Strange PyPlot Interaction Gives NaN with QR factorization

[jebej]

I have reduced a weird issue I noticed a little while ago. Displaying a pyplot figure, here, just before the call to qr, causes the factorization to return all NaNs.

If the figure isn't displayed, which I'm not sure how to do in the REPL, so the below is run in Atom with display commented out, the factorization works fine.

If the factorization is rerun, the result is fine as well.

using PyPlot
fig = figure(1);
x = rand(10,10);
plot(x); xlim(-12,0);
display(fig); # comment out in Atom so fig doesn't show
first(qr(x))[1]

[jebej]

also ref JuliaNLSolvers/LsqFit.jl#176

[stevengj]

Seems like a duplicate of #477.

Essentially, Matplotlib loads Numpy, which loads a BLAS library, and on some systems Numpy's BLAS is somehow conflicting with Julia's BLAS library. Would be good to comment with what BLAS library you are using with Julia and which you are using with NumPy, not to mention what operating system you have.