Another coding horror story was reported in the microsoft.public.dotnet.xml newsgroup:
I've been experiencing OutOfMemory errors on our prodution webserver for a few weeks now. I've finally managed to isolate (I think) the problem to our use of c# script blocks in our xsl files.
While debugging I discovered that the app domain for one of our sites had 13000+ assemblies loaded.
Cool. This is just a remainder for those who use XSLT scripting (msxsl:script) in .NET: watch out, this feature can be pure evil if used unwisely - it leaks memory and there is nothing you can do about it.
The problem is that when XSLT stylesheet is loaded in .NET, msxsl:script is compiled into an assembly via CodeDOM and then loaded into memory, into the current application domain. Each time the stylesheet is loaded above process is repeated - new assembly is being generated and loaded into the application domain. But it's impossible to unload an assembly from application domain in .NET!
Here is KB article on the topic. It says it applies to .NET 1.0 only, but don't be confused - the problem exists in .NET 1.1 and 2.0. Moreover I'm pretty much pessimistic about if it's gonna be fixed in the future.
The solution is simple - just don't use script in XSLT unless you really really really have to. Especially on the server side - XSLT script and ASP.NET should never meet unless you take full resonsibility for caching compiled XslCompiledTransform. Use XSLT extension objects instead.
Update. Of couse Yuriy reminds me that msxsl:script runs faster than an extension object, because msxsl:script is available at compile time and so XSLT compiler can generate direct calls, while extension objects are only available at run-time and so can only be called via reflection.
That makes msxsl:script a preferrable but danger solution when your stylsheet makes lots of calls to extension functions.
In a perfect world of course msxsl:script would be compiled into dynamic methods (just like XSLT itself), which are GC reclaimable, but I don't think CodeDOM is capable of doing this currently. I wonder if it's possible to compile C#/VB/J# method source into dynamic method anyway?
Also it's interesting how to improve extension objects performance - what if extension objects could be passed at compile time? They are usually available anyway at that time too. Or what if compiled stylesheet could be "JITted" to direct calls instead of reflection?
Sergey, Anton, can you please comment on this?