August 2010 Blog Posts
Our eye in the sky reports two angry evil (but devishly good looking) cyborg units, XSP 2000 and TRS-80, are fast approaching Black Rock City. They are considered very armed and dangerous. In fact, they are mostly armed and not much else.
These cyborgs do not come in peace. I repeat, they are to be considerd hostiles. However, we’ve received a secret communiqué that reveals a weakness built into these cyborg models. Due to a lack of TDD during development, a bug in their FOF system (friend or foe) causes them to view anyone offering a frosty beverage to be a friend, not foe.
Any attempts to engage with these hostiles will result in calamity unless you offer them an ice cold beverage. For the sake of your beloved city, I suggest stocking up.
Intelligence confirms they are headed towards their evil cyborg camp at 8:15 and Kyoto on the Playa and are predicted to arrive on Tuesday morning. If we band together, we may be able to save our fair city by, once again, offering frosty alcoholic beverages in order to confuse their FOF system.
You’ve been duly warned.
This blog (and associated Twitter account) will go quiet for at least a week as communication systems are nonexistent within the Black Rock City area.
On Twitter yesterday I made the following comment:
We're not here to write software, we're here to ship products and deliver value. Writing code is just a fulfilling means to that end :)
All I see now is blonde, brunette, redhead.
For the most part, I received a lot of tweets in agreement, but there were a few who disagreed with me:
While I agree in principle, the stated sentiment "justifies" the pervasive lack of quality in development
Doctors with this mentality don't investigate root causes, because patients don't define that as valuable
That's BS. If you live only, or even primarily, for end results you're probably zombie. We're here to write code AND deliver value.
I have no problem with people disagreeing with me. Eventually they’ll learn I’m always right. ;) In this particular case, I think an important piece of context was missing.
What’s that you say? Context missing in a 140 character limited tweet? That could never happen, right? Sure, you keep telling yourself that while I pop a beer over here with Santa Claus.
The tweet was a rephrasing of something I told a Program Manager candidate during a phone interview. It just so happens that the role of a program manager at Microsoft is not focused on writing code like developers. But that wasn’t the point I was making. I’ve been a developer in the past (and I still play at being a developer in my own time) and I still think this applies.
What I really meant to say was that we’re not paid to write code. I absolutely love writing code, but in general, it’s not what I’m paid to do and I don’t believe it ever was what I was paid to do even when I was a consultant.
For example, suppose a customer calls me up and says,
“Hey man, I need software that allows me to write my next book. I want to be able to print the book and save the book to disk. Can you do that for me?”
I’m not going to be half way through writing my first unit test in Visual Studio by the end of that phone call. Hell no! I’ll step away from the IDE and hop over to Best Buy to purchase a copy of Microsoft Word. I’ll then promptly sell it to the customer with a nice markup for my troubles and go and sip Pina Coladas on the beach the rest of the day. Because that’s what I do. I sip on Pina Coladas.
At the end of the day, I get paid to provide products to my customers that meet their needs and provides them real value, whether by writing code from scratch or finding something else that already does what they need.
Yeah, that’s a bit of cheeky example so let’s look at another one. Suppose a customer really needs a custom software product. I could write the cleanest most well crafted code the world has ever seen (what a guy like me might produce during a prototype session on an off night), but if it doesn’t ship, I don’t get paid. Customer doesn’t care how much time I spent writing that code. They’re not going to pay me, until I deliver.
Justifying lack of quality
Now, I don’t think, as one Twitterer suggested, that this “justifies a pervasive lack of quality in development” by any means.
Quality in development is important, but it has to be scaled appropriately. Hear that? That’s the sound of a bunch of eggs lofted at my house in angry disagreement. But hear me out before chucking.
A lot of people will suggest that all software should be written with the utmost of quality. But the reality is that we all scale the quality of our code to the needs of the product. If that weren’t true, we’d all use Cleanroom Software Engineering processes like those employed by the Space Shuttle developers.
So why don’t we use these same processes? Because there are factors more important than quality in building a product. While even the Space Shuttle coders have to deal with changing requirements from time to time, in general, the laws of physics don’t change much over time last I checked. And certainly, their requirements don’t undergo the level of churn that developers trying to satisfy business needs under a rapidly changing business climate would face. Hence the rise of agile methodologies which recognize the need to embrace change.
Writing software that meets changing business needs and provides value is more important than writing zero defect code. While this might seem I’m giving quality a short shrift, another way to look at it is that I’m taking a higher view of what defines quality in the first place. Quality isn’t just the defect count of the code. It’s also how well the code meets the business needs that defines the “quality” of an overall product.
The debunking of the Betamax is better than VHS myth is a great example of this idea. While Betamax might have been technically superior to VHS in some ways, when you looked at the “whole product”, it didn’t satisfy customer needs as well as VHS did.
Nate Kohari had an interesting insight on how important delivering value is when he writes about the lessons learned building Agile Zen, a product I think is of wonderful quality.
It also completely changed the way that I look at software. I’ve tried to express this to others since, but I think you just have to experience it firsthand in order to really understand. It’s a unique experience to build a product of your own, from scratch, with no paycheck or deferred responsibility or venture capital to save you — you either create real value for your customers, or you fail. And I don’t like to fail.
Update: Dare Obasanjo wrote a timely blog that dovetails nicely with the point I’m making. He writes that Google Wave and REST vs SOAP provide a cautionary tale for those who focus too much on solving hard technical problems and miss solving their customers actual problems. Sometimes, when we think we’re paid to code, we write way too much code. Sometimes, less code solves the actual problems we’re concerned with just fine.
Code is a part of the whole
The Betamax vs VHS point leads into another point I had in mind when I made the original statement. As narcissistic developers (c’mon admit it. You are all narcissists!), we tend to see the code as being the only thing that matters. But the truth is, it’s one part of the whole that makes a product.
There’s many other components that go into a product. A lot of time is spent identifying future business needs to look for areas where software can provide value. After all, no point in writing the code if nobody wants to use it or it doesn’t provide any value.
Not to mention, at Microsoft, we put a lot of effort into localization and globalization ensuring that the software is translated into multiple languages. On top of this, we have writers who produce documentation, legal teams who work on licenses, marketing teams who market the product, and the list goes on. A lot goes into a product beyond just the code. There are also a lot of factors outside the product that determines its success such as community ecosystem, availability of add-ons, etc.
I love to code
Now don’t go running to tell on me to my momma.
“Your son is talking trash about writing code!”
It’d break her heart and it’d be completely untrue. I love to code! There, I said it. In fact, I love it so much, I tried to marry it, but then got a much better offer from a very lovely woman. But I digress.
Yes, I love coding so much I often do it for free in my spare time.
I wasn’t trying to make a point that writing code isn’t important and doesn’t provide value. It absolutely does. In fact, I firmly believe that writing code is a huge part of providing that value or we wouldn’t be doing it in the first place. This importance is why we spend so much time and effort trying to elevate the craft and debating the finer points of how to write good software. It’s an essential ingredient to building great software products.
The mere point I was making is simply that while writing code is a huge factor in providing value, it’s not the part we get paid for. Customers pay to receive value. And they only get that value when the code is in their hands.
In my last blog post, I covered some challenges with versioning methods that differ only by optional parameters. If you haven’t read it, go read it. If I do say so myself, it’s kind of interesting. ;) In this post, I want to cover another very subtle versioning issue with using optional parameters.
At the very end of that last post, I made the following comment.
By the way, you can add overloads that have additional required parameters. So in this way, you are in the same boat as before.
However, this can lead to subtle bugs. Let’s walk through a scenario. Imagine that some class library has the following method in version 1.0.
public static void Foo(string s1, string s2, string s3 = "v1") {
Console.WriteLine("version 1");
}
And you have a client application which calls this method like so:
ClassName.Foo("one", "two");
That’s just fine right. You don’t need to supply a value for the argument s3 because its optional. Everything is hunky dory!
But now, the class library author decides to release version 2 of the library and adds the following overload.
public static void Foo(string s1, string s3 = "v2") {
Console.WriteLine("version 2");
}
public static void Foo(string s1, string s2, string s3 = "v1") {
Console.WriteLine("version 1");
}
Notice that they’ve added an overload that only has two parameters. It differs from the existing method by one required parameter, which is allowed.
As I mentioned before, you’re always allowed to add overloads and maintain binary compatibility. So if you don’t recompile your client application and upgrade the class library, you’ll still get the following output when you run the application.
version 1
But what happens when you recompile your client application against version 2 of the class library and run it again with no source code changes. The output becomes:
version 2
Wow, that’s pretty subtle.
It may not seem so bad in this contrived example, but lets contemplate a real world scenario. Let’s suppose there’s a very commonly used utility method in the .NET Framework that follows this pattern in .NET 4. And in the next version of the framework, a new overload is added with one less required parameter.
Suddenly, when you recompile your application, every call to the one method is now calling the new one.
Now, I’m not one to be alarmist. Realistically, this is probably very unlikely in the .NET Framework because of stringent backwards compatibility requirements. Very likely, if such a method overload was introduced, calling it would be backwards compatible with calling the original.
But the same discipline might not apply to every library that you depend on today. It’s not hard to imagine that such a subtle versioning issue might crop up in a commonly used 3rd party open source library and it would be very hard for you to even know it exists without testing your application very thoroughly.
The moral of the story is, you do write unit tests dontcha? Well dontcha?! If not, now’s a good time to start.
One nice new feature introduced in C# 4 is support for named and optional arguments. While these two features are often discussed together, they really are orthogonal concepts.
Let’s look at a quick example of these two concepts at work. Suppose we have a class with one method having the following signature.
// v1
public static void Redirect(string url, string protocol = "http");
This hypothetical library contains a single method that takes in two parameters, a required string url and an optional string protocol.
The following shows the six possible ways this method can be called.
HttpHelpers.Redirect("http://haacked.com/");
HttpHelpers.Redirect(url: "http://haacked.com/");
HttpHelpers.Redirect("http://haacked.com/", "https");
HttpHelpers.Redirect("http://haacked.com/", protocol: "https");
HttpHelpers.Redirect(url: "http://haacked.com/", protocol: "https");
HttpHelpers.Redirect(protocol: "https", url: "http://haacked.com/");
Notice that whether or not a parameter is optional, you can choose to refer to the parameter by name or not. In the last case, notice that the parameters are specified out of order. In this case, using named parameters is required.
The Next Version
One apparent benefit of using optional parameters is that you can reduce the number of overloads your API has. However, relying on optional parameters does have its quirks you need to be aware of when it comes to versioning.
Let’s suppose we’re ready to make version two of our awesome HttpHelpers library and we add an optional parameter to the existing method.
// v2
public static void Redirect(string url, string protocol = "http",
bool permanent = false);
What happens when we try to execute the client without recompiling the client application?
We get a the following exception message.
Unhandled Exception: System.MissingMethodException: Method not found:
'Void HttpLib.HttpHelpers.Redirect(System.String, System.String)'....
Whoops! By changing the method signature, we caused a runtime breaking change to occur. That’s not good.
Let’s try to avoid a runtime breaking change by adding an overload instead of changing the existing method.
// v2.1
public static void Redirect(string url, string protocol = "http");
public static void Redirect(string url, string protocol = "http",
bool permanent = false);
Now, when we run our client application, it works fine. It’s still calling the two parameter version of the method. Adding overloads is never a runtime breaking change.
But let’s suppose we’re now ready to update the client application and we attempt to recompile it. Uh oh!
The call is ambiguous between the following methods or properties:
'HttpLib.HttpHelpers.Redirect(string, string)' and
'HttpLib.HttpHelpers.Redirect(string, string, bool)'
While adding an overload is not a runtime breaking change, it can result in a compile time breaking change. Doh!
Talk about a catch-22! If we add an overload, we break in one way. If we instead add an argument to the existing method, we’re broken in another way.
Why Is This Happening?
When I first heard about optional parameter support, I falsely assumed it was implemented as a feature of the CLR which might allow dynamic dispatch to the method. This was perhaps very naive of me.
My co-worker Levi (no blog still) broke it down for me as follows. Keep in mind, he’s glossing over a lot of details, but at a high level, this is roughly what’s going on.
When optional parameters are in use, the C# compiler follows a simple algorithm to determine which overload of a method you actually meant to call. It considers as a candidate *every* overload of the method, then one by one it eliminates overloads that can’t possibly work for the particular parameters you’re passing in.
Consider these overloads:
public static void Blah(int i);
public static void Blah(int i, int j = 5);
public static void Blah(string i = "Hello");
Suppose you make the following method call: Blah(0).
The last candidate is eliminated since the parameter types are incorrect, which leaves us with the first two.
public static void Blah(int i);
public static void Blah(int i, int j = 5);
public static void Blah(string i = "Hello");
At this point, the compiler needs to perform a conflict resolution. The conflict resolution is very simple: if one of the candidates has the same number of parameters as the call site, it wins. Otherwise the compiler bombs with an error.
In the case of Blah(0), the first overload is chosen since the number of parameters is exactly one.
public static void Blah(int i); //WINNER!!!
public static void Blah(int i, int j = 5);
public static void Blah(string i = "Hello");
This allows you to take an existing method that doesn’t have optional parameters and add overloads that have optional parameters without breaking anybody (except in Visual Basic which has a slightly different algorithm).
But what happens if you need to version an API that already has optional parameters? Consider this example:
public static void Helper(int i = 2, int j = 3); // v1
public static void Helper(int i = 2, int j = 3, int k = 4); // added in v2
And say that the call site is Helper(j: 10). Both candidates still exist after the elimination process, but since neither candidate has exactly one argument, the compiler will not prefer one over another. This leads to the compilation error we saw earlier about the call being ambiguous.
Conclusion
The reason that optional parameters were introduced to C# 4 in the first place was to support COM interop. That’s it. And now, we’re learning about the full implications of this fact.
If you have a method with optional parameters, you can never add an overload with additional optional parameters out of fear of causing a compile-time breaking change. And you can never remove an existing overload, as this has always been a runtime breaking change.
You pretty much need to treat it like an interface. Your only recourse in this case is to write a new method with a new name.
So be aware of this if you plan to use optional arguments in your APIs.
UPDATE: By the way, you can add overloads that have additional required parameters. So in this way, you are in the same boat as before. However, this can lead to other subtle versioning issues as my follow-up post describes.
I’ve been working on a lovely little prototype recently but ran into a problem where my code receives a collection of attributes and needs to change them in some way and then pass the changed collection along to another method that consumes the collection.
I want to avoid changing the attributes directly, because when you use reflection to retrieve attributes, those attributes may be cached by the framework. So changing an attribute is not a safe operation as you may be changing the attribute for everyone else who tries to retrieve them.
What I really wanted to do is create a copy of all these attributes, and pass the collection of copied attributes along. But how do I do that?
CustomAttributeData
Brad Wilson and David Ebbo to the rescue! In a game of geek telephone, David told Brad a while back, who then recently told me, about a little class in the framework called CustomAttributeData.
This class takes advantage of a feature of the framework known as a Reflection-Only context. This allows you to examine an assembly without instantiating any of its types. This is useful, for example, if you need to examine an assembly compiled against a different version of the framework or a different platform.
Copying an Attribute
As you’ll find out, it’s also useful when you need to copy an attribute. This might raise the question in your head, “if you have an existing attribute instance, why can’t you just copy it?” The problem is that a given attribute might not have a default constructor. So then you’re left with the challenge of figuring out how to populate the parameters of a constructor from an existing instance of an attribute. Let’s look at a sample attribute.
[AttributeUsage(AttributeTargets.All, AllowMultiple = true)]
public class SomethingAttribute : Attribute {
public SomethingAttribute(string readOnlyProperty) {
ReadOnlyProperty = readOnlyProperty;
}
public string ReadOnlyProperty { get; private set; }
public string NamedProperty { get; set; }
public string NamedField;
}
And here’s an example of this attribute applied to a class a couple of times.
[Something("ROVal1", NamedProperty = "NVal1", NamedField = "Val1")]
[Something("ROVal2", NamedProperty = "NVal2", NamedField = "Val2")]
public class Character {
}
Given an instance of this attribute, I might be able to figure out how the constructor argument should be populated by assuming a convention of using the property with the same name as the argument. But what if the attribute had a constructor argument that had no corresponding property? Keep in mind, I want this to work with arbitrary attributes, not just ones that I wrote.
CustomAttributeData saves the day!
This is where CustomAttributeData comes into play. An instance of this class tells you everything you need to know about the attribute and how to construct it. It provides access to the constructor, the constructor parameters, and the named parameters used to declare the attribute.
Let’s look at a method that will create an attribute instance given an instance of CustomAttributeData.
public static Attribute CreateAttribute(this CustomAttributeData data)
{
var arguments = from arg in data.ConstructorArguments
select arg.Value;
var attribute = data.Constructor.Invoke(arguments.ToArray())
as Attribute;
foreach (var namedArgument in data.NamedArguments) {
var propertyInfo = namedArgument.MemberInfo as PropertyInfo;
if (propertyInfo != null) {
propertyInfo.SetValue(attribute, namedArgument.TypedValue.Value, null);
}
else {
var fieldInfo = namedArgument.MemberInfo as FieldInfo;
if (fieldInfo != null) {
fieldInfo.SetValue(attribute, namedArgument.TypedValue.Value);
}
}
}
return attribute;
}
The code sample demonstrates how we use the information within the CustomAttributeData instance to figure out how to create an instance of the attribute described by the data.
So how did we get the CustomAttributeData instance in the first place? That’s pretty easy, we called the CustomAttributeData.GetCustomAttributes() method. With these pieces in hand, it’s pretty straightforward now to copy the attributes on a type or member. Here’s a set of extension methods I wrote to do just that.
public static IEnumerable<Attribute> GetCustomAttributesCopy(this Type type) {
return CustomAttributeData.GetCustomAttributes(type).CreateAttributes();
}
public static IEnumerable<Attribute> GetCustomAttributesCopy(
this Assembly assembly) {
return CustomAttributeData.GetCustomAttributes(assembly).CreateAttributes();
}
public static IEnumerable<Attribute> GetCustomAttributesCopy(
this MemberInfo memberInfo) {
return CustomAttributeData.GetCustomAttributes(memberInfo).CreateAttributes();
}
public static IEnumerable<Attribute> CreateAttributes(
this IEnumerable<CustomAttributeData> attributesData) {
return from attributeData in attributesData
select attributeData.CreateAttribute();
}
And here’s a bit of code I wrote in a console application to demonstrate the usage.
foreach (var instance in typeof(Character).GetCustomAttributesCopy()) {
var somethingAttribute = instance as SomethingAttribute;
Console.WriteLine("ReadOnlyProperty: " + somethingAttribute.ReadOnlyProperty);
Console.WriteLine("NamedProperty: " + somethingAttribute.NamedProperty);
Console.WriteLine("NamedField: " + somethingAttribute.NamedField);
}
And there you have it, I can grab the attributes from a type and produce a copy of those attributes.
With this out of the way, I can hopefully continue with my original prototype which led me down this rabbit hole in the first place. It always seems to happen this way, where I start a blog post, only to start writing a blog post to support that blog post, and then a blog post to support that one. Much like a dream within a dream within a dream. ;)
In ASP.NET MVC 3 Preview 1, we introduced some syntactic sugar for creating and accessing view data using new dynamic properties.
Sugar, it’s not just for breakfast.
Within a controller action, the ViewModel property of Controller allows setting and accessing view data via property accessors that are resolved dynamically at runtime. From within a view, the View property provides the same thing (see the addendum at the bottom of this post for why these property names do not match).
Disclaimer
This blog post talks about ASP.NET MVC 3 Preview 1, which is a pre-release version. Specific technical details may change before the final release of MVC 3. This release is designed to elicit feedback on features with enough time to make meaningful changes before MVC 3 ships, so please comment on this blog post if you have comments.
Let’s take a look at the old way and the new way of doing this:
The old way
The following is some controller code that adds a string to the view data.
public ActionResult Index() {
ViewData["Message"] = "Some Message";
return View();
}
The following is code within a view that accesses the view data we supplied in the controller action.
<h1><%: ViewData["Message"] %></h1>
The new way
This time around, we use the ViewModel property which is typed as dynamic. We use it like we would any property.
public ActionResult Index() {
ViewModel.Message = "Some Message";
return View();
}
And we reference it in a Razor view. Note that this also works in a WebForms View too.
Note that View.Message is equivalent to View["Message"].
Going beyond properties
However, what might not be clear to everyone is that you can also store and call methods using the same approach. Just for fun, I wrote an example of doing this.
In the controller, I defined a lambda expression that takes in an index and two strings. It returns the first string if the index is even, and the second string if the index is odd. It’s very simple.
The next thing I do is assign that lambda to the Cycle property of ViewModel, which is created on the spot since ViewModel is dynamic.
public ActionResult Index() {
ViewModel.Message = "Welcome to ASP.NET MVC!";
Func<int, string, string, string> cycleMethod =
(index, even, odd) => index % 2 == 0 ? even : odd;
ViewModel.Cycle = cycleMethod;
return View();
}
Now, I can dynamically call that method from my view.
<table>
@for (var i = 0; i < 10; i++) {
<tr class="@View.Cycle(i, "even-css", "odd-css")">
<td>@i</td>
</tr>
}
</table>
As a fan of dynamic languages, I find this technique to be pretty slick. :)
The point of this blog post was to show that this is possible, but it raises the question, “why would anyone want to do this over writing a custom helper method?”
Very good question! Right now, it’s mostly a curiosity to me, but I can imagine cases where this might come in handy. However, if you re-use such view functionality or really need Intellisense, I’d highly recommend making it a helper method. I think this approach works well for rapid prototyping and maybe for one time use helper functions.
Perhaps you’ll find even better uses I didn’t think of at all.
Addendum: The Property name mismatch
Earlier in this post I mentioned the mismatch between property names, ViewModel vs View. I also talked about this in a video I recorded for MvcConf on MVC 3 Preview 1. Originally, we wanted to pick a nice terse name for this property so when referencing it in the view, there is minimal noise. We liked the property View for this purpose and implemented it for our view page first.
But when we went to port this property over to the Controller, we realized it wouldn’t work. Anyone care to guess why? Yep, that’s right. Controller already has a method named View so it can’t also have a property named the same thing. So we called it ViewModel for the time being and figured we’d change it once we came up with a better name.
So far, we haven’t come up with a better name that’s both short and descriptive. And before you suggest it, the acronym of “View Data” is not an option.
If you have a better name, do suggest it. :)
Addendum 2: Unit Testing
Someone on Twitter asked me how you would unit test this action method. Here’s an example of a unit tests that shows you can simply call this dynamic method directly from within a unit test (see the act section below).
[TestMethod]
public void CanCallCycle() {
// arrange
var controller = new HomeController();
controller.Index();
// act
string even = controller.ViewModel.Cycle(0, "even", "odd");
// assert
Assert.AreEqual("even", even);
}