On Fri, Mar 24, 2023 at 08:00:44PM -0700, Andres Freund wrote:
> I don't understand what we're optimizing for here. These functions are very
> very very far from being a hot path. The xact functions are barely ever
> used. Compared to the cost of query evaluation the cost of iterating throught
> he subxacts is neglegible.
I was wondering about that, and I see why I'm wrong. I have quickly
gone up to 10k subtransactions, and while I was seeing what looks like
difference of 8~10% in runtime when looking at
pg_stat_xact_all_tables, the overval runtime was still close enough
(5.8ms vs 6.4ms). At this scale, possible that it was some noise,
these seemed repeatable still not to worry about.
Anyway, I was looking at this patch, and I still feel that it is a bit
incorrect to have the copy of PgStat_TableStatus returned by
find_tabstat_entry() to point to the same list of subtransaction data
as the pending entry found, while the counters are incremented. This
could lead to mistakes if the copy from find_tabstat_entry() is used
in an unexpected way in the future. The current callers are OK, but
this does not give me a warm feeling :/
--
Michael