Permalink
Switch branches/tags
Nothing to show
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
277 lines (167 sloc) 17.2 KB

F# Tooling RFC FST-1001 - FSharp.Core delivery via NuGet packages

NOTE: This is a draft and remains to be reviewed by key contributors. Please contribute to the discussion and submit adjustments to the draft.

The purpose of this RFC document is to develop a shared plan of record for the delivery of FSharp.Core in NuGet package form that takes into account the needs of F# Core Engineering contributors. This is a long-term technical plan to converge to an optimal solution that satisfies both users of F# and those delivering tooling for it.

This is not as simple as readers may expect. It is probably not interesting except to those contributing to F# Core Engineering. Also, this does not cover the actual functionality contained in FSharp.Core.

Discussion thread: https://github.com/fsharp/fslang-design/issues/197

Background

Historically FSharp.Core.dll was compiled and delivered by Microsoft as a .NET 4.x binary installed under "Reference Assemblies". The DLL was signed and strong-named by Microsoft. Way, way back in time it was called fslib.dll.

Subsequently, FSharp.Core.dll has been recompiled for "trimmed down" platforms: portable profiles, Xamarin profiles, .NET Core, .NET Standard and .NET Framework. There is even a version of FSharp.Core for the "Xamarin iOS TV" profile.... When provided by Microsoft, these were signed, strong-named and installed under "Reference Assemblies" on Windows. For other F# compilation environments like Xamarin, Mono, Cloud Sharper, FSharp.Formatting, Azure Notebooks, Azure Functions there was a bit of a mess (made considerably worse by the separation of sigdata/optdata files, now addressed)

The F# Core Engineering Group created a NuGet package called FSharp.Core, partly to avoid a proliferation of "homebrew" packages appearing at that time. This package has now grown to be a "one-stop-shop" package for FSharp.Core for all different target platforms. A single, unified FSharp.Core NuGet package has advantages:

  • Simplicity: Users don't have to think at all - they just reference the package and that's that

  • Multi-targeting: One package supports building against multiple targets

The F# Core Engineering group also publish notes and guidance on FSharp.Core.dll.

Equally, the FSharp.Core NuGet pacakge has problems, see below.

Today

As of 2017, FSharp.Core.dll is very stable in design and the FSharp.Core NuGet package is in good shape for the main scenarios envisaged by the F# Core Engineering Group. It has prevented F# users posting new, random packagings of FSharp.Core, and become a trusted part of the F# library ecosystem. Frequently, the package reference is managed by Paket, though the package also works well with NuGet tooling in IDE environments.

As of 2017, Microsoft have a set of time-critical objectives to deliver quality core tooling for F# as a default part of the dotnet SDK, see this RFC for discussion and pros/cons.

Up to this point, Microsoft have found it difficult to commit to a dependency on the community-provided FSharp.Core package in the dotnet SDK tooling for a number of reasons

  • The package contains some delay-signed DLLs (e.g. the Xamarin variations are delay-signed)

  • The package is now relatively large (33MB unzipped, 8MB zipped) and may get bigger (e.g. embedded PDBs).

  • The package has been prepared and pushed in an adhoc way

  • The package is not pre-installed with tooling (preventing some offline development scenarios)

  • The package is not easily buildable from a source-tarball (a legal requirement for Microsoft in some commercial settings)

Some of these problems are easily solvable, others are more work. As a result, early versions of F# tooling for .NET Core 1.x have used a smaller, signed package called Microsoft.FSharp.Core.netcore. This contains nothing but the .NET Standard 1.6 build of FSharp.Core and is small. As a first small step, Microsoft have agreed to rename this pack FSharp.Core.netstandard (or possibly FSharp.Core.netstandard1.6)

This means that prior to this PR the FSharp.Core NuGet package is not yet in sufficiently good shape to become a default assumption for the F# support embedded in the .NET SDK. That doesn't mean that the problems aren't solvable. There is long-term value in a unified, simple FSharp.Core package, and we can always iterate towards a better solution. That's just what we need to do.

Assumptions

For the purposes of this RFC we will assume

  • Portable profiles will eventually be legacy in favour of .NET Standard. (People building PCL DLLs will still be able to reference an earlier FSharp.Core package)

  • Xamarin programmability will eventually iterate towards .NET Standard, at least for the purposes of FSharp.Core. (People building Xamarin apps today will still be able to reference an earlier FSharp.Core package)

  • The runtime dependencies in the F# library ecosystem are based around DLL identity (not package identity).

  • The compile-time and script-execution-time dependencies in the F# library ecosystem are based on package identity.

FSharp.Core is a "root dependency" in the F# library ecosystem, both as a DLL with a version and strong name dependency, and, increasingly, as a package. Currently, F# libraries tend to assume either

  • The Profile 259 version of FSharp.Core, or

  • The fatter .NET Framework 4.x version of FSharp.Core

as their "root dependency". the general principle that libraries should have a "most portable sufficient" root dependency, it makes sense to move to a .NET Standard dependency.

It is very important to note that NuGet packages can easily be progressed from a current state to a different state without breaking existing consumers of specific versions:

  • a future version of a package can become "empty" and refer to a different package as an identity, effectively renaming the package (without breaking consumers of existing versions)

  • a future version of a package can emit a warning (without breaking consumers of existing versions)

  • a future version of a package can drop or add platforms and dependencies (without breaking consumers of existing versions)

  • a future version of a package can be signed by a different authority (without breaking consumers of existing versions)

Together this means that, no matter what the situation at any particular point in time, we can always iterate towards a better world.

Shared long term goals

All core engineering participants share some common long term goals

  • a simple experience for F# tooling users

  • a simpler set of FSharp.Core DLLs centered around a .NET Standard version of the library

  • the availability of a unified FSharp.Core package

  • ongoing binary compatibility for all existing users

  • a unified, sensible, healthy, "non-bifurcated" F# library ecosystem

  • mutual cooperation to see F# tooling succeed in many different scenarios

  • a healthy ecosystem of "innovative" F# tooling

Scenarios

Please contribute scenarios to the list below.

Using FSharp.Core NuGet package to simplify .NET Framework development. The F# community regularly use the existing FSharp.Core NuGet package to simplify and remove edge cases in .NET Framework development.

  1. A C# programmer wants to consume an F# library (which is assumed not to package FSharp.Core). Adding a reference to the package doesn't add an FSharp.Core reference to the C# project. THe F# programmer solves this situation by adding an FSharp.Core NuGet reference to their project, and republishing.

  2. Likewise, a C# programmer wants to consume multiple F# libraries where FSharp.Core dependencies need to be resolved by Paket or NuGet.

Proposal A

The following steps are proposed for the next few months, until about September 2017:

  1. The FSharp.Core package continue to be fully available and usable

  2. Microsoft publish FSharp.Core.netstandard (deprecating Microsoft.FSharp.Core.netcore). If time permits, Microsoft also publish FSharp.Core.netfx containing the .NET 4.x DLLs.

  3. If technically feasible (see below), F# Core Engineering add these packages as dependencies of a future version of FSharp.Core and drop the direct inclusion of DLLs. (As noted above these are not necesssarily permanent dependencies)

  4. FSharp.Core.netstandard is pre-loaded as a part of dotnet SDK tooling, making some degree of offline development possible.

  5. The default build logic for the intial set of F# project templates in the dotnet SDK will be to have a dependency only the FSharp.Core.netstandard package. It will, however, be possible to add FSharp.Core or other packages as a dependency instead.

Library authors will have a choice of depending on FSharp.Core(unified, fat), FSharp.Core.netstandard (minimal, needs some thought) or FSharp.Core.netfx (not quite minimal, needs some thought) as their FSharp.Core package reference.

.NET Framework Library authors may also simply have no package reference, as is common today.

Looking beyond ~September 2017, we propose:

  1. Iterate to improve the FSharp.Core package and ensure its continued usability.

  2. Assess the viability of making the .NET Standard version of FSharp.Core be the "basic assumed library" for the F# library ecosystem.

Proposal B

  1. Visual Studio deployed templates and the dotnet cli templates reference the FSharp.Core nuget package rather than Microsoft.FSharp.Core.netcore
  2. Visual F# Compiler and tools OSS repo (https://github.com/Microsoft/visualfsharp) host, builds and signs the FSharp.Core nuget packages and publishes them to nuget.org
  3. Visual F# compiler and tools OSS repo continues to update and publish FSharp.Core.nuget, 4.1.xxx with all of the PCLs including the Xamarin specific FSharp.Core.dlls until they are deprecated in the OSS repo (currently planned at end of year 2017).
  4. Visual F# compiler and tools OSS repo publishes FSharp.Core.nuget, 4.2.xxx This release contains the net45 and the netstandard 1.6 build of FSharp.Core.dll.

Guidance for developers

  • Existing packages targeting pcls, net20, or net40 use FSharp.Core.nuget versions 4.1.xxx
  • Existing desktop libraries or projects ... either package is fine, prefer FSharp.Core.nuget versions 4.2.xxx where feasible.
  • New desktop projects, Xamarin projects, or netstandard projects use: FSharp.Core.nuget versions 4.2.xxx
  • Library developers --- target as low a version of dotnet standard as your API consumption allows. netstandard1.6 is ideal for libraries not including type providers. Provide a net45 and netstandard build of your libraries, to enable developers who need to deploy to a wide range of existing Windows dotnet installs. FSharp.Core.nuget versions 4.2.xxx
  • TP developers you will need to target dotnet standard 2.0 and/or net45 --- but the netstandard1.6 profile of FSharp.Core will be ideal to build against use: FSharp.Core.nuget, 4.2.xxx

This approach appears to meet

  • Community scenario requirements

  • Microsoft scenario requirements

  • Microsoft size requrements

  • Microsoft publication criteria

  • Community simplicity requirements (for a single, unified FSharp.Core package with no additional dependencies).

Given the guidance that library authors should target lower versions of FSharp.Core in order to make their library more useful in more scenarios this seems like a reasonable tradeoff. Effectively it would be saying PCL library development is fine until .NET Standard library development is fully supported by all tooling. However please stick to referencing NuGet package FSharp.Core 4.1.17 or before, and by the way you will get slightly greater reach for your library if you use NuGet package 4.0.0.1 anyway.

Problems (Proposal A)

F# Core Engineering previously tried to make FSharp.Core be a unifying package by using a dependency on FSharp.Core.netstandard (then called Microsoft.FSharp.Core.netcore). For some reason that approach failed technically. It is a priority to determine what, if anything, goes wrong with doing that.

Specifically, this comment indicates that tooling can incorrectly interpret dependencies "FSharp.Core --> FSharp.Core.netstandard" in an incorrect way

It is technically very very hard to use different packages of FSharp.Core.dll for the same target framework. Is a special case not handled, and result in messed transitive dependencies and multiple FSharp.Core.dll referenced.

We should determine the exact nature of this problem and solve it. If it is not solvable in a reasonable time, we will need to assess the implications of that.

Open Questions (General)

There are a number of open questions about the way FSharp.Core is referenced by dotnet SDK templates and tooling. These are somewhat orthogonal to the package structure and delivery. See this comment.

  • Is the FSharp.Core package reference pinned or not in templates?

    • PackageReference support version ranges. Package versions in NuGet are major.minor.patch (semver), that's the official versioning scheme.
    • Pinning to an exact specific version, es 4.1.9 mean if i need to update it, user need to do it manually. And is bad for resilience to bugs. And is more complicated if implicit.
    • Pinning to a wilcard .path (eg 4.1.*) mean it's possibile to update it later.
    • Pinning to wilcard minor (eg 4.*) mean a more strict contract for the package.
    • Implicit version may help just give the minimal supported version.
    • All bundles (VS/Mono/cli) support offline packages (to not downlaod additional stuff). So this doesnt preclude open ranges, just mean new version need to be downloaded if needed.
  • Is the FSharp.Core reference explicit or implicit in .NET SDK project files? See again this comment

    • an Fsharp.Core.dll is always required. Template can implicit reference it or not.
    • If implicit, it must be possible to disable it.
    • Implicit is one line less in template.
    • Other package managers (like Paket) can just add that property as default, to manage FSharp.Core himself.
    • Explicit is easier to understand, and less surprises changing sdk version.

Open Questions (Proposal A)

  • Once FSharp.Core.netfx is available, will Mono, Xamarin and Visual Studio templates reference this?

  • At what point does the FSharp.Core package drop the inclusion of PCL versions of FSharp.Core (as mentioned above, PCL library development will still always be available by referencing older versions of the package)

    • Answer: when support for .NET Standard package references is widespread, stable and fully accepted
  • At what point does the FSharp.Core package drop the inclusion of Xamarin-specific versions of FSharp.Core (as mentioned above, Xamaring library development will still always be possible by referencing older versions of the package)

    • Answer: when Xamarin no longer needs these

Open Questions (Proposal B)

  • Is Xamarin development substantially affected?

Example F# Libraries

Here are a list of some sample F# libraries with .NET Standard or .NET Core compilations and an existing compile-time dependency on the FSharp.Core NuGet package

There are many others - searching github for nuget FSharp.Core in paket.dependencies is one way to find them. These libraries can be used to assess the suitability of proposals here. For example, does changing these libraries to an FSharp.Core.netstandard dependency bifurcate the world of F# libraries, or can a "FSharp.Core --> FSharp.Core.netstandard" dependency work to make the dependency chain commute?

Alternatives

  • The dotnet SDK takes a dependency on the existing versions of the FSharp.Core package.

    • Response: See reasons above for issues and what needs to be solved.
  • The F# world give up on a unified FSharp.Core package.

    • Response: No! See reasons above for why this is not sensible - there are just too many scenarios where is it is just "way simpler and much easier" to instruct users to reference this package.

Notes

  • file size more info and some stats, the 8MB nupkg is:
    • the 6.5% of .NET Core sdk bundle, and 0.28% of VS local nupkg feed
    • big (3-4 times) for a single library package

Acknowledgements

Thanks to Enrico Sada, Steffen Forkmann, Kevin Ransom, Phillip Carter for discussions leading up to the first draft of this RFC.