Difference in Random numbers between 1.5 & 1.6

General Usage
1

The HISTORY.md file does not show any changes to Random number generation between v1.5 and v1.6, but I still see different numbers generated with the same seed/RNG:

Julia 1.5.4:

julia> VERSION
v"1.5.4"

julia> using Random

julia> observed=rand(MersenneTwister(1234), 0:100, 5)
5-element Array{Int64,1}:
 10
 37
 74
 12
 95

Julia 1.6.3:

julia> VERSION
v"1.6.3"

julia> using Random

julia> observed=rand(MersenneTwister(1234), 0:100, 5)
5-element Vector{Int64}:
 48
 60
 94
 60
 12

Could anyone help explain what’s happening?

Thank you.

2

The TLDR is don’t assume random numbers will be the same between versions. If you want random numbers that are the same between versions use StableRNGs.

2 Likes
3

This is documented and expected behavior.

As a heads up: julia 1.7 will switch from MersenneTwister to Xoshiro (a much faster and easier to parallelize pseudo RNG), as well as have a different RNG object per task, which will also change the stream of random numbers.

4

Thanks. I knew about the change in 1.7, but we’re planning a move to 1.6 since it will be the next LTS. I didn’t expect 1.6 to be different, but I understand why.

5

If you’re planning such a large version jump, also note that due to performance improvements and improvements to numerical accuracy, exact bitpatterns for floating point results are not guaranteed between versions.

If there was such a guarantee, almost any change to floating point algorithms would be breaking.

1 Like
6

I didn’t think 1.5.4 → 1.6.3 was a big jump, it’s just the next available version even though it is a minor version jump.

FWIW, I have over 70000 unit tests and the only one that fails after changing versions is the one that assumed random would be the same.

3 Likes