SampleDatasetUtils is in disarray

[rogancarr]

There is a project called SamplesUtils sitting inside the core ML.NET codebase. I believe this was created before the [BestFriend] concept came around.

This project contains one namespace, SampleDatasetUtils that has one static class, DatasetUtils. Here are the following issues:

1. Everything is public and nothing is sealed.
2. This was built for Samples, but now Tests have taken a dependence on a couple of methods.
3. The file itself has no organization, everything is everywhere.

[rogancarr]

I suggest the following fixes:

1. Mark everything as internal and all non-static classes as sealed.
2. Move this library class into Samples and out of the core codebase.
3. Mark everything appropriate as BestFriend and let Tests have access to them. I do think these will also be useful to Tests, and we do explicitly test some of our Sample code.

@sfilipi @TomFinley , what do you think about this?

[sfilipi]

This class was created and is shipping inside the main Microsoft.ML package, so that the users can copy/paste the API references samples and they will just run.

#2277 has some similar questions.

We should definitely seal the classes, but leave public visibility to the methods that download the datasets, and other utilities, because they need to be referenced from outside our own assemblies.

We don't want to use the [BestFriend] annotation from the samples project.

We could organize the SamplesUtils file, as it grows.

[rogancarr]

Thanks Senja!

I'll start with the move to sealing, but I do share @TomFinley 's concerns from #2277. If we ship this in the main NuGet with public APIs, then people, including us as evidenced by our Tests usage, will begin taking a dependence on these methods and classes, and then we'll have to support them forever.

[sfilipi]

The alternative would be to have a separate package for the datasets. We have discussed that here: #2393 1137 and the decision was to not do it. Reconsider?

Scikit has a package for dataset, so does R.

[TomFinley]

As already expressed in #2277 and, as @rogancarr has already noted, I feel fairly strongly that this class and the assembly containing it should be part of a separate nuget. It is clearly not core functionality of ML.NET, and it's perfectly fine for our samples to rely on multiple nugets (as, of course, they already do). Yet it is in Microsoft.ML nuget. So weird.

[Did you mean to link #1137, instead of #2393? (The latter appears to be irrelevant.) In #1137, it was established (correctly) that we should not bundle the datasets in the nuget but download them in some other way. It said nothing about the merits of making that samples class part of the main nuget.

[sfilipi]

Seems like we're in agreement, to summarize this issue can be closed after making the SamplesUtils a separate package.

[eerhardt]

We are not shipping SamplesUtils as a stable package. The SamplesUtils assembly is now in a separate package - Microsoft.ML.SampleUtils - as of #2543.

Moving this out of Project 13, since it is not a stable package.