Sean Kerwin dot org

Imagine that video rental shops worked like libraries; there's a whole network of affiliated branches, and if this library doesn't have it, they can get it at the cousin library across town and deliver it for you. And as long as we're imagining, take it a step further: they can get it for you for no extra cost, absolutely instantly.

So that's the way video rentals work, and they all stock VHS cassettes, and that's that. But now a sea change occurs: somebody invents DVDs. All of the video rental shops start building their stock of DVD titles, and while doing so they take advantage of their affiliation network, so if your favored shop doesn't have Joe Dirt on DVD, they can get it from the shop across town.

But one little shop bucks the trend — even though DVD is the future, they decide to stick with VHS. And that works out okay for them, because while the other rental shops are investing heavily in building a DVD library, they're making room by selling off their VHS cassettes, and this little dark-horse shop is gobbling them up like popcorn, and their library is huge.

Now it's ten years later. Every video rental shop around the world has switched to DVD, but this little shop is still on VHS. It's impossible to buy a VHS deck in stores, so this shop rents or sells you a VHS player when you sign up for membership, but that's okay actually — some people find that kind of convenient.

But now something interesting happens — a big, new, hit movie comes out... but only for DVD. This little shop is caught-flat footed, because they simply can't compete here. So what do they do? They advertise!

This shop starts running ads, talking about how the local Blockbuster has a tiny library of titles, and it's absolutely terrible how small their library is. Now, in a sense that's true; any individual Blockbuster may have a small library available, but keep in mind the affiliate network — when you get a DVD at Blockbuster, you may be getting it from any one of dozens of DVD rental places all over the world.

---

So what does any of that have to do with anything?

Well, VHS is the creaky old CDMA digital cellular technology. DVDs are GSM, the modern cellular network architecture that almost every cellular carrier around the world uses. Our hit DVD is obviously the iPhone.

And the tiny dark-horse video shop, that's Verizon.

Now, don't read this as a defense of AT&T. Personally I haven't had all that much trouble with them, but I know plenty of people who reflexively spit upon hearing those letters, and it's pretty clear that AT&T has been using their iPhone windfall to make money hats rather than investing in their network infrastructure, and that's clearly subpar. But while I dislike incompetence and don't care for spin, I absolutely loathe deception, and that's what Verizon's 'map' ads amount to, in my opinion.

The 'Verizon map' in those ads is, by necessity, a map of Verizon-owned-and-operated coverage zones; they have to be Verizon-owned-and-operated, because nobody else is running CDMA these days, just like nobody's using VHS or Windows 95. The 'AT&T map' to which they compare themselves, meanwhile, is of AT&T-owned-and-operated coverage areas, which is not at all representative because AT&T has reciprocal carriage agreements with GSM carriers throughout the US, and roaming agreements with GSM carriers around the world.

Personally, I can't stand Verizon — they have a policy of installing Verizon-controlled firmware on every phone they support, which tends to cripple the phone by removing built-in features to encourage you to use Verizon's expensive and inferior network-based equivalents. Ever try to copy a ringtone to your Verizon phone using standard Bluetooth file transfer? If you did, I'm betting it didn't go so well. But Verizon does have a lot to offer customers, and if they want to be taken seriously they should compete by explaining the benefits of Verizon rather than deceiving people about the non-deficiencies of their competitors.

Ask each of your developers who the three best developers in the company are. Only ask the developers — managers can't answer because they're in no good position to know. Average out the votes and produce a consensus 'top three' list. Hold on to it, and repeat the survey every six month.

If over a year has passed and two of those three people are still on the list, your company is probably screwed. You might not actually have realized it yet, but that doesn't make it any less true. Your management or business metrics may actually be positive, but if your development team isn't improving, the technology underlying everything is stagnating and sooner or later the smell will reach your customers.

Unless you already employ the best programers in the world, of course.

And if you really and truly believe that you employ the best programmers in the world, then unless your company name rhymes with 'Foogle', 'Schnapple', or 'Sicromoft', you're probably really screwed.

Normal function calls are easy to write; you call DoSomething( ), it executes and returns, and you continue on your merry way.

Asynchronous function calls seem, at first blush, only a little more difficult — instead of calling DoSomething() and executing your follow-up code after it returns, you pass in a callback: BeginDoSomething( Action callback ).

So, problem solved, let's go home. Unless you need a return value, that is. But even then it seems simple; to turn a synchronous method like int CalculateSomething( ) into an asynchronous method, you just pass it a delegate that takes a parameter: void BeginCalculateSomething( Action<int> callback ).

So is that it? Nope. Because all of that is wrong.

Even though the original DoSomething( ) method had no return type, it still had a return path — it could throw an exception. Let's imagine that BeginDoSomething looks something like this:

public void BeginDoSomething(Action callback)
{
    PrepareForLongRunningOperation( );
    
    ThreadPool.QueueUserWorkItem(
        delegate
            {
                LongRunningOperation();

                callback();
            });
}

A handy way to think about this sort of thing is to figure out where a thrown exception would emerge.

If something goes wrong within the call to PrepareForLongRunningOperation, that happens in the same context as the calling code — any exceptions will throw up to the calling code and come out of its call to BeginDoSomething. The same applies to the call to ThreadPool.QueueUserWorkItem — no problem there.

But what if LongRunningOperation throws?

LongRunningOperation would throw up into whatever internal part of the ThreadPool implementation actually launched it. That exception can't come out of ThreadPool.QueueUserWorkItem, because by the time the asynchronous anonymous delegate is running ThreadPool.QueueUserWorkItem has already returned. And since the exception can't come out of ThreadPool.QueueUserWorkItem, it also can't come out of BeginDoSomething — which means there's no way for the calling code to get the exception.

There are two main approaches to this problem — error handlers and completion calls.

Instead of passing your begin method one callback, pass two: a callback to be invoked if everything goes to plan, and an exception-accepting callback to which errors will be passed.

public void BeginDoSomething(Action callback, Action<Exception> errorHandler)
{
    PrepareForLongRunningOperation( );
    
    ThreadPool.QueueUserWorkItem(
        delegate
            {
                try
                {
                    LongRunningOperation();
                    callback();
                }
                catch(Exception ex)
                {
                    errorHandler(ex);
                }
            });
}

// Sample usage:

BeginDoSomething(
    delegate
    {
		// Do something now that we're done
    },
    delegate(Exception ex)
    {
		// Do something with the error
    });

There are strengths and weaknesses to this approach. The biggest strength of this model is that it forces the calling code to think about error handling — the prompt for it is right there in the method signature. Error handling tends to fall through the cracks in any sort of code, but it's especially easy to overlook in an asynchronous context (It's also a lot more dangerous in an asynchronous context, because often dropping a callback invocation will cause a process to spin forever, sucking down resources and accomplishing nothing).

Separating the success case from the failure case may be either a strength or a weakness, depending on the particular task. Sometimes it makes your code much cleaner, but it often happens that your success and error handler need to share context and implementation, which can make for some very ugly code.

Instead of just invoking a parameter-less Action callback or a single-parameter Action<TReturn> callback, your code calls a single-parameter callback and passes it an Action or Func<TReturn> that the callback in turn invokes.

public void BeginDoSomething(Action<Action> callback)
{
    PrepareForLongRunningOperation( );
    
    ThreadPool.QueueUserWorkItem(
        delegate
            {
                try
                {
                    LongRunningOperation();
                    callback(delegate { });
                }
                catch(Exception ex)
                {
                    callback(delegate { throw ex; });
                }
            });
}

// Sample Usage:

BeginDoSomething(
    delegate(Action complete)
    {
        try
        {
            complete();
        }
        catch (Exception ex)
        {
            // Do something with the error
            return;
        }
        // Do something now that we're done
	})

At first blush, this seems like a much clumsier solution; you're essentially trusting the calling code to call your completion method. That's true, at least in this case.

Where completion calls really shine are for asynchronous calls returning values; instead of calling their callback and handing in an Action, you call their callback and give it a Func<TReturn>, which they then must invoke to get their result. That gives you an opportunity to throw exceptions that they can't cleverly bypass:

public void BeginCalculateSomething(Action<Func<int>> callback)
{
    PrepareForLongRunningCalculation();

    ThreadPool.QueueUserWorkItem(
        delegate
        {
            try
            {
                int result = LongRunningCalculation();
                callback(() => result);
            }
            catch (Exception ex)
            {
                callback(delegate { throw ex; });
            }
        });
}

// Sample Usage:

BeginCalculateSomething(
    delegate(Func<int> complete)
    {
        int result;
        try
        {
            result = complete();
        }
        catch (Exception ex)
        {
            // Do something with the error
            return;
        }
        // Do something with the content of result;
	})

Personally I prefer completion calls, mainly because the the pattern works so well for return values. In practice, any time you need this sort of a pattern it's because you care about return values; if you need to ensure that X happens after Y, it's generally because X depends on the result of Y. If X doesn't depend on Y, that's often a sign that you're being too linear in your analysis and that the tasks should be happening in parallel.

Microsoft seems to have collectively reached the same conclusion; IHttpAsyncHandler, the asynchronous methods off of SqlCommand, and the asynchronous forms of WebRequest all use the completion call approach.

public void BeginGetSingle(TIdentityCriteria identityCriteria, CompletionCallback<TItem> callback)
{
    TFilterCriteria filterCriteria =
        CriteriaUtilities.UpgradeCriteria<TIdentityCriteria, TFilterCriteria>(identityCriteria);

    RestClient.BeginGet<TItem>(
        CriteriaUtilities.CriteriaToUrl(
            filterCriteria,
            m_Map,
            m_ServiceUrlBase,
            r => typeof (TItem).FullName.Equals(r.OutputPayloadClass) && r.AllowedVerbs.Contains(Verb.Get)),
        completionFunction => callback(() => completionFunction().Payload));        
}

My function takes a function and when done it calls that function, passing a function that the calling function calls for the result.

Now in fact my function calls another function taking a function accepting a function to call for its result, and to that function it passes a function which when called calls the first function passing a new function that when called calls the function that was passed to the function that my function passed to the other function, thereby returning the result of that function to the function that called my function.

And they say you can't write Lisp code in C#.

Still a work in progress, but pretty cool looking already in my rarely-humble opinion. The trick is to dim the red and green components while leaving the blue intact.

I'm trying to decide the best way to show text bubbles. Everything's pretty scrupulously iPhone-friendly this time around (that's actually what got me started on Xclan again, figuring out how to work with UDP CFSockets in a run loop), but as I understand it uploading new text textures every few frames will bog an iPhone down pretty quickly. I'm thinking of switching my NSOpenGLView to a CAOpenGLLayer and putting a bunch of CATextLayers on top for the bubbles.

There's actually an NSTextField in a layer in the game window; on Leopard it floats above the NSOpenGLView, but in Snow Leopard (where the screen shots were taken) it's invisible. I haven't pursued it because I was planning on replacing the OGL view with a layer anyway, but it's a strange difference.

Dear Reader,

Your code sucks.

It's okay though, so does everyone else's code, including mine. But a big part of improving is recognizing what you're doing wrong, and a big part of that is having random bald guys berate you. So in the spirit of improving the world, here are five thoughts on writing code that sucks less. Heavily C#-centric because that's what I've been reading lately, but I think most of it's meaningful in any environment modulo some terminological adjustments.

1. Dumb Comments Are Dumb

A comment that doesn't say anything useful is worse than worthless; it's actively harmful.

If you have a property named Marklar with a comment telling me that it 'gets or sets the Marklar', you haven't told me anything of value, and reading your useless comment has cost me precious seconds of my life that I'll never get back. You've also probably raised my blood pressure a bit, but I suppose that's more on me than on you.

Now if a Marklar is an odd or confusing type of thing, and you can't reasonably expect everyone to know what it is or why it's there, by all means explain it. Similarly, if this property has unusual behavior in some respect — perhaps it's lazily loaded — mention that in a comment. But don't just put a bloody comment on every property because some never-been-employed professor once told you that that's the right thing to do. It isn't. If someone can't tell you that the EyeColor property lets you get or set the color of the receiver's eyes, they have no business writing code in the first place.

You should only put comments where they're needed, and generally if a comment is legitimately needed it's a sign that you're writing bad code. Don't name the property Age and then add a comment explaining that it's in seconds; name it AgeInSeconds and skip the comment entirely. That saves you typing, saves me reading, and has the further benefit of being completely impossible for future programmers to miss unless they're severely mentally impaired (in which case, see above, they have no business writing code in the first place).

Similarly, if you feel compelled to add a comment explaining that the next five lines of code are stochastically approximating the chronosynclastic infundibulumation of the yada, yada... that's your sign that those five lines should be extracted into a method with an appropriate name like 'AproximateTechnobabble'. That makes it easier for me to read the original method because I can temporarily ignore the hard stuff, while also making it easier for me to understand the hard stuff later on if I need to (since it's now isolated with clear, labeled inputs).

Finally, dumb comments are dumb not just in the code, but everywhere. If your checkin note is "I changed the twelfth character of line 37 from an 'e' to a 'z'", you have so completely and cleanly missed the point of checkin notes that I can't even scrape together a proper disparaging simile. Checkin notes are there so I can get a high-level understanding of what you've checked in, without having to open the bloody file. "Ran a match along a patch of bark" is a lousy checkin note when what you actually did was "burned down the forest". Tell me about the forest, I couldn't care less about the trees.

2. Every Time You Type '#region' I Die A Little Inside

Regions suck. If I were king I'd ban 'em outright, so sincere and total is my dislike. They're a trick you play on yourself to let you pretend that your bad code is good, and they tell the world that you're a good enough coder to know a three-thousand-line file stinks, but that you're either not good enough to fix it or you're just plain too lazy. In either situation the solution is to step up, not to add a region and sweep part of the bad code under the rug.

A special note on regions within methods: wrapping a region around your twelve constructors to hide the excess complexity is a venal sin. Putting a region inside your method to hide 3,000 of its 4,000 lines is a mortal sin.

Some people might ask why extracting methods (like AproximateTechnobabble above) is good but collapsing regions within a method is evil. The simple answer is encapsulation; a region isn't a new scope, so it has access to all the fields and variables of the method it's within. That greatly increases the difficulty of understanding the code, because now I need to actively read through and figure out how the code inside the region interacts with the code outside the region. An extracted method doesn't have the problem, because I can look at exactly what parameters are passed in and exactly what comes out, without having to trace through and worry about side effects. Regions superficially seem to simplify the complexity of your code; extracted methods actually do.

3. Static Methods Are Your Friends, And You Should Want More Of Them

Static methods are pure functions, in the mathematical sense; they're black boxes that take some input and return some output. Remember where I said methods are better than regions because they simplify complexity? The same logic recommends static methods over non-static methods. Static methods help you make sure you're not weaving a tangled web of side effects and dependencies with each method call.

Imagine coming upon a call to myComplexMarklar.CalculateInfundibulumation(). Now, if you don't know what infundibulumation is (incidentally, the state of being or the act of becoming funnel-like or akin to a funnel), and have no idea how you would calculate it (no idea), that method doesn't tell you a whole lot. Now imagine stepping through that rather lengthy method and figuring out how it works. You are potentially entering a world of pain.

Now imagine a static method, Marklar.CalculateInfundibulumation(float topRadius, float bottomRadius, float height), and further imagine that the non-static method works by calling the static method and passing along the three appropriate values. It has instantly become a whole lot easier for me to understand what infundibulumation is, because now I know that no matter how gnarly the math is it's just a function of those three values. Now not only can I better understand that particular method, but a lot of the meaning of the Marklar class in general has become more clear.

In an object-oriented language, every instance method is getting more parameters than you've listed in the declaration line; it's also getting access to all of the instance's member fields. Imagine seeing that written out — would you deliberately write a method that takes a hundred parameters but only uses two of them? That might be a good sign that it's time for some judiciously-applied static methods.

Static methods have other benefits. For one, they're a lot more reusable; if the static CalculateInfundibulumation( tR, bR, h ) method is general, you might find you need it in another class. Depending on your situation, you could either make it public and call it right out of Marklar, move it to a new common base class, or move it to a utility class. Each of those solutions would be far more complicated if it were non-static. I've been in situations where I needed a complex non-static method from a pre-existing class that I couldn't reasonably modify, and ended up constructing an instance with a bunch of bogus data, setting the handful of fields that the calculation actually depends on, and calling the instance method on that. That's hideous — ugliness wrapped in brittleness wrapped in unmaintainable complexity. Don't make me do that again.

Static methods also lend themselves incredibly well to a form of optimization known as memoization. The root of memoization is basically the realization that the next time you run the same calculation, you'll reach the same result (yeah, I know, 'duh', right?). The tie-in to static methods is that static methods let you know precisely what the inputs are to the calculation you're running, while instance methods don't. Memoizing a static method is 'easy' (in the sense of being a solved problem), while memoizing instance methods can be a buggy pain in the butt.

3b. Subverting Static Methods Is Bad And People Who Do It Should Be Forced To Write Pascal

Sometimes you'll see a static class, full of static fields and static variables and static methods. That's not what I'm talking about when I say static is good. That's not good, it's the other thing. Bad. It's a crappy-ass degenerate-global-variable-singleton. Don't do that. Smart people hate singletons.

4. Properties Are Properties, Methods Are Methods

A property is intended to represent an attribute of an object, while a method models an action that the object may take or that may be taken upon the object. These are different concepts, and you need to think about which is appropriate for a particular case. If you have a property with a verb in the name, you're almost certainly thinking about things the wrong way.

Some people will tell you that properties should just be get/set wrappers around your fields. Others will say it's okay to do some field validation. Others go hog wild. It's hard to draw bright lines because the simple fact is that it's sometimes a judgement call. A good heuristic is to think like someone using your class. The default assumption is that properties are simple get/set wrappers around a private field, and everything you do to violate that assumption nudges the scale towards building a method.

Let's say your class has a ServerConnection property that returns some sort of communications channel to a remote server. As a caller, my mindset is that this connection already exists, and the property is giving me access to it. But if only a subset of the calling code ever needs this connection, it might make sense to lazily connect when the property is first accessed rather than connecting in the constructor and keeping around a connection that may never actually be needed. But that's changing the semantics of the property a little bit, because the mindset of the calling code is still 'this already exists, the property is just handing it to me', and the calling code may not expect this property to take a long time to execute. Maybe now there should be a progress bar, or the call should happen in another thread.

Like I said, judgement call. In situations like that one, asking ten developers will get you fifteen different opinions. There isn't a 'right' answer, it's just a matter of getting it right enough. All anyone can ask is that you think.

But in other cases, it's totally cut and dried. I've seen code with a LastSaveDate property that saves to the database and returns DateTime.Now. If you're not physically ill at the thought of that, you need to figure why and get sick fast.

5. Public Methods Are Contracts, Which Should Not Be Entered Lightly

When you make a method public, you're inviting the world to use it and you're promising that it will continue to do exactly what it does now. Forever.

Let's say you're creating a CarFactory class that creates new Car objects. CarFactory.CreateCar( ) was starting to get a little long, so you started extracting out helper methods: CreateTransmission( ), CreateBody( ), CreateWheel( ), etc. That's good, keep doing it.

Now let's say you decided that since CreateCar( ) was public that you might as well make CreateTransmission( ) et al public too. That's no-so-much with the good. Unless you know right now that someone else needs to create a transmission without creating a full car, leave that method private.

To the calling code, a CarFactory and a Car are black boxes. The calling code doesn't need to know that a car contains a transmission, and next year when you change CarFactory and Car to build fully-electric vehicles, which don't need a transmission, you won't have to worry about supporting that public method that a bunch of your coworkers have just spent twelve months linking against.

There's a more general lesson here: Share the hot dog, not the making thereof.

Which, come to think of it, is a good moral on which to end any story.

I 'got' Twitter pretty quickly, because I was actually using it before it existed. I used to carry on lossy low-priority conversations with people via AIM status messages; it was a good way to say stuff that wasn't important enough to warrant a message and also to broadcast things in the vein of "hey, here's a fun link" without needing to directly spam people.

I'm pretty sure this is roughly what the designers of Twitter had in mind, incidentally. If you take a look at the XML or JSON feeds provided by the Twitter API, each individual tweet is actually called a 'status' — the URI for a particular tweet is <username>/status/<some_big_number> — the <OL> element that contains the tweets on Twitter HTML page is of CSS class 'statuses'... the list goes on. Obviously the original intent of the tool doesn't define what the tool is actually used for — I don't think AIM away messages were intended as an out-of-band communications protocol, at least at first — but it's interesting nonetheless.

What's also interesting is that Twitter users seem to fall into two separate and largely non-overlapping groups. Group A are people using it much like I am, as a non-ACKed broadcast medium for random thoughts that aren't important enough for a reliable messaging protocol like IM, email, or a blog post. Group B are the people who see Twitter as a marketing/communication tool that will allow them to get the word out to members of Group A, for whatever value of 'the word' is meaningful to them.

I see three ways this can go: either the two groups diverge completely and group B develops a self-sustaining ecosystem of their own; or group B recognizes that marketing in this context is futile and dwindles and dies; or group B is so persistent in their efforts that group A abandons Twitter and finds a new bicycle.

If you get the 'new bicycle' thing, you're definitely in group A, though not getting it doesn't necessarily relegate you to group B.

But if you know you're in group B and want to understand group A, then you should figure out what it means, why you don't get it, and what you can do to make sure you get to autotune the new new bicycle... and then despair of ever marketing effectively on something like Twitter.