On Thu, Nov 7, 2024 at 12:34 AM Andrey M. Borodin <x4...@yandex-team.ru> wrote: > > > > > On 7 Nov 2024, at 12:42, Masahiko Sawada <sawada.m...@gmail.com> wrote: > > > > On Wed, Nov 6, 2024 at 10:14 AM Andrey M. Borodin <x4...@yandex-team.ru> > > wrote: > >> > >> > >> > >>> On 5 Nov 2024, at 23:56, Andrey M. Borodin <x4...@yandex-team.ru> wrote: > >>> > >>> <v30-0001-Implement-UUID-v7.patch> > >> > >> Some more thoughts on this patch version: > >> > >> 0. Comment mentioning nanoseconds, while we do not need to carry anything > >> /* Convert TimestampTz back and carry nanoseconds. */ > >> > >> 1. There's unnecessary &3 in > >> uuid->data[7] = uuid->data[7] | ((uuid->data[8] >> 6) & 3); > >> > >> 2. Currently we store 0..999 microseconds in 10 bits, so values 1000..1023 > >> are unused. We could use them for overflow. That would slightly increase > >> non-overflowing capacity when generating more than million UUIDs per > >> second on one backend. However, given current performance of our CSPRNG I > >> do not think this feature worth code complexity. > >> > > > > While using only 10 bits microseconds makes the implementation simple, > > I'm not sure if 10 bits is enough to generate UUIDs at microsecond > > granularity without losing monotonicity. Since 10-bit microseconds are > > used as is in rand_a space, 1000 UUIDs can be generated per > > millisecond without losing monotonicity. > > We won’t loose monotonicity on one backend. We will just accumulate time > shift. > See > + us = tv.tv_sec * SECS_PER_DAY * USECS_PER_SEC + tv.tv_usec; > + if (previous_us >= us) > + us = previous_us + 1;
IIUC the microsecond part is working also as a counter in a sense. It seems fine to me but I'm slightly concerned that there is no guidance of such implementation in RFC 9562. > BTW if we just continue to use nanoseconds patch, zero bits will act exactly > as counters. Yes, but we will lose some randomness on macOS as the nanosecond part is 0 in most cases. Regards, -- Masahiko Sawada Amazon Web Services: https://aws.amazon.com
