Re: [PATCH] Change wait_time column of pg_stat_lock to double precision

On Mon, Jun 29, 2026 at 04:27:55AM +0000, Bertrand Drouvot wrote:
> Hi,
> 
> On Mon, Jun 29, 2026 at 11:11:06AM +0900, Michael Paquier wrote:
>> Relying on long here is not reliable on WIN32, where the value could
>> overflow for long wait times (where sizeof(long) == 4).  I'd suggest a
>> int64 or a uint64 instead for the API to be able to store wait times
>> in usecs longer than the overflow threshold.  That was another design
>> oversight.  My oversight here (aka I want to abolish completely the
>> use of long in the core code; it leads to insanity).
> 
> That makes sense and the lock time that would produce the overflow is realistic
> (about 35 minutes) if my math is correct.

Dealing with the "long" change, I have arrived at the conclusion that
using int64 or even uint64 felt stupid because we already use a cast
to PgStat_Counter in pgstat_count_lock_waits(), and some other stats
just use PgStat_Counter when reporting usec times.  So I have just
switched the code to use PgStat_Counter.

>> The counter is stored in usecs, displayed in msecs.  Documenting it as
>> stored in msecs is incorrect, no?
> 
> Yeah, so this change.

I need to re-learn how to read.

And backpatched down to v19, catversion bump oblige.
--
Michael