Re: autovacuum can't keep up, bloat just continues to rise

On 21/07/17 15:58, Joshua D. Drake wrote:

> On 07/19/2017 07:57 PM, Tom Lane wrote:
>> Peter Geoghegan <pg(at)bowt(dot)ie> writes:
>>> My argument for the importance of index bloat to the more general
>>> bloat problem is simple: any bloat that accumulates, that cannot be
>>> cleaned up, will probably accumulate until it impacts performance
>>> quite noticeably.
>>
>> But that just begs the question: *does* it accumulate indefinitely, or
>> does it eventually reach a more-or-less steady state? The traditional
>> wisdom about btrees, for instance, is that no matter how full you pack
>> them to start with, the steady state is going to involve something like
>> 1/3rd free space. You can call that bloat if you want, but it's not
>> likely that you'll be able to reduce the number significantly without
>> paying exorbitant costs.
>>
>> I'm not claiming that we don't have any problems, but I do think it's
>> important to draw a distinction between bloat and normal operating
>> overhead.
>
> Agreed but we aren't talking about 30% I don't think. Here is where I
> am at. It took until 30 minutes ago for the tests to finish:
>
> name | setting
> -------------------------------------+-----------
> autovacuum | on
> autovacuum_analyze_scale_factor | 0.1
> autovacuum_analyze_threshold | 50
> autovacuum_freeze_max_age | 200000000
> autovacuum_max_workers | 3
> autovacuum_multixact_freeze_max_age | 400000000
> autovacuum_naptime | 60
> autovacuum_vacuum_cost_delay | 20
> autovacuum_vacuum_cost_limit | -1
> autovacuum_vacuum_scale_factor | 0.2
> autovacuum_vacuum_threshold | 50
> autovacuum_work_mem | -1
> log_autovacuum_min_duration | -1
>
>
> Test 1: 55G /srv/main
> TPS: 955
>
> Test 2: 112G /srv/main
> TPS: 531 (Not sure what happened here, long checkpoint?)
>
> Test 3: 109G /srv/main
> TPS: 868
>
> Test 4: 143G
> TPS: 840
>
> Test 5: 154G
> TPS: 722
>
> I am running the query here:
>
> https://wiki.postgresql.org/wiki/Index_Maintenance#Summarize_keyspace_of_a_B-Tree_index
>
>
> And will post a followup. Once the query finishes I am going to launch
> the tests with autovacuum_vacuum_cost_limit of 5000. Is there anything
> else you folks would like me to change?
>
>
>
>

I usually advise setting autovacuum_naptime = 10s (or even 5s) for
workloads that do a lot of updates (or inserts + deletes) - as on modern
HW a lot of churn can happen in 1 minute, and that just makes vacuum's
job harder.

regards
Mark