Am 21.07.22 um 10:41 schrieb Dean Rasheed:
>
> It's important to mark these new functions as VOLATILE, not IMMUTABLE,
> otherwise they won't work as expected in queries. See
> https://www.postgresql.org/docs/current/xfunc-volatility.html
>
> It would be better to use pg_prng_uint64_range() rather than rand() to
> pick elements. Partly, that's because it uses a higher quality PRNG,
> with a larger internal state, and it ensures that the results are
> unbiased across the range. But more importantly, it interoperates with
> setseed(), allowing predictable sequences of "random" numbers to be
> generated -- something that's useful in writing repeatable regression
> tests.
>
> Assuming these new functions are made to interoperate with setseed(),
> which I think they should be, then they also need to be marked as
> PARALLEL RESTRICTED, rather than PARALLEL SAFE. See
> https://www.postgresql.org/docs/current/parallel-safety.html, which
> explains why setseed() and random() are parallel restricted.
>
Here is an updated patch that marks the functions VOLATILE PARALLEL
RESTRICTED and uses pg_prng_uint64_range() rather than rand().