Stop compensating for early Timer in libuv

[jakobnissen]

It appears to have been fixed on the libuv side, so now Julia's timers are one millisecond too slow.

Closes #57263

[Keno]

Do we speculate this is fixed or do we know that it is fixed? The reason I ask is because the argument is supposed to be a hard floor, so if it wasn't fixed, this would give incorrect behavior (timing out late is suboptimal but not incorrect according to the spec of the function).

[jakobnissen]

I only speculate, after seeing on my computer that Julia's timer is 1 ms too long. I'll try to look into libuv and see if I can find anything.

[vtjnash]

It was introduced by 7310f55, I think as a workaround because of problems with the original implementation of sleep rounding timeouts down (471828a)?

[martinholters]

It seems that from e5496e0 (#49937) on, the extra 1ms is no longer needed, at least on my system. So indeed, there might have happened a fix in libuv, somewhere between 1.44.2 and 1.48.0. Would still be good to identify the fix over there and verify it is robust and cross-platform.

[vtjnash]

• win,unix: change execution order of timers (Trevor Norris)
  Perhaps? Nothing else mentions timers that I see

[martinholters]

Bisecting libuv brought me to libuv/libuv@00357f8 (libuv/libuv#4026), which BTW is not contained in any libuv releases, yet. Someone else (@vtjnash ?) has to judge whether that's a proper, reliable fix for the timer-firing-early bug.

[jakobnissen]

This appears to have been fixed on Julia 1.12, so closing

[jakobnissen]

Nevermind, it was just a change in the result of BenchmarkTools

[vtjnash]

libuv bug (with reproducer script) filed as libuv/libuv#4773