During the last year, I’ve been a part of launching two production sites that run on ASP.NET Core, and as a company, we’ve had enough dealings with the budding framework that we arranged a full day’s seminar on the topic.
Needless to say, using a framework in anger at this point of its development has led to all kinds of interesting discoveries, the kind that you typically only ever make on the bleeding edge.
Where have my assemblies gone?
One of the major changes in .NET Core compared to the full .NET Framework is that there is no more Global Assembly Cache. All assemblies – including most if not all of the runtime itself – will be shipped as NuGet packages, which means that the assembly loading story is a fairly major departure from the way things used to be. However, .NET Core is not always a viable platform: for instance, currently there is no production-ready server-side image processing capability since System.Drawing is not cross-platform*. Given that constraint, we’ve ended up deploying our production ASP.NET Core applications on the full .NET framework, and the full FX still has the GAC.
Currently, ASP.NET Core on the full FX loads assembly dependencies by hooking up AppDomain.AssemblyResolve to work its magic. When your code tries to interact with an assembly that is not yet loaded, the runtime looks for the assembly in your NuGet packages. However, there’s a key phrase in the documentation for the event: “Occurs when the resolution of an assembly fails.” This means that regular assembly binding rules are attempted first.
Typically, this would not be a problem. When you deploy your application, you deploy the NuGet dependencies, and the GAC only contains the framework’s assemblies. However, sometimes you will have a rogue application on your system that installs something to the GAC, and things may go a little pear-shaped.
DocumentDB deserialization woes
Consider this example: our app uses Azure DocumentDB as one of its data stores. The .NET DocumentDB client library uses JSON as its serialization format, and in particular, Newtonsoft.Json as its serialization library. One of the things you can do with that combination is specify that the serialized name of your property is different from the one declared in code, by annotating the property with JsonPropertyAttribute. Now, our app opted to use one of the latest builds of Newtonsoft.Json (7.x), and for the most part, everything worked beautifully. However, my development system had an installed app that, unbeknownst to me, registered an older version of Newtonsoft.Json into the GAC.
Unfortunately, the runtime assembly version of the GAC assembly matched the requirements of the DocumentDB client library exactly, so the runtime loaded that assembly for the DocumentDB client. The practical effect was that when the DocumentDB client (de)serialized objects, it never noticed the JsonPropertyAttribute that we were using. The net result? A single property in that class was never (de)serialized correctly.
It took me a while to figure out what was happening, but the key insight was to look at the loaded modules in the debugger and notice that indeed, we now had two copies of Newtonsoft.Json in memory: the version from the GAC and the version we were specifying as a dependency. Our own code was using the JsonPropertyAttribute from version 7.x whereas the older version of Newtonsoft.Json was looking for the JsonPropertyAttribute specified in that assembly. While the properties were identical in function, they were different in identity, so the attribute we were using was ignored entirely.
Wait, isn’t this a solved problem already?
If you’re a seasoned .NET developer, at this point you are probably thinking “binding redirects”. At least we were – but the question was, where to put them? Another major change in ASP.NET Core is that your application configuration is entirely decoupled from both the configuration of the runtime and the configuration of your web server. Which means that in a fresh out-of-the-box web application, you do have a web.config, but it is only used to configure the interaction between IIS and your application server, Kestrel.
Since Kestrel is running in a process outside IIS, it’s reasonable to expect that Web.config doesn’t affect the behavior of the runtime in that process. And indeed, it doesn’t. But the new configuration system doesn’t have a way to specify the configuration of the .NET runtime either. So what does that leave us?
After a little bit of to-and-fro with the ASP.NET Core team, the answer finally came up: the runtime configuration still exists, but its naming conventions are different from what we are used to. If you create a file called App.config (yes, even when it is a web application) and specify your binding redirects there, they will be picked up, and all is well in the world again.
The configuration file has the same schema as you would expect from a Web.config or a standalone executable’s App.config. The resulting file looks like this:
<?xml version=”1.0″ encoding=”utf-8″?>
<assemblyIdentity name=”Newtonsoft.Json” culture=”neutral” publicKeyToken=”30ad4fe6b2a6aeed” />
<bindingRedirect oldVersion=”0.0.0.0-18.104.22.168″ newVersion=”22.214.171.124″ />
Hope this helps anyone else encountering the same problem, however uncommon it may be!
(* My colleagues pointed out that I neglected to mention the fact that System.Drawing is not a production-ready server-side image processing solution either, given that it uses GDI+ which uses process-wide locks, and therefore essentially makes the image-processing parts of your app single-threaded.)