On 08/16/2013 10:44 AM, Vlad Arkhipov wrote:
> On 08/15/2013 03:27 AM, Jim Nasby wrote:
>> On 8/14/13 12:31 AM, Vlad Arkhipov wrote:
>>> I used to use VACUUM FULL periodically to resolve the issue, but the
>>> problem arises again in 2-3 months.
>>> Here is the statistics (from pgstattuple). I run VACUUM FULL on
>>> 2013-08-07.
>>
>> How much non-temporary DDL do you do? It's possible that you end up
>> with a tuple at the end of the table for a non-temporary object. One
>> of those would stay valid for quite some time, and if you're unlucky
>> then you'll end up with another long-lived row farther down the
>> table, etc, etc.
>>
>> Depending on how frequently you're creating temp objects, autovac
>> might not be able to keep up. Assuming that a manual vacuum doesn't
>> take too long it might be a good idea to cron a manual vacuum (NOT
>> FULL) of that table once a minute.
> Not much. 1-2 tables per day.