Special-cased support for ML.NET

[abock]

While we still need to figure out a better general solution for NuGet packages that depend on native libraries (since there is no standard/conventional way to know what to load, see #204), we should special-case ML.NET support just as we do for a number of other interesting NuGet packages that natively integrate with the target platform.

See dotnet/machinelearning#157 for more bug details. We should continue any discussion here.

We are gearing up for a minor bug fix release against the 1.4 series and a fix for ML.NET support should be trivial enough to include.

Thanks @cosmincatalin for filing the original issue.

[sandyarmstrong]

One thing to note: ML.NET does not have support for 32-bit platforms, and our .NET Core agent is still 32-bit on Windows. I don't see that changing in 1.4.x.

So this special-cased support will only be for:

• Console (.NET Framework) on 64-bit Windows
• WPF on 64-bit Windows
• Console (Mono) on Mac
• .NET Core on Mac
• Xamarin.Mac

[migueldeicaza]

Why can we not do .NET Core agent as 64 on Windows?

[sandyarmstrong]

@migueldeicaza because we still support 32-bit Windows. Our client, console agent, and WPF agent are all AnyCPU. But there is no AnyCPU equivalent for .NET Core publishing.

In order to support 64-bit .NET Core on Windows, we would need to ship a second .NET Core agent app, which would add something like 50MB to the install footprint.

Long-term, we plan to make agent apps download on-demand, which would be the preferred way to fix this.

[abock]

I think we're going to have to eat the size overhead again. I would much rather see ML.NET support on .NET Core 64 bit Windows.

[sandyarmstrong]

@abock fine by me. Do you want it for 1.4.1?

[abock]

Yes.

[migueldeicaza]

64-bit windows is on 93.68% of machines, so we are talking about only 7% of machines without it.

I would rather switch the .NET Core agent to 64, and for 32, point people to the desktop version.