Re: pg_stat_statements max size clarification

Peter Geoghegan <pg(at)bowt(dot)ie> writes:
> On Sat, Jul 15, 2017 at 5:31 PM, <baron(at)xaprb(dot)com> wrote:
>> I wonder if &quot;least-executed&quot; is correct. I&#39;m not an expert and haven&#39;t
>> convinced myself of this by examining the code, but I think after N distinct
>> queryid&#39;s have been seen, then any additional ones are ignored. But that may
>> not be &quot;least-executed&quot; at all. It&#39;s &quot;most-recent&quot; instead. I think we need
>> a new phrase here.

> It's most executed since tracking for the entry began, with a special
> heuristic for queries that take a long time to execute, and might
> therefore consistently be evicted before execution finishes and costs
> are tallied (see "sticky entries" stuff for full details). Most
> executed means the total number of calls, which may not be the best
> thing to evict on the basis of, but certainly isn't too bad.

> The way it actually works is that either 5% of all entries or 10
> entries are evicted (whichever amount is greatest) once
> pg_stat_statements.max entries are reached. You're right that this
> means that the most marginal of entries cannot be usefully tracked,
> but I doubt that that's much of a problem in practice. It's the usual
> "recency versus frequency" cache eviction problem, but for query cost
> tracking purposes if 5,000 entries or 10,000 entries is truly
> insufficient, then pg_stat_statements probably isn't the right tool.

The short answer, really, is that the algorithm is too complicated to be
worth explaining in the documentation --- and it's subject to change,
anyway. But "least-executed" is a reasonable short description, since
frequency of use is a major factor in the decisions. Certainly
"most-recent" is flat out wrong.

I am not sure whether this complaint is actually meant as a bug report
that the algorithm didn't seem to work well on the OP's use case. If so,
we'd need a lot more details to have any hope of improving it (and the
documentation comments aren't the right submission forum, either).

regards, tom lane