Hi,
On 2024-07-03 18:47:15 +0900, Michael Paquier wrote:
> While looking at a different patch from Tristan in this area at [1], I
> still got annoyed that this patch set was not able to support the case
> of custom fixed-numbered stats, so as it is possible to plug in
> pgstats things similar to the archiver, the checkpointer, WAL, etc.
> These are plugged in shared memory, and are handled with copies in the
> stats snapshots. After a good night of sleep, I have come up with a
> good solution for that, among the following lines:
> - PgStat_ShmemControl holds an array of void* indexed by
> PGSTAT_NUM_KINDS, pointing to shared memory areas allocated for each
> fixed-numbered stats. Each entry is allocated a size corresponding to
> PgStat_KindInfo->shared_size.
I am dubious this is a good idea. The more indirection you add, the more
expensive it gets to count stuff, the more likely it is that we end up with
backend-local "caching" in front of the stats system.
IOW, I am against making builtin stats pay the price for pluggable
fixed-numbered stats.
It also substantially reduces type-safety, making it harder to refactor. Note
that you had to add static casts in a good number of additional places.
Greetings,
Andres Freund
