On Sat, Nov 23, 2024 at 12:20 AM Andrey M. Borodin <x4...@yandex-team.ru> wrote: > > > > > On 23 Nov 2024, at 10:58, Masahiko Sawada <sawada.m...@gmail.com> wrote: > > > > I've attached an updated patch that squashed changes I made for v33. > > We're still discussing increasing entropy on Windows and macOS, but > > the patch seems to be in good shape. > > +1, thanks! > > PFA version with improved comment.
Thank you for updating the patch! In the following code, we use "defined(__darwin__) || defined(_MSC_VER)": +#if defined(__darwin__) || defined(_MSC_VER) +#define SUBMS_MINIMAL_STEP_BITS 10 +#else +#define SUBMS_MINIMAL_STEP_BITS 12 +#endif #define SUBMS_BITS 12 -#define SUBMS_MINIMAL_STEP_NS ((NS_PER_MS / (1 << SUBMS_BITS)) + 1) +#define SUBMS_MINIMAL_STEP_NS ((NS_PER_MS / (1 << SUBMS_MINIMAL_STEP_BITS)) + 1) on the other hand, we use "defined(__darwin__) || defined(WIN32)" here: +#if defined(__darwin__) || defined(WIN32) + /* + * On MacOS real time is truncted to microseconds. Thus, 2 least + * significant are dependent on other time-specific bits, thus they do not + * contribute to uniqueness. To make these bit random we mix in two bits + * from CSPRNG. + * + * SUBMS_MINIMAL_STEP is chosen so that we still guarantee monotonicity + * despite altering these bits. + */ + uuid->data[7] = uuid->data[7] ^ (uuid->data[8] >> 6); +#endif Is there a reason for using different macros? In get_real_time_ns_ascending(), we use _MSC_VER so we use clock_gettime() on MinGW. > > Sergey Prokhorenko just draw my attention to the new release of MariaDB [0]. > They are doing very similar UUID v7 generation as we do [1]. > Thank you for the references. It made me think that we can use the function name uuid_v7() rather than uuidv7(). Regards, -- Masahiko Sawada Amazon Web Services: https://aws.amazon.com
