tomas@tuxteam.de writes:
> WARNING: I don't really know what I'm talking about -- but considering
> that vaccuming a large table across several "maintainance windows" is
> one of the envisioned scenarios, it might make sense to try to update
> the statistics (i.e. to do partially step 7) on each partial run.
> Otherwise the table's stats might wander off for quite long times?
You can handle that by issuing ANALYZE as a separate operation; I see no
need to complicate this proposal even further by having it make magic
changes to the behavior of VACUUM ANALYZE.
Or were you speaking of the pg_class.reltuples count? Keeping that
from diverging indefinitely far from reality will indeed be a problem
with this sort of thing. We've already seen some issues related to the
stats collector's similar count.
regards, tom lane