Robert Treat <xzilla@users.sourceforge.net> writes:
> The thing with vacuum is at some point you should reach a condition
> where space on the disk increases only when the total number of inserts
> is greater than the total number of deletions over a given time frame.
> If it increases when inserts is less than deletes, your not vacuuming
> enough. If you can reach this level of "symbiosis" with regular vacuums,
> you should never have to do vacuum full.
It's simpler than that: the vacuum frequency determines the amount of
overhead you have to tolerate. Judging from the numbers Julian quoted
us to start with, he's got the frequency set so that vacuum reclaims
about 10% of the tuples on each run. That seems like a good ballpark to
me: 10% space overhead is quite reasonable.
If space usage increases when the total amount of live data is staying
about the same, then the problem has got to be that the space getting
reclaimed by vacuum isn't being re-used effectively. And the only cause
for that, AFAIK, is the FSM settings being too small to keep track of
all the free space in the database.
regards, tom lane