On 04/11/2019 06:41 PM, Alvaro Herrera wrote:
> On 2019-Apr-11, rihad wrote:
>
>> On 04/11/2019 06:20 PM, Tom Lane wrote:
>>> rihad <rihad(at)mail(dot)ru> writes:
>>>> Thanks! Our autovacuum_work_mem = 1GB, so this probably means any space
>>>> would be available for reuse only at the end of the vacuum?
>>> It's six bytes per dead tuple, last I checked ... you do the math.
>>>
>>>> Are there
>>>> any downsides in decreasing it to, say, 64MB? I see only pluses )
>>> Well, usually people prefer to minimize the number of passes over
>>> the indexes.
>> Yup, it's just that n_dead_tuples grows by several hundred thousand (the
>> table sees much much more updates than inserts) and disk usage grows
>> constantly between several hour long vacuum runs. Running vacuum full isn't
>> an option.
> Perhaps it'd be better to vacuum this table much more often.
>
Each run took 5-6 hours, now it takes 2-3 hours after I've tweaked some
cost-based vacuum knobs.