An ASP.NET vNext Build and Repo Structure Braindump

aspnet, dotnet, autofac, github comments edit

Alex and I are working on switching Autofac over to ASP.NET vNext and as part of that we’re trying to figure out what the proper structure is for a codeline, how a build should look, and so on.

There is a surprisingly small amount of documentation on the infrastructure bits. I get that things are moving quickly but the amazing lack of docs of any detail creates for a steep learning curve and a lot of frustration. I mean, you can read about the schema for project.json but even that is out of date/incomplete so you end up diving into the code, trying to reverse-engineer how things come together.

Below is a sort of almost-stream-of-consciousness braindump of things I’ve found while working on sorting out build and repo structure for Autofac.

No More MSBuild - Sake + KoreBuild

If you’re compiling only on a Windows platform you can still use MSBuild, but if you look at the ASP.NET vNext repos, you’ll see there’s no MSBuild to be found.

This is presumably to support cross-platform compilation of the ASP.NET libraries and the K runtime bits. That’s a good goal and it’s worth pursuing - we’re going that direction for at least core Autofac and a few of the other core libs that need to change (like Autofac.Configuration). Eventually I can see all of our stuff switching that way.

The way it generally works in this system is:

  • A base build.cmd (for Windows) and build.sh (for Linux) use NuGet to download the Sake and KoreBuild packages.
  • The scripts kick off the Sake build engine to run a makefile.shade which is platform-agnostic.
  • The Sake build engine, which is written in cross-platform .NET, handles the build execution process.

The Sake Build System

Sake is a C#-based make/build system that appears to have been around for quite some time. There is pretty much zero documentation on this, which makes figuring it out fairly painful.

From what I gather, it is based on the Spark view engine and uses .shade view files as the build scripts. When you bring in the Sake package, you get several shared .shade files that get included to handle common build tasks like updating assembly version information or running commands.

It enables cross-platform builds because Spark, C#, and the overall execution process works both on Mono and Windows .NET.

One of the nice things it has built in, and a compelling reason to use it beyond the cross-platform support, is that a convention-based standard build lifecycle that runs clean/build/test/package targets in a standard order. You can easily hook into this pipeline to add functionality but you don’t have to think about the order of things. It’s pretty nice.

The KoreBuild Package

KoreBuild is a build system layered on top of Sake that is used to build K projects. As with Sake, there is zero doc on this.

If you’re using the new K build system, though, and you’re OK with adopting Sake, there’s a lot of value in the KoreBuild package. KoreBuild layers in Sake support for automatic NuGet package restore, native compile support, and other K-specific goodness. The _k-standard-goals.shade file is where you can see the primary set of things it adds.

The Simplest Build Script

Assuming you have committed to the Sake and KoreBuild way of doing things, you can get away with an amazingly simple top-level build script that will run a standard clean/build/test/package lifecycle automatically for you.

var AUTHORS='Your Authors Here'

use-standard-lifecycle
k-standard-goals

At the time of this writing, the AUTHORS value must be present or some of the standard lifecycle bits will fail… but since the real authors for your package are specified in project.json files now, this really just is a placeholder that has to be there. It doesn’t appear to matter what the value is.

Embedded Resources Have Changed

There is currently no mention of how embedded resources are handled in the documentation on project.json but if you look at the schema you’ll see that you can specify a resources element in project.json the same way you can specify code.

A project with embedded resources might look like this (minus the frameworks element and all the dependencies and such to make it easier to see):

{
    "description": "Enables Autofac dependencies to be registered via configuration.",
    "authors": ["Autofac Contributors"],
    "version": "4.0.0-*",
    "compilationOptions": {
        "warningsAsErrors": true
    },
    "code": ["**\\*.cs"],
    "resources": "**\\*.resx"
    /* Other stuff... */
}

Manifest Resource Path Changes

If you include .resx files as resources, they correctly get converted to .resources files without doing anything. However, if you have other resources, like an embedded XML file…

{
    "code": ["**\\*.cs"],
    "resources": ["**\\*.resx", "Files\\*.xml"]
}

…then you get an odd generated path. Easiest to see with an example. Say you have this:

~/project/
  src/
    MyAssembly/
      Files/
        Embedded.xml

In old Visual Studio/MSBuild, the file would be embedded and the internal manifest resource stream path would be MyAssembly.Files.Embedded.xml - the folders would represent namespaces and path separators would basically become dots.

However, in the new world, you get a manifest resource path Files/Embedded.xml - literally the relative path to the file being embedded. If you have unit tests or other stuff where embedded files are being read, this will throw you for a loop.

No .resx to .Designer.cs

A nice thing about the resource system in VS/MSBuild was the custom tool that would run to convert .resx files into strongly-typed resources in .Designer.cs files. There’s no automatic support for this anymore.

However, if you give in to the KoreBuild way of things, they do package an analogous tool inside KoreBuild that you can run as part of your command-line build script. It won’t pick up changes if you add resources to the file in VS, but it’ll get you by.

To get .resx building strongly-typed resources, add it into your build script like this:

var AUTHORS='Your Authors Here'

use-standard-lifecycle
k-standard-goals

#generate-resx .resx description='Converts .resx files to .Designer.cs' target='initialize'

What that does is add a generate-resx build target to your build script that runs during the initialize phase of the standard lifecycle. The generate-resx target dependes on a target called resx which does the actual conversion to .Designer.cs files. The resx target comes from KoreBuild and is included when you include the k-standard-goals script, but it doesn’t run by default, which is why you have to include it yourself.

Gotcha: The way it’s currently written, your .resx files must be in the root of your project (it doesn’t use the resources value from project.json). They will generate the .Designer.cs files into the Properties folder of your project. This isn’t configurable.

ASP.NET Repo Structure is Path of Least Resistance

If you give over to Sake and KoreBuild, it’s probably good to also give over to the source repository structure used in the ASP.NET vNext repositories. Particularly in KoreBuild there are some hardcoded assumptions in certain tasks that you’re using that repo structure.

The structure looks like this:

~/MyProject/
  src/
    MyProject.FirstAssembly/
      Properties/
        AssemblyInfo.cs
      MyProject.FirstAssembly.kproj
      project.json
    MyProject.SecondAssembly/
      Properties/
        AssemblyInfo.cs
      MyProject.SecondAssembly.kproj
      project.json
  test/
    MyProject.FirstAssembly.Test/
      Properties/
        AssemblyInfo.cs
      MyProject.FirstAssembly.Test.kproj
      project.json
    MyProject.SecondAssembly.Test/
      Properties/
        AssemblyInfo.cs
      MyProject.SecondAssembly.Test.kproj
      project.json
  build.cmd
  build.sh
  global.json
  makefile.shade
  MyProject.sln

The key important bits there are:

  • Project source is in the src folder.
  • Tests for the project are in the test folder.
  • There’s a top-level solution file (if you’re using Visual Studio).
  • The global.json points to the src file as the place for project source.
  • There are build.cmd and build.sh scripts to kick off the cross-platform builds.
  • The top-level makefile.shade handles build orchestration.
  • The folder names for the source and test projects are the names of the assemblies they generate.
  • Each assembly has…
    • Properties with AssemblyInfo.cs where the AssemblyInfo.cs doesn’t include any versioning information, just other metadata.
    • A .kproj file (if you’re using Visual Studio) that is named after the assembly being generated.
    • A project.json that spells out the authors, version, dependencies, and other metadata about the assembly being generated.

Again, a lot of assumptions seem to be built in that you’re using that structure. You can save a lot of headaches by switching.

I can see this may cause some long-path problems. Particularly if you are checking out code into a deep file folder and have a long assembly name, you could have trouble. Think…

C:\users\myusername\Documents\GitHub\project\src\MyProject.MyAssembly.SubNamespace1.SubNamespace2\MyProject.MyAssembly.SubNamespace1.SubNamespace2.kproj

That’s 152 characters right there. Add in those crazy WCF-generated .datasource files and things are going to start exploding.

Assembly/Package Versioning in project.json

Part of what you put in project.json is your project/package version:

{
    "authors": ["Autofac Contributors"],
    "version": "4.0.0-*",
    /* Other stuff... */
}

There desn’t appear to be a way to keep multiple assemblies in a solution consistently versioned. That is, you can’t put the version info in the global.json at the top level and I’m not sure where else you could store it. You could probably come up with a custom build task to handle centralized versioning, but it’d be nice if there was something built in for it.

XML Doc Compilation Warnings

The old compiler csc.exe had a thing where it would automatically output compiler warnings for XML documentation errors (syntax or reference errors). The K compiler apparently doesn’t do this by default so they added custom support for it in the KoreBuild package.

To get XML documentation compilation warnings output in your build, add it into your build script like this:

var AUTHORS='Your Authors Here'

use-standard-lifecycle
k-standard-goals

#xml-docs-test .clean .build-compile description='Check generated XML documentation files for errors' target='test'
  k-xml-docs-test

That adds a new xml-docs-test target that runs during the test part of the lifecycle (after compile). It requires the project to have been cleaned and built before running. When it runs, it calls the k-xml-docs-test target to manually write out XML doc compilation warnings.

Runtime Update Gotchas

Most build.cmd or build.sh build scripts have a line like this:

CALL packages\KoreBuild\build\kvm upgrade -runtime CLR -x86
CALL packages\KoreBuild\build\kvm install default -runtime CoreCLR -x86

Basically:

  • Get the latest K runtime from the feed.
  • Set the latest K runtime as the ‘default’ one to use.

While I think this is fine early on, I can see a couple of gotchas with this approach.

  • Setting the ‘default’ modifies the user profile. When you call kvm install default the intent is to set the aliast default to refer to the specified K runtime version (in the above example, that’s the latest version). When you set this alias, it modifies a file attached to the user profile containing the list of aliases - it’s a global change. What happens if you have a build server environment where lots of builds are running in parallel? You’re going to get the build processes changing aliases out from under each other.
  • How does backward compatibility work? At this early stage, I do want the latest runtime to be what I build against. Later, though, I’m guessing I want to pin a revision of the runtime in my build script and always build against that to ensure I’m compatible with applications stuck at that runtime version. I guess that’s OK, but is there going to be a need for some sort of… “binding redirect” (?) for runtime versions? Do I need to specify some sort of “list of supported runtime versions?”

Testing Means XUnit and aspnet50

At least at this early stage, XUnit seems to be the only game in town for unit testing. The KoreBuild stuff even has XUnit support built right in, so, again, path of least resistance is to switch if you’re not already on it.

I did find a gotcha, though, where if you want k test to work your assemblies must target aspnet50.

Which is to say… in your unit test project.json you’ll have a line to specify the test runner command:

{
    "commands": {
        "test": "xunit.runner.kre"
    },
    "frameworks": {
        "aspnet50": { }
    }
}

Specifying that will allow you to drop to a command prompt inside the unit test assembly’s folder and run k test to execute the unit tests.

In early work for Autofac.Configuration I was trying to get this to work with the Autofac.Configuration assembly only targeting aspnetcore50 and the unit test assembly targeting aspnetcore50. When I ran k test I got a bunch of exceptions (which I didn’t keep track of, sorry). After a lot of trial and error, I found that if both my assembly under test (Autofac.Configuration) and my unit test assembly (Autofac.Configuration.Test) both targeted aspnet50 then everything would run perfectly.

PCL Support is In Progress

It’d be nice if there was a portable class library profile that just handled everything rather than all of these different profiles + aspnet50 + aspnetcore50. There’s not. I gather from Twitter conversations that this may be in the works but I’m not holding my breath.

Also, there’s a gotcha with Xamarin tools: If you’re using a profile (like Profile259) that targets a common subset of a lot of runtimes including mobile platforms, then the output of your project will change based on whether or not you have Xamarin tools installed. For example, without Xamarin installed you might get .nupkg output for portable-net45+win+wpa81+wp80. However, with Xamarin installed that same project will output for portable-net45+win+wpa81+wp80+monotouch+monoandroid.

Configuration Changes

Obviously with the break from System.Web and some of the monolithic framework, you don’t really have web.config as such anymore. Instead, the configuration system has become Microsoft.Framework.ConfigurationModel.

It’s a pretty nice and flexible abstraction layer that lets you specify configuration in XML, JSON, INI, or environment variable format. You can see some examples here.

That said, it’s a huge change and takes a lot to migrate.

  • No appSettings. I’m fine with this because appSettings always ended up being a dumping ground, but it means everything you have originally tied to appSettings needs to change.
  • No ConfigurationElement objects. I can’t tell you how much I have written in the old ConfigurationElement mechanism. It had validation, strong type parsing, serialization, the whole bit. All of that will not work in this new system. You can imagine how this affects things like Autofac.Configuration.
  • List and collection support is nonexistent. I’ve actually filed a GitHub issue about this. A lot of the configuration I have in both Autofac.Configuration and elsewhere is a list of elements that are parameterized. The current XML and JSON parsers for config specifically disallow list/collection support. Everything must be a unique key in a tree-like hierarchy. That sort of renders the new config system, at least for me, pretty much unusable except for the most trivial of things. Hopefully this changes.
  • Everything is file or in-memory. There’s no current support for pulling in XML or JSON configuration that comes from, say, a REST API call from a centralized repository. Even in unit testing, all the bits that actually run the configuration parsing on a stream of XML/JSON are internals rather than exposed - you have to load config from a file or manually create it yourself in memory by adding key/value pairs. There’s a GitHub issue open for this, too.

As a workaround, I’m considering using custom object serialization and bypassing the new configuration system altogether. I like the flexibility of the new system but the limitations are pretty overwhelming right now.

Comments