurev.com> writes:
> The logical, amended solution would then be to have the "writer"
> session perform, after INSERTion but before COMMITTing, the same
> calculation that the autovacuum daemon currently performs inside
> relation_needs_vacanalyze, based on the same configuration parameters
> that the autovacuum daemon uses, before deciding whether to explicitly
> ANALYZE or not the affected application tables. Ideally one would not
> want to duplicate the relation_needs_vacanalyze logic (not to mention
> having to guess the value of last_anl_tuples, which is not exposed via
> any storedproc -- one would have to assume that it is more or less
> equivalent to pg_class.reltuples.)
I don't know if you actually looked at that code, but it's driven off of
statistics counters that are only updated at commit; and furthermore are
maintained in a different process altogether. So what you have in mind
isn't going to work ...
regards, tom lane