Inconsistent behavior when saving `TaylorN`s with JLD2 v0.4

[LuEdRaMo]

When saving a long Vector{TaylorN{Float64}} with JLD2 v0.4 (required minor version in Project.toml) I get inconsistent results:

v = set_variables(Float64, "dx"; order = 5, numvars = 6)
x = [exp(sum(rand(6) .* v)) for _ in 1:10_000]
begin
    for _ in 1:10
        jldsave("test.jld2"; x)
        y = JLD2.load("test.jld2", "x")
        println(count(x .!= y))
    end
end
# 3451, 1967, 965, 0, 0, 0, 2051, 0, 1663, 1305

But with JLD2 v0.5 (latest minor version), everything works:

# 0, 0, 0, 0, 0, 0, 0, 0, 0, 0

¿@lbenet can you reproduce this?

[lbenet]

What version of TaylorSeries are you using?

I'm getting the same inconsistencies when I use TaylorSeries v0.18.1 and JLD2 v0.4..53. Yet, if I install JLD2 v0.5 it downgrades TaylorSeries to v0.17.3, and then everything seems consistent.

So, there seems to be an issue with the version and Project.toml in TaylorSeries v0.18.1...

[lbenet]

It seems to be an issue in TaylorSeries v0.18.1: with JLD2 v0.4..53 and TaylorSeries to v0.17.3 everything seems ok. Can you confirm this?

[lbenet]

As we discussed offline, the problem is solved by upgrading to JLD v0.5; #374 should fix it.