> Ideally, autovacuum would only request a new copy of the file if the> one it got was considerably out of date.
Obviouslya tenth of a> second is not old enough.
I've tried to look at it and found that's already implemented - see
autovac_refresh_stats(). STATS_READ_DELAY which is set to 1s. Am I
reading the code correctly? If so then 1s is not enough for big clusters.
I guess it would be feasible to crank STATS_READ_DELAY up a little bit,
say to 10s. What do you think?
Kuba
Dne 16.2.2010 19:59, Alvaro Herrera napsal(a):
> Jakub Ouhrabka wrote:
>>> Maybe you should decrease naptime a bit.
>>
>> That did the trick, thanks!
>>
>>> Yes. There were some changes that needed to be done to autovacuum so
>>> that it didn't read the stats file too often, but I don't recall if I
>>> got around to it.
>>
>> I looked at the strace output and there are *writes* to the file not
>> reads. Why? Is it a consequence of this optimization?
>>
>> Release notes 8.4:
>>
>> Reduce I/O load of writing the statistics collection file by writing
>> the file only when requested (Martin Pihlak)
>>
>> Was autovacuum requesting to write this 20MB file 650x per minute?
>
> Yes, exactly.
>
> Ideally, autovacuum would only request a new copy of the file if the one
> it got was considerably out of date. Obviously a tenth of a second is
> not old enough.
>
