define Random.GLOBAL_RNG = TaskLocalRNG()

[rfourquet]

GLOBAL_RNG is a relic from pre-v1.3 when our global RNG was not thread-safe:

const GLOBAL_RNG = MersenneTwister()

GLOBAL_RNG was never really specified, but was referred to in docstrings (of rand etc.), and people had to use it in packages in order to provide a default RNG to their functions. So we have to keep defining Random.GLOBAL_RNG for the time being.

In v1.3, we didn't have anymore a unique global RNG, but instead one per thread; in order to keep code using GLOBAL_RNG working, we got this new definition as a clever hack:

struct _GLOBAL_RNG <: AbstractRNG end
const GLOBAL_RNG = _GLOBAL_RNG()

I.e. GLOBAL_RNG is just a handle, which refers to the actual threaded-RNG when passed to rand functions. This entails defining most function taking a default_rng() to also accept _GLOBAL_RNG. See #41123, and #41235 which is not even merged.

But since v1.7, we have Random.default_rng() (our now official way to refer to the default RNG) returning TaskLocalRNG(), which is itself a "handle" (singleton). So we can as well redefine GLOBAL_RNG to be TaskLocalRNG() instead of _GLOBAL_RNG(), with the advantage that all the relevant methods are already defined for TaskLocalRNG.

The only expected difference in behavior is seed!(GLOBAL_RNG, seed), which previously would update Random.GLOBAL_SEED (a hack used by @testset), but now is equivalent to seed!(TaskLocalRNG(), seed), which doesn't update this global seed. This precise behavior was never documented (GLOBAL_SEED is purely internal), so this should be fine.

[fingolfin]

Really? I think this feature is really useful. I actually was just recently looking into ways to add more things to be reset at the start and end of each @testset (think additional RNGs and other global state). Unfortunately the only way I could see was to create my own AbstractTestSet subtype and then change every single @testset invocation to explicitly specify that type, which is rather a burden :-(

[rfourquet]

Oh that's the first time i see positive feedback on this feature, it has usually been "meh" ! So this comment about removing this was more of a reflexion of what I thought was the general sentiment; and also the fact that since we don't use MersenneTwister as the global RNG, there is a mismatch between what we say:

Before the execution of the body of a @testset, there is an implicit call to Random.seed!(seed) where seed is the current seed of the global RNG.

and what we can do, because the "global RNG" doesn't have a seed (there are multiple global RNGs, one per task, and we currently store only one seed.

But anyway, i'm very glad to see someone else finding this feature useful :) I will remove this comment, and i have a plan to make the feature more robust.