When I pick up a new framework or library, there's usually that learning curve where I get familiar with its API, find what works, what doesn't work, etc. One habit I've gotten into is that I create a TestFixture for everything i think i should be able to do and build a test for that assumption. The purpose of these tests is both to make sure the code does what I expect it to, but also to serve as a record of what I've already learned. If i later on wonder how some call would function, i first check my test signatures, to see if i've already tested that behavior. If there is an appropriate test, I immediately know what the behavior will be, plus i now have working sample code or how to do it.
For example, I was playing around with setting up Moq's through Autofac and wanted to come up with a registration that would give me a container scoped Moq object that i could set up before executing a particular test. The resulting test looked like this:
public interface IMockWithAccessor
{
IMockAccessorValue Accessor { get; }
}
public interface IMockAccessorValue
{
string Foo { get; }
}
[Test]
public void Create_nested_mock_so_it_can_be_altered_in_container_scope()
{
var builder = new ContainerBuilder();
builder.Register(c => new Mock<IMockAccessorValue>())
.As<Mock<IMockAccessorValue>>().ContainerScoped();
builder.Register(c => c.Resolve<Mock<IMockAccessorValue>>().Object)
.As<IMockAccessorValue>().ContainerScoped();
builder.Register(c =>
{
var mockBuilder = new Mock<IMockWithAccessor>();
mockBuilder.Setup(x => x.Accessor)
.Returns(c.Resolve<IMockAccessorValue>());
return mockBuilder.Object;
}).As<IMockWithAccessor>().ContainerScoped();
using (var container = builder.Build().CreateInnerContainer())
{
var mockAccessorBuilder = container
.Resolve<Mock<IMockAccessorValue>>();
mockAccessorBuilder.Setup(x => x.Foo).Returns("bar");
Assert.AreEqual("bar", container
.Resolve<IMockWithAccessor>().Accessor.Foo);
}
}
Sometimes, of course, my expectations are not met and the code does not allow me to do what i set out to do. These test are even more valuable for future reference, as long as i make sure to rename the test to reflect the failed expectation, and alter the asserts to reflect the actual behavior.
I was trying to figure out parametrized component registrations in Autofac. The example showed it being used with FactoryScope. I wondered whether, in default (Singleton) scope, Autofac would use the parameters to create singletons per parameter set. My original test was named Parametrized_resolve_creates_different_singleton_per_parameter_Value. Well, it turned out that, no, autofac does not vary singletons, and parametrized registrations only make sense in FactoryScope. The final test looks like this:
public class ParametrizedSingleton { }
[Test]
public void Parametrized_resolve_without_factory_scope_is_always_a_singleton()
{
var builder = new ContainerBuilder();
builder.Register((c, p) => new ParametrizedSingleton());
using (var container = builder.Build())
{
var foo1 = container.Resolve<ParametrizedSingleton>(
new NamedParameter("type", "foo"));
var foo2 = container.Resolve<ParametrizedSingleton>(
new NamedParameter("type", "foo"));
var bar1 = container.Resolve<ParametrizedSingleton>(
new NamedParameter("type", "bar"));
Assert.AreSame(foo1, foo2);
Assert.AreSame(foo1, bar1);
}
}
I usually keep these test fixtures in a separate test project in my solution as permanent reference, as i continue to develop the main code. It's proven useful a number of times when coming back to some old code and having to reacquaint myself with a third party bit of code.
In my last post I proposed using delegates instead of interfaces to declare dependencies for injection. While delegates are limited to a single function call, this is often sufficient for service dependencies. In addition, this is not a wholesale replacement for traditional IoC, since at the end of the day, you have to have some class instances that provide the methods bound by the delegates and want to return instances require those delegates, so our container will still need to resolve class instances.
The main benefit of using delegates instead of interfaces is that delegates do not impose any requirements on the providing class, and thereby dependencies can be defined by what the client needs rather than what a service provides.
To illustrate this mapping, let's bring back the classes defined in the last post:
public class MessageQueue
{
public void Enqueue(string recipient, string message) { ... }
public string TryDequeue(string recipient) { ... }
}
public class Producer : IProducer
{
public delegate void EnqueueDelegate(string recipient, string message);
public Producer(EnqueueDelegate dispatcher) { ... }
}
public class Consumer : IConsumer
{
public delegate string TryDequeueDelegate(string recipient);
public Consumer(TryDequeueDelegate inbox) { ... }
}
What we need is a way to map a delegate to a method:
Aside from the fact that the above is not a legal C# syntax, the above has the implicit assumption that we can resolve a canonical instance of MessageQueue, since MessageQueue.Enqueue really refers to a method on an instance of MessageQueue. I.e. our container must function like a regular IoC container so we can resolve that instance.
The above solves the scenario of a mapping one delegate to one implementation. In addition, we'd probably want the flexibility to map a particular implementation to a particular client class, such that:
The usage scenarios described are simple enough to understand. If our container were initialized using strings or some external configuration file (Xml, custom DSL, etc.), the actual reflection required for the injection of mappings isn't too complex either. However, I abhor defining typed mappings without type-safety. This is isn't so much about making mistakes in spelling, etc. that the compiler can't catch. It is mostly about being able to navigate the mappings and the dependencies they describe and the ability to refactor while keeping the mappings in sync.
It would be great if we could just say:
Note: I'm purposely using a syntax that mimics Autofac, since the implementation later on will be done as a hack on top of Autofac
That looks fine and could even work in the face of polymorphism, since knowing the signature of Producer.EnqueueDelegate we could reflect the proper overload of MessageQueue.Enqueue. However, C# has no syntax for getting at the MethodInfo of a method via Generics (it is not a type). There isn't even an equivalent to typeof(T) for members, the reason for which was well explained by Eric Lippert in In Foof We Trust: A Dialogue. The only way to get MethodInfo relies on string based reflection.
Fortunately, C#3.0 introduced a syntax that allows us to capture method calls as expression trees that we can decompose programatically. This lets us to express our method call like this:
This expression conveniently infers the type of a and b. As a sidenote, Producer.EnqueueDelegate does not mean that EnqueueDelegate is a member of Producer. It's just a syntax artifact of nested declarations in C#, which in this case conveniently makes the delegate look attached to the class.
Unfortunately, there we can't just include MessageQueue in the parameter list of the lambda. If we were to include it in the argument list, it could not be inferred, and if we were to define MessageQueue explicitly as a lambda parameter, then we'd be forced to declare all arguments. We want to express the above and only explicitly define MessageQueue. To accomplish this we need to create a composite expression that previously was told about MessageQueue:
Expression<Func<MessageQueue, Expression<Producer.EnqueueDelegate>>> expr
= x => (a,b)=> messageQueue.Enqueue(a,b);
Now we have enough syntactic sugar to describe our two registration scenarios in terms of the container builder. First, the global registration of the delegate against an implementation:
builder.Define<Consumer.TryDequeueDelegate>()
.As<MessageQueue>(x => a => x.TryDequeue(a));
builder.Define<Producer.EnqueueDelegate>()
.As<MessageQueue>(x => (a, b) => x.Enqueue(a, b));
And alternately, the registration of delegates and their implementation in the context of a particular class:
builder.Register<Consumer>().As<IConsumer>()
.With<Consumer.TryDequeueDelegate>()
.As<MessageQueue>(x => a => x.TryDequeue(a));
builder.Register<Producer>().As<IProducer>()
.With<Producer.EnqueueDelegate>()
.As<MessageQueue>(x => (a, b) => x.Enqueue(a, b));
Next time, I'll go over the implementation of the above to get it working as an injection framework.
I've been on a bit of a tear about declaring dependency contracts and injecting only what is required. While examining the use of Interfaces in IoC and their shortcomings, I decided that taken to the extreme, dependencies come down to call dependencies, which could be modeled with delegates rather than interfaces. Instead of writing a novel, as I've been prone to, i thought I'd do a shorter post on my approach to this solution, and expand on the implementation in later posts.
To recap, in the SOLID principles, the Interface Segregation Principle states: Clients should not be forced to depend upon interfaces that they do not use. This means that interfaces should be fine-grained enough to expose no more than one responsibility. Taken to the extreme, this could be taken to mean that each interface only has a single method. There are valid SRP scenarios where a responsibility is modeled by more than one call, but let's start with the simplest scenario first, then see how well it applies to more complex responsibilities later.
In C# we have delegates, which describe a single method call. A delegate instance is a reference to a method that encapsulates a specific instance of a class, without exposing the underlying class (unless your delegate is a static method). A delegate can even be used to expose internal, protected and private methods.
Instead of declaring a list of interfaces that the IoC container should inject, classes would define their dependencies as delegates. Taking the example from my duck typing post, we would get the following dependency declarations.
First, we have the same service provider, MessageQueue, which still doesn't need to implement an interface:
public class MessageQueue
{
public void Enqueue(string recipient, string message) { ... }
public string TryDequeue(string recipient) { ... }
}
Next, we have the new Producer, now declaring its dependency has a delegate:
public class Producer : IProducer
{
public delegate void EnqueueDelegate(string recipient, string message);
public Producer(EnqueueDelegate dispatcher) { ... }
}
And finally, we have the new Consumer, also declaring a delegate for construction time injection:
public class Consumer : IConsumer
{
public delegate string TryDequeueDelegate(string recipient);
public Consumer(TryDequeueDelegate inbox) { ... }
}
Think of the delegate as your Method Interface. You could define your dependencies as Func's and Action's, but that would obfuscate your dependencies beyond recognition in most scenarios. By using an explicit delegate, you get to attach the dependency to the class that has the dependency, in addition to having a descriptive signature.
Now, if we were to wire this up manually we'd get something like this:
var queue = new MessageQueue();
IProducer producer = new Producer(queue.Enqueue);
IConsumer consumer = new Consumer(queue.TryDequeue);
That's simple enough, but not really very scalable, once you get a lot of dependencies to wire up. What we really need is an IoC container that let's us register delegates against classes, instead of having to have instances at dependency declaration time. Delegates can't be cast from one to another and are not, strictly speaking, types, which posts some challenges with creating a type-safe registration interface. There are a number of ways to accomplish this syntax, which I will elaborate on in my next post.
Interfaces are the method by which we get multiple inheritance in C#, Java, etc. They are contracts without implementation. We don't get the messy resolution of which code to use from multiple base classes, because there's only one inheritance chain that includes code.
They're useful to let us provide one contract and multiple implementations or simply describe a contract for our code and allow someone else to come along and replace our implementation entirely.
In practice, though, I almost always use them purely for decoupling the contract from the code, so that I can replace the implementation with a mock version in unit tests. Hmm.. So, I use interfaces just to get around the yoke of the type system? Wait, why I am so in love with statically typed languages again?
Right... While the above sounds like the interface exists just so that I can mock my implementation, the real purpose of the interface is the ability for the consuming code to express a contract for its dependencies. It's unfortunate that interface implementation forces them to be attached at the implementation side, which is why I say that Interface attachment is upside down. And it's deep rooted in the language, after all it's called Interface Inheritance not Dependency Contract Definition.
So, it's not surprising that there is no CLR-level way around this limitation. Fortunately, you can always create a dynamic proxy that wraps your class and implements the Interface. Both Castle's DynamicProxy and LinFu's DynamicProxy are excellent frameworks for writing your own proxies. I've never tested them against each other, but have used both in production and neither showed up as culprits when time for profiling came about.
With a dynamic proxy, you can generate an object that claims to implement an interface but under the hood just has a interceptors that provide the call signature to let you respond correctly or proxy the call on to a class you are wrapping. I've previously covered how you can even use them to have a class inherit from more than one base class via a proxy. This is necessary if you want to remote an object, which requires a base of MarshalByRefObject, but you already have a base class.
However, proxies require a fair bit of hand-rolling so they are not the most lightweight way, development time wise, to attach an Interface.
What would be really useful would be the ability to cast an object to an interface:
IUser user = new User() as IUser;
The above code would even be compile time verifiable, since we can simply see if the User implements the call signatures promised by IUser. This would be provide us strongly typed Duck Typing -- an object that can quack ought to be able to be treated as a duck.
This is where LinFu goes a step further than just DynamicProxy and provides duck typing as well:
IUser user = DynamicObject(new User()).CreateDuck<IUser>();
DynamicObject's constructor takes an instance of a class to wrap. You can then create a duck from that dynamic object which automatically proxies the given interface and will call the appropriate method on the wrapped class on demand.
Saying that you may have a class that has the perfect method signature but doesn't implement an interface you already have, does sound rather contrived. However, forcing a class to implement an interface of your choosing does have some real benefits, aside from being able to abstract an existing class into a mockable dependency:
One problem with interfaces is that they tell you everything an implementation can do. And often a class acts as a service that provides functionality to more than one client class, but provides just a single interface. That single interface may expose capabilities that you don't care about.
Instead, interfaces should be fine-grained to only include the methods appropriate to the client. But that's not always feasible. Aside from having a class implement lots of tiny interfaces, the service class does not know about the client's requirements, so it really doesn't know what the interfaces should include. The client, on the other hand, does know and can tailor exactly the right interface it wants as a contract with the dependency.
Suppose we have message queue for passing data between decoupled classes:
public class MessageQueue
{
public void Enqueue(string recipient, string message) { ... }
public string TryDequeue(string recipient) { ... }
}
Proper interface segregation would have us create a Dispatcher interface for our message Producer
public interface IMessageDispatcher
{
void Enqueue(string recipient, string message);
}
public class Producer : IProducer
{
public Producer(IMessageDispatcher dispatcher) { ... }
}
and an inbox interface for our message Consumer
public interface IMessageBox
{
string TryDequeue(string recipient);
}
public class Consumer : IConsumer
{
public Consumer(IMessageBox inbox) { ... }
}
Assuming that MessageQueue does not implement our interfaces (yes, in this case it would not have been a problem to have the class implement them both, but this is a simplified example with obvious segregation lines), we can now configure our IoC container (example uses AutoFac) to create the appropriately configured IProducer and IConsumer, each receiving exactly those capabilities they should depend on:
var queue = new MessageQueue();
var builder = new ContainerBuilder();
builder.Register<Producer>().As<IProducer>();
builder.Register<Consumer>().As<IConsumer>();
builder.Register(c => new DynamicObject(queue).CreateDuck<IMessageDispatcher>()).As<IMessageDispatcher>();
builder.Register(c => new DynamicObject(queue).CreateDuck<IMessageBox>()).As<IMessageBox>();
using (var container = builder.Build())
{
var producer = container.Resolve<IProducer>();
var consumer = container.Resolve<IConsumer>();
}
While I think Dynamic objects in C# 4.0 are very cool, as of right now, they seem to have skipped over duck typing, at least in a strongly typed fashion.
Sure, once you have a dynamic instance, the compiler will let you call whatever signature you wish on it and defers checking until execution time. But that means we have no contract on it, if used as a dependency, nor can we use it to dynamically create objects that provide implementations for existing contracts. So, you've have to wrap a dynamic object with a proxy, in which case, LinFu's existing duck typing already provides a superior solution.
The lack of casting to an interface, imho was already oversight with C# 3.0, which introduced anonymous classes that are so convenient for Linq projections, but can't be passed out of the scope of the current method, due to a lack of type.
So don't expect C# 4.0 to do anything to let you more easily attach your contracts at the dependency level. For the foreseeable future, this remains the territory of Dynamic Proxy.
However, there is another way to deal with dependency injection that provides a fine-grained contract and imposes no proxies nor other requirement on the classes providing the dependency: Injection of the required capabilities as delegates
I've been experimenting with a framework to make this as painless as traditional IoC containers make dependency resolution. It's still a bit rough around the edges, but I hope to have some examples to write about soon.
Over the years I've hopped back and forth between static and dynamically typed languages, trying to find the sweet spot. I currently still favor managed, static languages like C# and Java. But I agree that sometimes I have to write a whole lot of code to express my intent to the compiler without any great benefit. And no, i don't think that code-generation is a way out of this.
I won't go over the usual arguments for dynamic, which basically boil down to "you can do what you want without having to explain it to the type system first". I'll stipulate that that is why most people choose dynamic, but it's not a significant a pain point with static for me. I did spend a good many years in dynamic land and switched to static of my own free will. Instead, I want to concentrate on some specific cases.
I generally don't care whether i am dealing with int or long, or int64, or double vs. decimal. For the most part, number would do just fine. I think these types can be useful optimizations for both speed and memory, but certainly something that would be better optimized by a tracing rather than declaratively at compile time. And having to call special converters all over the place to go between these various types, is just not useful. I think type inference can handle these scenarios just fine.
I'm not saying that there aren't areas where speed isn't important enough to drop down to C/C++ levels, but anywhere where you are willing to use a statically typed managed language, a dynamic language can either perform right now or is only a short time away from being performant. After all, i already sacrifice performance to be in managed land, so a little more sacrifice for the development benefits seems arbitrary. Besides, recent javascript optimizations paint a pretty good picture that tracing and JIT compilation can make dynamic code fast enough for most scenarios.
So what's the pain point of dynamic for me? I care neither about the locking down of a class to handle only statically defined things, nor about the guarantee that a type is really a particular type. Frankly "types" are not important to me. However, declaration of dependencies in a discoverable fashion is!
What do I mean by that? A class should tell me via a machine discoverable contract what it expects the passed instance to be capable of. If I use a class that has service dependencies at construction time, or instance dependencies at method invocation time, I want to be able to discover this in code, rather than by looking at documentation or going by naming convention. After all, hasn't documentation been deemed a code smell? Why is it then, that in dynamic languages the expected capabilities of the object to be passed is not expressed in a fashion that can be discovered without breaking encapsulation and looking at what the code expects to do with the passed instance?
Sure, dynamic languages pride themselves on not requiring an IDE. This is often held up as a strength and a key reason why they are faster to develop in. In my experience, however, I find dynamic languages faster for small things but as the project grows, my velocity decreases:
I have to memorize more and more code and refer back to docs more, instead of letting the IDE guide me on available signatures
Instead of navigating by concrete signatures, i do lots of string searches
Instead of using long, self-documenting class and method names, I use terse syntax to ease typing and recollection
Instead of symbolic refactoring, I do regex replaces, hoping there isn't a syntax collision between classes that replaces the wrong thing
And handing off code or integrating someone else's code is slower because domain knowledge moves out of the code into tribal knowledge and documentation
All the above is not a problem in static languages, but at the cost of inflexible, rigid types. Types are a solution that are a trojan horse of limitations that are completely orthogonal to the problem of dependency discovery. A class requiring an object of type User should have no dependence on the implementation details of User. Having such a dependence would be a clear violation of encapsulation. The class should simply want an instance of an object that has the capabilities of a User, i.e. it has a requirement for an object that exposes certain capabilities. The class should be able to declare a contract for the instance to be passed in.
In C# and similar languages this contract is an Interface. Interfaces allow the declaration of capabilities without an implementation. In interface inheritance, a class commits to providing the contract expressed by the interface. So a class requiring an a specific interface can declare its requirements without any knowledge of implementation. All right, problem solved! Right?
Interfaces unfortunately do not solve the problem, because the way the are attached to implementation inverts the dependence hierarchy. I.e. User implements an interface its author declared, called IUser. Now IUser becomes my dependency, which is still a declaration outside of my control. I should not care where the implementation comes from. But an interface, puts the burden on a third party to implement my interface, which means I cannot use anything pre-existing, since it wouldn't have implemented my interface, or the burden is put on the third party to provide an interface tome to use, which means another third party solving the same problem, provides their own interface.
This may be wonderful for mocking and unit testing, but it still ties me to a contract not of my own making and usually violates the Interface Segregation Principle: Clients should not be forced to depend upon interfaces that they do not use. So interfaces provide a solution, but they still enforce rigidity that has no benefit to the definition of dependency contracts.
At the end of the day, I have less to quarrel about dynamic vs. static, than tribal definition (naming conventions, documentation, etc.) of dependencies vs. declarative definition of dependencies. Until I can discover what a class expects as its input without being told or cracking open the man page, I will suffer the yoke of interfaces. Especially since I can still use Dynamic Proxies to fake a class implementing an interface -- in yet another "more code than you'd think" way of working, tho.
Are there any static or dynamic languages that have tackled declarative contracts that are not attached at the implementation side that I'm not aware of? It seems like a sweet spot that isn't yet addressed.
Update: I realize i left those wanting a solution to the interface issue in C# wanting. There are two ways to solve the problem that I'm aware of, Duck Typing as offered by LinFu and delegate injection, both of which I will cover in future posts.
The using statement/block in C# (not the one used to pull in namespaces) is meant to aid in the IDisposable pattern, i.e. cleaning up resources that won't be handled by garbage collection and to do so in a deterministic fashion. I.e. everything that finalization is not. It really is just syntactic sugar to avoid try/finally all the time. I.e. this
using(var disposable = new DisposableObject())
{
// do something with disposable
}
is pretty much the same as
var disposable = new Disposable();
try
{
// do something with disposable
}
finally
{
disposable.Dispose();
}
But there is a common pitfall with using, err,using. It's in the first line above: The disposable object is created outside the try/finally block! Now, a constructor failing shouldn't ever have allocated disposable resources, so you're usually safe here. But beware if the construction of the Disposable object is a Method.
using(var disposable = CreateDisposable())
{
// do something with disposable
}
If CreateDisposable() fails after it has created Disposable, you'll end up with a resource leak!
You can easily avoid this by catching failure in your method and cleaning up, but you can't use using for this purpose, since success would return an already disposed instance. A safe implementation of CreateDisposable() looks like this:
public Disposable CreateDisposable()
{
Disposable disposable = new Disposable();
try {
// do some extra initialization of disposable
}
catch
{
if( disposable != null )
disposable.Dispose();
throw;
}
return disposable;
}
IDisposable is an important pattern in .NET, but because it is a pattern rather than a construct enforced by the compiler, it is a common source of "leaks" in .NET. using is very useful for handling the disposable, but it is important to remember that the only code covered by the automatic disposition logic is the code inside the using block, not the code inside the using statement.
A while back, I wrote about searching through a tree using linq to objects. That post was mostly snippets of code about delegates, lambda's, yield and how it applies to linq -- more a technical exploration than an example. So I thought I'd follow it up with concrete extension methods to make virtually any tree searchable by Linq.
All that is required to search a tree with Linq is creating a list of all nodes in the tree. Linq to Objects can operate on IEnumerable. Really, Linq to objects is a way of expressing operations we've been doing forever in loops with if/else blocks. That means there isn't any search magic going on, it is a linear traversal of all elements in a set and examining each to determine whether it matches our search criteria.
To turn a tree into a list of node we need to walk and collect all children of every node. A simple task for a recursive list that carries along a list object to stuff every found node into. But there is a better way, using yield to return each item as it is encountered. Now we don't have to carry along a collection. Iterators using yield implement a pattern in which a method can return more than once. For this reason, a method using yield in C# must return an IEnumerable, so that the caller gets a handle to an object it can traverse the result of the multiple return values.
IEnumerable is basically an unbounded set. This is also the reason why unlike collections, it does not have a Count Property. It is entirely possible for an enumerator to return an infinite series of items.
Together IEnumerable and yield are a perfect match for our problem, i.e. recursively walking a tree of nodes and return an unknown number of nodes.
In depth-first traversal, the algorithm will dig continue to dig down a nodes children until it reaches a leaf node (a node without children), before considering the next child of the current parent node.
In breadth-first traversal, the algorithm will return all nodes at a particular depth first before considering the children at the next level. I.e. First return all the nodes from level 1, then all nodes from level 2, etc.
public static class TreeToEnumerableEx
{
public static IEnumerable<T> AsDepthFirstEnumerable<T>(this T head, Func<T, IEnumerable<T>> childrenFunc)
{
yield return head;
foreach (var node in childrenFunc(head))
{
foreach (var child in AsDepthFirstEnumerable(node, childrenFunc))
{
yield return child;
}
}
}
public static IEnumerable<T> AsBreadthFirstEnumerable<T>(this T head, Func<T, IEnumerable<T>> childrenFunc)
{
yield return head;
var last = head;
foreach(var node in AsBreadthFirstEnumerable(head,childrenFunc))
{
foreach(var child in childrenFunc(node))
{
yield return child;
last = child;
}
if(last.Equals(node)) yield break;
}
}
}
This static class provides two extension methods that can be used on any object, as long as it's possible to express a function that returns all children of that object, i.e. the object is a node in some type of tree and has a method or property for accessing a list of its children.
Let's use a hypothetical Tree model defined by this Node class:
public class Node
{
private readonly List<Node> children = new List<Node>();
public Node(int id)
{
Id = id;
}
public IEnumerable<Node> Children { get { return children; } }
public Node AddChild(int id)
{
var child = new Node(id);
children.Add(child);
return child;
}
public int Id { get; private set; }
}
Each node simply contains a list of children and has an Id, so that we know what node we're looking at. The AddChild() method is a convenience method so we don't expose the child collection and no node can ever be added as a child twice.
The calling convention for a depth-first collection is:
The lambda expression n => n.Children is the function that will return the children of a node. It simply states given n, return the value of the Children property of n. A simple test to verify that our extension works and to show us using the extension in linq looks like this:
[Test]
public void DepthFirst()
{
// build the tree in depth-first order
int id = 1;
var depthFirst = new Node(id);
var df2 = depthFirst.AddChild(++id);
var df3 = df2.AddChild(++id);
var df4 = df2.AddChild(++id);
var df5 = depthFirst.AddChild(++id);
var df6 = df5.AddChild(++id);
var df7 = df5.AddChild(++id);
// find all nodes in depth-first order and select just the Id of each node
var IDs = from node in depthFirst.AsDepthFirstEnumerable(x => x.Children)
select node.Id;
// confirm that this list of IDs is in depth-first order
Assert.AreEqual(new int[] { 1, 2, 3, 4, 5, 6, 7 }, IDs.ToArray());
}
For breadth-first collections, the calling convention is:
Again, we can test that the extension works like this:
[Test]
public void BreadthFirst()
{
// build the tree in breadth-first order
var id = 1;
var breadthFirst = new Node(id);
var bf2 = breadthFirst.AddChild(++id);
var bf3 = breadthFirst.AddChild(++id);
var bf4 = bf2.AddChild(++id);
var bf5 = bf2.AddChild(++id);
var bf6 = bf3.AddChild(++id);
var bf7 = bf3.AddChild(++id);
// find all nodes in breadth-first order and select just the Id of each node
var IDs = from node in breadthFirst.AsBreadthFirstEnumerable(x => x.Children)
select node.Id;
// confirm that this list of IDs is in depth-first order
Assert.AreEqual(new int[] { 1, 2, 3, 4, 5, 6, 7 }, IDs.ToArray());
}
The tree used in the example is of course extremely simple, i.e. it doesn't even have any worthwhile data to query attached to a node. But these extension methods could be used on a node of any kind of tree, allowing the full power of Linq, grouping, aggregation, sorting, projection, etc. to be used on the tree.
As a final note, you may wonder, why bother with depth-first vs. breadth first? After all, in the end we do examine every node! There is however one particular case where the choice of algorithm can be very important: You are looking for one match or a particular number of matches. Since we are using yield, we can terminate the traversal at any time. Using the FirstOrDefault() extension on our Linq expression, the traversal would stop as soon as one match is found. And if have any knowledge where that node might be in the tree, the choice of search algorithm can be a significant performance factor.
Since i keep overwriting my App.Config with revision control configs and promptly forgetting how to set up filters, I figured i might was well write a brief article on filters here, so i have a place to look it up next time :)
My basic tenet with logging is that lots of debug statements is good. Now some may say that it just gets too noisy after a while, so don't put them in unless you need to debug. The problem is that you usually don't know when you'll need to debug, and if the code is deployed on a server or worse with a customer, generating a new release so they can run the debug is a burden you shouldn't have to shoulder. And commenting out log statements (or even conditional compiles) are a code smell reeking of Console.Writeline debugging
But it does get noisy! And noisy also means slow! However with log4net, noisy and performance degradation are non-arguments, since aside from levels, it has excellent filtering, which not only reduces the noise, but also cuts out 99% of the logging overhead. Worst case debugging example I've had was tracking down behavior in the motion control software for Full Motion Racing. The physics calculations in this software ran between 60Hz and 100Hz. When i added debug logging in that physics loop, the rate dropped down to about 20Hz because of I/O overhead, and this was with either RollingFile or Udp appenders. Needless to say, motion became jerky and unusable for a rider. But I got the debug data i needed. Disabling those logging statements with filters rather than removing left no appreciable degradation in the performance of the physics loop.
So, again, lots of debug logging == good. Because when you need that data, you need that data. But you may want to ship your code with a log4net configuration that pre-filters the loggers you know to be noisy, so that a user turning on debug logging doesn't overwhelm them.
The basic deal with log4net filters is that they are applied in order and the first filter that matches short-circuits the matching logger. I.e. if the first filter is a DenyAllFilter, nothing else will even be considered, since it matches all loggers. That means there are generally two approaches to filtering, whitelisting and blacklisting. It also means that if you match a logger and a subsequent filter would remove that logger, the subsequent filter is never reached, since consideration of the filter chain stops at the first match
LoggerMatchFilter filters default to acceptOnMatch being true, i.e. if omitted, the filter is a accepts (includes in logging) on match. The above will only emit logging statements for the Only.Logger.To.Match logger, since all others will hit the DenyAllFilter and be excluded.
This filter will show all logging statements except those for Logger.To.Filter.Out.
LoggerMatchFilter also matches on partial namespaces, which is very useful when you have a noisy namespace, but one logger in that namespace that you do want in your logs such as:
So, don't get locked into thinking that your choice for verbosity lies only in log levels and once committed to a level it's either all the noise or none of it. But to keep things sane, pre-populate your config with filters, because you are the one that knows best which loggers are of general use and which are special case only.
Once again, there's been extended silence over here. I have several article drafts that keep getting the short end of my time in favor of coding. In the meantime, I have blogged a couple of article's over on the MindTouch Dev blog.
There's been a lot of chatter of late about SOLID. It started with Uncle Bob talking on a couple of podcasts about the SOLID principles, but it really got the chatter going when Joel Spolsky and Jeff Atwood started talking smack about Uncle Bob on the Stackoverflow podcast. Since then battlelines seem to have been drawn between the TDD/SOLID folks and those who finally found a champion fighting to get them out from under the pattern yoke. Ok, that's a bunch of hyperbole, but it seems that the main objection to SOLID seems to be that having a list of Principles like that just feels bureaucratic and dogmatic, which rubs free thinking developers the wrong way. Since I've been practicing SOLID longer than I've been aware of, I wanted to walk through how I got here and to illustrate that the principles espoused by Uncle Bob are not a yoke, but rather helpful guidelines that will save you a lot of grief down the line.
The overall goal of development should be delivering software that solves the stated problem. Beyond that I do have some personal guiding principles for programming, i.e. what I personally want to get out of it once the raison d'__être is accomplished:
I want to learn new things rather than maintain old things
I never want to have to fix the same bug twice
I don't want to be prevented from doing something better by legacy decisions
To me, 1) means writing code that's easy to maintain, so that maintenance does not become a time suck interfering with new things, 2) means that i protect myself from regressions and 3) means that my code should allow me to refactor it without screwing up 1) or 2). So far that seems pretty non-controversial.
As I go on, the common theme will be testing, not because testing is some higher goal and end in itself, but because testing, in my experience, let's me prove that code does what it claims and I didn't break anything else by the addition/change. For those who think that writing tests is a lot of tedium that only leads to test maintenance rather than code maintenance, I can only respond "you're probably doing it wrong" and address that statement under Pain points of TDD below.
In the early days of the web, testing was what the programmer did to make sure things didn't break and then you relied on the customer telling you if it didn't work correctly -- the wild west days of CGI scripts. Once I started at MP3.com, testing became more refined via QA. Sure, everybody tested their web apps before handing them off to QA (or at least they should), but there wasn't really any formalized testing on the development side. It was all manual functional testing of firing up the app, trying out the things that should work and looking at logs. QA was responsible for test plans, regression tests, etc.
Early on I switched from web apps to running the databases and with that came dealing with the pain of maintaining schemas when everyone had raw SQL in strings throughout their code. So I set out to write a DB API. Being an OO geek, it quickly morphed from an API into an ORM instead, which abstracted the DB and built the SQL on demand. This gave the DB group more freedom to refactor the database as needed without having to have every developer track down their SQL.
Developing the ORM did mean that I was now out of the QA loop, since my deliverables went to developers and had to work long before QA ever got involved. So I developed test suites that I could run from the shell whenever i changed something. These tests gave me confidence that I didn't just break live apps with code for a new app.
Part of having these type of tests, however, was a giant WTF in itself. Why was I constantly risking the codebase by futzing around in the guts? Yes, the ORM suffered horribly from fragile baseclass problems and I had designed myself into a number of corners that could only be addressed by modifying the base. Learning this lesson, I spent a lot more time trying to build object hierarchies to provide the proper hooks to let subclasses extend the functionality without affecting or overriding the base functionality. Little did I know that I had started practicing SOLID's O, or the Open/Closed Principle (OCP).
When I started at Printable Technologies a number years later, I became part of an effort to migrate the existing application from ASP to ASP.NET. Like most legacy ASP applications it was the usual single code file per page mixing data access, business logic and html rendering. It was something we did not want to repeat. We set out to separate our logical layers carefully so that we would get greater re-use and transparency of what was going on in the application. I wanted to start off on the right foot and played around with NUnit to try out unit testing our new code base. This was before TestDriven.Net or similar tools for integration into the IDE. But at least it gave me an automated test suite, rather than a series of console apps. I thought, "Now I'm doing unit testing!".
Except, like many test adoptees, I really wasn't. I was doing functional tests with a test running harness. I was hitting test databases to check my DB Abstraction Layer, and as you moved up the object hierarchy, the graphs of dependent code supporting the code to be tested got deeper and deeper. Testing something that was at the front end, really tested all pieces beneath it. It certainly gave us good test coverage, but the tests were fragile, took a lot time to write, and it was often difficult to dig out what actually failed. That didn't seem right and made me wonder if this unit testing thing was really so great. But the test coverage did help to achieve my personal guiding principles, so I wasn't ready to give up on it just yet. I just need to work through the pain points.
The two major problems we had in our tests were that each test really tested many things at once and that the setup to get to test running was tedious and pulled in too many dependencies.
The first was a problem with our class design. We needed to break classes up into smaller functional pieces that each could be tested before testing the whole. While testability was the driving force for this change, it requiring this change was really a symptom of how badly coupled things were, i.e. that the design was flawed. This is a pattern in testing that has since repeated itself many times: If your test is fragile or difficult, it's generally the fault of the design of the code to be tested not the testing process.
There were a bunch of monolithic classes that did lots of things at once, which meant a test failure could be one of a hundred things. I started to break up classes into smaller pieces, each dedicated to one functional area. Now I could take those helper classes that the main class was composed of and test them independently. When the composite broke but none of the components did, I knew where to look for failure. This compartmentalization just happens to be Single Responsibility Principle (SRP).
So far so good, but our second problem still dogged us. Tests were still annoying to write the further you got into business logic, since everything built on the supporting infrastructure. I had heard about mocking and started looking into it hoping for some magic bullet that could just create me fakes (which had been easy back in the perl days). This was before TypeMock hit the scene, so I couldn't create a fake version of my concrete type. It was either making everything virtual (yuck) or using interfaces instead of concrete classes. Interfaces won out, but because I had yet to discover the D in SOLID, introducing a lot of interfaces also led us to a pattern that itself became a major pain point. This pattern was the use of singletons and the static factory methods, both well meaning static accessors to get around the inability of new'ing up an interface. But before realizing this separate morrass, using interfaces had lead to using the L, Liskov Substitution and I, Interface Segregation principles.
Mocking out classes with interfaces exposed a couple of places where we had an abstract baseclass and code accepting the base class using typeof() to determine what class was actually provided. Well, with an interface being passed in instead of the abstract baseclass, the typeof() logic still worked, until the first test with a mock object was run. That failure illustrated what a bad idea that bit of code was. If we say we require an object implementing an interface, any object implementing that interface should work, and that right there is the Liskov Substitution Principle(LSP). Making sure that our interfaces really represented the required functionality enabled mocking and cleaned out some inappropriate knowledge embedded in code.
Another aspect of mocking (rolling mocks by hand rather than using a mocking framework) was that lazyness dictated that you didn't want to implement a lot of things just to get a test working. So large interfaces got widdled down to just the methods required by the object taking in that interface. And that happens to be the Interface Segregation Principle (ISP).
Many of the above principles were only partially applied because they imposed a new pain and a whole new set of plumbing code that was tedious to write and maintain. The issue with separation of concerns and abstracting those concerns with interfaces was two-fold:
You can't new up an interface, so you needed factories everywhere
Suddenly half the code seemed to be plumbing to wire up increasingly complex object graphs.
As I said, a lot of this was dealt with via Singleton's and static factory methods. Both are really just degenerate implementations of the Service Locator Pattern, but we weren't even aware of that. Since this plumbing had it's own set of pain points, we often skipped the abstractions unless we really needed them to keep life simpler. Generally that meant we paid for that convenience in maintenance debt.
When I started writing code for Full Motion Racing, like every new project, I wanted to take the lessons learned from Printable and avoid the pains I had come across. I once again had need for object graphs that required access to singleton type service objects, but wanted to avoid statics as much as I could because of previous experience. I built a repository of objects that I could stuff instances into, providing me a Service Locator. Looking more into how other people were doing this, I came across talk about service locator still being an inappropriate coupling, since it itself is a dependency that had nothing to do with the responsibility of the consuming objects. Instead, services should be passed in at construction time whenever possible. Wow, really? That just seemed to take the pain of wiring up object graphs to unprecedented heights. That just couldn't be how people were writing their code.
Wanting to understand how this way of building decoupled systems could actually work in the real world, I learned about the D of SOLID, or the Dependency Inversion Principle(DIP) also (and maybe more accurately) known as Inversion of Control. In my opinion, DIP may be the last principle mentioned, but in many ways it is the enabling plumbing without which the remaining principles are all well in theory but often feel worse than the disease they aim to cure.
Early use of IoC for Full Motion Racing still relied on a singleton container that factory classes could use to create their dependencies for creating transient objects. Only over time did I learn to trust the container to build up all my objects for me, and learned how to register factories to support lifestyles other than singleton via the container.
It wasn't until I started at Bunkspeed, that I really saw IoC used properly and was able to reap the true benefits of this design pattern. If you've ever seen Bunkspeed's HyperShot or HyperDrive in action you know that the visualizations they create are mind blowing, especially once you realize it's real-time. Needless to say, sitting down with this codebase was initially intimidating. It still is the largest single codebase i've worked on. Maybe some of the distributed web apps I've worked on had more code in total, but they were disparate systems that largely had no interdependence. The main Visual Studio Solution I worked on at Bunkspeed was one application with hundreds of projects all loaded at once.
I assumed this meant lots of branching, lots of areas of expertise where certain people would be responsible for a subset of the code. This was not the case. Everyone was trusted and had authority to modify, extend and refactor everything as they required it. Making a change that required a change much lower down in the system could be made by anyone. Tests ran with every build in addition to a full set of CI servers building various configurations on each check-in. And it all ran smoother than any other shop I'd been in and was easier to ramp up on then other, far simpler projects.
Bunkspeed employs every one of the SOLID principles. Systems were composed of lots of small classes with very limited responsiblity, each being abstracted by an interface. One of the reasons for the many projects rather than fewer, larger projects was that areas of responsiblity were segregated, including their interfaces being in separate DLLs so that low level changes wouldn't cause rebuilds of the entire system. Deep reaching refactors were not the norm but rather an indication that some inappropriate coupling had been done at some previous time and the refactor served to rectify the situation so that technical debt was accruing at much slower rates than is the norm. This was not some academic application of patterns from a book, but a truly agile development shop able to make significant changes with a small team in record time.
SOLID wasn't some rule set put before me, but a natural evolution of trying to make development easier. It wasn't until about 9 months ago that I read Robert C. Martin and Micah Martin's Agile Principles, Patterns, and Practices in C# because it was sitting on a co-worker's desk and for the first time I put names to the patterns I'd been applying this entire time.
There is definitely a different cost occured by applying SOLID to design. Most of this cost is in navigating the granularity of the design and in this tooling is an important aid to make this not only painless but more productive than the alternatives. The issue is that Visual Studio really isn't all that well suited to navigating large object hierarchies, especially when using interfaces for abstraction. There are those who will point to this pain as evidence of SOLID being a bad practice. "I don't want to have to get special tools just to do development." But if tooling really is your enemy then you probably shouldn't be working with a language like C# in the first place, because it already does rely on the many crutches VS offers up. Try writing C# without an IDE and you'll quickly understand why people love the simple and terse syntax of Ruby and other dynamic Languages. Saying "well, Visual Studio is as much tooling as I accept, beyond that it's ridiculous" is not an argument I can relate to, so if that's the objection to SOLID, I'll have to admit that I can't convince you.
The issue just is that VS does not provide efficient ways to navigate from class to class and from interface to implementers and from implementers to usage of the interface. This is where ReSharper entered the picture for me, and after adding more keybindings for some of their extended commands, the number of classes and abstractions really becomes a non-issue. I simply couldn't do development in C# without ReSharper at this point and that's more because of Visual Studio's shortcomings than anything else (eclipse, for example, provides most of the same features I rely on out of the box).
The other pain created by lack of tooling is that the larger surface area of code and the increased abstraction means that refactoring ususally touch more code than before, meaning that refactoring takes more work. However with proper tooling this also happens automatically. In addition, the flexibility of loose coupling introduced by SOLID generally makes code more pliable to refactoring.
Finally there is the whole concept of TDD itself which to many seems like such a "making work" paradigm. Usually the examples of TDD failure pointed to are fragile tests and lost productivity due to writing tests instead of code.
Fragile tests refers to a simple change breaking a lot of tests and thereby incurring extra work instead of saving work. But if a simple change breaks a lot of tests, it's a clear indication that your design needs another look, because there seems to be coupling that is getting in the way. The only time (ideally at least) that more than one test breaks is when you change the expected behavior of a class, and in that case, refactoring the expected behavior would have included the refactoring of all dependencies, including the tests.
The lost productivity argument only holds water if you are not responsible for extension or maintenance. And all you are doing then is pushing the work that should have been yours in the first place on the poor sucker that's inheriting your legacy. It has been my experience that any time I find a bug or break something with a new feature, it's because I didn't have test coverage on the affected code. Which means I get to lose productivity when I am likely already under time pressue rather than up front.
Another part of the lost productivity argument usually refers to the amount of code required to test a particular object vs. just using that object in production code. Between DIP for wiring things up and the numerous mocking frameworks available to declare your expectations on dependencies, wiring up your test harness should be short. If there is a lot of plumbing required just to set up the test conditions something's wrong with the design.
Since I arrived at practicing the SOLID principles, without being aware of them, I just can't see the dogma or beaurecracy in them. The recommendation in those 5 principles when taken as a whole is about making your life easier not forcing some philosophy down your throat.
If someone had put them front of me and told me "this is the way you must code, because good programmers do this", I'd likely be dismissive as well. Being thrown into implementing some process from a whitepaper without having seen it in practice and understood why it was useful, has a high likelyhood of leading to improper implementation which makes the presumed failure a self-fullfilling prophecy. However if taken as guidance, there is a lot of useful information there that can make projects, especially large projects, a lot less painful.