ConcurrentDictionary in .NET 4, not what you would expect

.NET 4 brings some awesome support for multithreaded programming as part of Task Parallel Library, Concurrent collections and few other types that live now in the framework.

Not all behaviour they expose is… not unexpected thought. What would be the result of the following code:

        
private static void ConcurrentDictionary()
{
	var dict = new ConcurrentDictionary<int, string>();
	ThreadPool.QueueUserWorkItem(LongGetOrAdd(dict, 1));
	ThreadPool.QueueUserWorkItem(LongGetOrAdd(dict, 1));
}
 
private static WaitCallback LongGetOrAdd(ConcurrentDictionary<int, string> dict, int index)
{
	return o => dict.GetOrAdd(index,i =>
										{
											Console.WriteLine("Adding!");
											Thread.SpinWait(1000);
 
											return i.ToString();
										}
					);
}

We call GetOrAdd method from ConcurrentDictionary<, > on two threads trying to concurrently obtain an item from the dictionary, creating it using provided delegate in case it’s not yet present in the collection. What would be the result of running the code?

Surprisingly both delegates will be invoked, so that we’ll see the message “Adding!” being printed twice, not once. This unfortunately makes the class unusable for a whole range of scenarios where you must ensure that creation logic is only ran once. One scenario where I wished I could use this class was Castle Dynamic Proxy and its proxy type caching. Currently it uses coarse grained double checked locking. The code could benefit from more elegant API like the one exposed by ConcurrentDictionary and fine grained locking.

However looks like ConcurrentDictionary is not an option, not unless coupled with Lazy<>, which while I suppose would work, is far less elegant and would not perform so good, since both classes would use their own, separate locking, so we’d end up using two locks instead of one, and having to go through additional, unnecessary layer of indirection.

So the message of the day is – new concurrent classes in .NET 4 are awesome but watch out for issues anyway.