> On 21 Mar 2024, at 20:21, Jelte Fennema-Nio <
postgres@jeltef.nl> wrote:
>
> On Wed, 20 Mar 2024 at 19:08, Andrey M. Borodin <
x4mmm@yandex-team.ru> wrote:
>> Timer-based bits contribute to global sortability. But the real timers we have are not even millisecond adjusted. We can hope for ~few ms variation in one datacenter or in presence of atomic clocks.
>
> I think the main benefit of using microseconds would not be
> sortability between servers, but sortability between backends.
Oh, that’s an interesting practical feature!
Se, essentially counter is a theoretical guaranty of sortability in one backend, while microseconds are practical sortability between backends.
> However, I don't really think it is incredibly important to get the
> "perfect" approach to filling in rand_a/rand_b right now. As long as
> we don't document what we do, we can choose to change the method
> without breaking backwards compatibility. Because either approach
> results in valid UUIDv7s.
Makes sense to me. I think both methods would be much better than UUIDv4 for practical reasons. And even not using extra bits at all (fill them with random numbers) would work for 0.999 cases.
Best regards, Andrey Borodin.