Unexpected benefits of Precompilation of LINQ

I once had a manager who told me – I can solve any maintenance problem by adding a layer of abstraction.  I can solve any performance problem by removing a layer of abstraction.

I think LINQ to SQL is a wonderful way to abstract the persistence layer elegant, easy to use, easy to manipulate, and easy to maintain lines of code.  Instead of writing SQL which amounts to “how to retrieve” the data – you manipulate an expression tree that gets closer to specifying “what data I want”.  The upside of this is huge – you can change the expression tree at any level of your code, and let .NET decide how to best write the SQL at the last possible moment – which effectively gits rid of dealing with intermediate results and inefficiently written SQL.  Unfortunately, this abstraction does indeed cause a performance hit – the translation/compilation of the tree to SQL – and it’s probably much bigger than you would think.  See http://peterkellner.net/2009/05/06/linq-to-sql-slow-performance-compilequery-critical/ to see what I mean.  In my analysis (using ANTS Profiler), when using uncompiled LINQ – the performance hit is usually about 80% compilation and only 20% retrieving the data!  Thankfully, .NET does allow you to precompile a LINQ query and save the compilation to use over and over again.

Your natural tendency when hearing those kind of numbers might be to precompile every single LINQ query you write.  There’s a big downside to doing that, though – you lose the ability to manipulate the compiled query in other parts of your code.  Another downside is that the precompilation code itself is fairly ugly and hard to read/maintain.

I’m a big believer in avoiding “premature optimization”.  What happens if you precompile everything, and in a version or two Microsoft resolves the issue and caches compilations for you behind the scenes?  You have written a ton of ugly code that breaks a major benefit of LINQ to SQL and is totally unnecessary.

Instead, I recommend you go after the low hanging fruit first – precompile the most frequently accessed queries in your application and the ones that gain no benefit from manipulating the expression tree.  In the applications I work on – there is a perfect case that fits both of these – the “get” method that returns the LINQ object representation of a single row in the database.  These are hit quite often – and there is absolutely no case where the expression tree is further refined.

The old way it was written:

	public static Item Get(int itemid) {
		return (from i in DataContext.Items
			where i.ItemID == itemid
		select i).First();
	}

The new way with Precompiled LINQ:

	private static Func<modeldatacontext, int,item="">
		GetQuery = CompiledQuery.Compile(
			(ModelDataContext DataContext,int itemid) =>

				(from i in DataContext.Items
					where i.ItemID == itemid
				select i).First()

				);

	public static Item Get(int itemid) {
		return GetQuery.Invoke(DataContext,itemid);
	}

Applying this fairly simple change to the application, I’d estimate we got 80%+ of the benefits of compiled LINQ, at the expense of a few extra lines of code per object/table and absolutely no loss of the tree manipulation.