It happens every so often. Someone is trying to use the WebDriver .NET language bindings, and it happens that their project also uses the excellent Json.NET library created by James Newton-King. They decide they want to use NuGet for managing their dependencies. Everything moves along swimmingly until there's an update to the Json.NET library. NuGet dutifully updates the dependency, but when they go to run their WebDriver tests, and everything blows up. After a cursory investigation, they see that the NuGet package for the Selenium WebDriver .NET bindings have a dependency on the Json.NET package, but the package declares the dependency on a specific version of Json.NET. When people see this, they inevitably point out that there must be an error in the authoring of the Selenium.WebDriver package. Hopefully, I'll be able to explain exactly why this isn't an error in the package authoring, and why naive proposed solutions won't solve this.
First, a little background. The build process of the .NET bindings is integrated with the build process for the other portions of the Selenium project. This means the released binaries aren't built through Visual Studio, and aren't built via MSBuild. The Selenium build script uses Rake as it's build engine, via the Albacore project. We have some unique requirements that don't allow us to just build the .csproj file and be done with it. As part of the build process, we need to embed the outputs of other targets (which are not built via Visual Studio or MSBuild) as resources into the .NET assembly. To accomplish this, we end up calling csc.exe with the appropriate command-line switches. This is a nicely stable build process, follows the patterns of the build for the other portions of the Selenium project, and doesn't require hand-development and maintenance of a custom MSBuild project file.
As part of the build process, we also sign the assembly to give it a strong name. This was a feature request of users who wished to reference the WebDriver .NET bindings from a project that also had a strong name. You see, when you create a strong-named assembly, any assemblies that you reference must also be strong-named. Part of the strong name of such an assembly is its version number. When your strong-named assembly tries to resolve the reference to another strong-named assembly, but the only copy you have of the referenced assembly is of a different version, the .NET Framework cheerfully tells you, "I don't have a copy of the assembly you're looking for." Which you don't, because the name of the assembly (which contains the version information because it's a strong name) doesn't match.
Complicating this behavior is that if your assembly is unsigned, it is free to reference assemblies with and without strong names. Furthermore, the unsigned assembly will load referenced assemblies matched by file name only, happily ignoring the version information in the referenced assembly. Since many people don't bother giving their assemblies strong names, they never have to think about the versions of the assemblies they reference.
I'm sure those of you who've encountered this before are already far ahead of me. Both the WebDriver .NET bindings assembly and the Json.NET assembly have strong names. That means that the WebDriver assembly can use only the version of the Json.NET assembly that it was compiled against. When building the NuGet package for the .NET bindings, we must restrict the packages we reference to the exact versions we compile against, otherwise the .NET bindings won't be able to load the referenced assembly if NuGet downloads the latest version of Json.NET, and it's not the version we've compiled against. Of course, that's a problem if your project is unsigned and also requires a reference to Json.NET, and your reference isn't similarly scoped with respect to version, because now you have a version conflict. This is why simply changing the version specification for the Json.NET reference in the WebDriver .nuspec file to be a "greater than or equal to version x" vs. "be exactly x" won't solve the problem.
Some people suggest that migrating the project to use NuGet to manage its dependencies is a viable solution to this problem. Given all that we already require to successfully build the project, I remain unconvinced that adding one more technology that prospective contributors have to know about is not the friendliest approach. This is particularly so if the prospective contributor in question has no need or desire to use NuGet in his or her own projects. Getting people to engage and contribute is hard enough without throwing other roadblocks in their way (but that's another rant for another day).
It should be noted that this is a well-known issue with creating NuGet packages with strong-named assemblies. As far as I'm aware, no one has come up with a good solution. The most promising solution would be to change the release of the WebDriver assemblies
to be unsigned. However, this would amount to a change in functionality for anyone who
is referencing them from a signed assembly, and would likely frustrate them just as much as those who might be frustrated by the current state of affairs in the NuGet package.
Update 6 July 2012: Apparently, James Newton-King, the author of the Json.NET library, has been experimenting with a solution to this problem. I'll be modifying the .NET bindings NuGet package to take advantage of this soon. There's still a fundamental design flaw in the mismatch between packaging and strong-named assemblies, but for a widely-used, frequently-updated library like Json.NET, this is likely the best solution.
Update 6 July 2012: Apparently, James Newton-King, the author of the Json.NET library, has been experimenting with a solution to this problem. I'll be modifying the .NET bindings NuGet package to take advantage of this soon. There's still a fundamental design flaw in the mismatch between packaging and strong-named assemblies, but for a widely-used, frequently-updated library like Json.NET, this is likely the best solution.
No comments:
Post a Comment