Decimal vs Double and Other Tips About Number Types in .NET

I've been coding with .NET for a long time. In all of that time, I haven't really had a need to figure out the nitty-gritty differences between float and double, or between decimal and pretty much any other type. I've just used them as I see fit, and hope that's how they were meant to be used.

Until recently, anyway. Recently, as I was attending the AngleBrackets conference, and one of the coolest parts of attending that conference is getting to be in an in-depth workshop. My particular workshop was called I Will Make You A Better C# Programmer by Kathleen Dollard, and my reaction was thus:

One of the most interesting things I learned at Kathleen's session was that the .NET number types don't always behave the way I think they do. In this post, I'm going to walk through a few (a VERY few) of Kathleen's examples and try to explain why .NET has so many different number types and what they are each for. Come along as I (with her code) attempt to show what the number types are, and what they are used for!

The Number Types in .NET

Let's start with a review of the more common number types in .NET. Here's a few of the basic types:

  • Int16 (aka short): A signed integer with 16 bits (2 bytes) of space available.
  • Int32 (aka int): A signed integer with 32 bits (4 bytes) of space available.
  • Int64 (aka long): A signed integer with 64 bits (8 bytes) of space available.
  • Single (aka float): A 32-bit floating point number.
  • Double (aka double): A 64-bit floating-point number.
  • Decimal (aka decimal): A 128-bit floating-point number with a higher precision and a smaller range than Single or Double.

There's an interesting thing to point out when comparing double and decimal: the range of double is ±5.0 × 10−324 to ±1.7 × 10308, while the range of decimal is (-7.9 x 1028 to 7.9 x 1028) / (100 to 28). In other words, the range of double is several times larger than the range of decimal. The reason for this is that they are used for quite different things.

Precision vs Accuracy

One of the concepts that's really important to discuss when dealing with .NET number types is that of precision vs. accuracy. To make matters more complicated, there are actually two different definitions of precision, one of which I will call arithmetic precision.

  • Precision refers to the closeness of two or more measurements to each other. If you measure something five times and get exactly 4.321 each time, your measurement can be said to be very precise.
  • Accuracy refers to the closeness of a value to standard or known value. If you measure something, and find it's weight to be 4.7kg, but the known value for that object is 10kg, your measurement is not very accurate.
  • Arithmetic precision refers to the number of digits used to represent a number (e.g. how many numbers after the decimal are used). The number 9.87 is less arithmetically precise than the number 9.87654332.

We always need mathematical operations in a computer system to be accurate; we cannot ever expect 4 x 6 = 32. Further, we also need these calculations to be precise using the common term; 4 x 6 must always be precisely 24 no matter how many times we make that calculation. However, the extent to which we want our systems to be either arithmetically precise has a direct impact on the performance of those system.

If we lose some arithmetic precision, we gain performance. The reverse is also true: if we need values to be arithmetically precise, we will spend more time calculating those values. Forgetting this can lead to incredible performance problems, problems which can be solved by using the correct type for the correct problem. These kinds of issues are most clearly shown during Test 3 later in this post.

Why Is Int The Default?

Here's something I've always wondered. If you take this line of code:

var number = 5;  

Why is the type of number always an int? Why not a short, since that takes up less space? Or maybe a long, since that will represent nearly any integer we could possibly use?

Turns out, the answer is, as it often is, performance. .NET optimizes your code to run in a 32-bit architecture, which means that any operations involving 32-bit integers will by definition be more performant than either 16-bit or 64-bit operations. I expect that this will change as we move toward a 64-bit architecture being standard, but for now, 32-bit integers are the most performant option.

Testing the Number Types

One of the ways we can start to see the inherent differences between the above types is by trying to use them in calculations. We're going to see three different tests, each of which will reveal a little more about how .NET uses each of these types.

Test 1: Division

Let's start with a basic example using division. Consider the following code:

private void DivisionTest1()  
{
    int maxDiscountPercent = 30;
    int markupPercent = 20;
    Single niceFactor = 30;
    double discount = maxDiscountPercent * (markupPercent / niceFactor);
    Console.WriteLine("Discount (double): ${0:R}", discount);
}

private void DivisionTest2()  
{
    byte maxDiscountPercent = 30;
    int markupPercent = 20;
    int niceFactor = 30;
    int discount = maxDiscountPercent * (markupPercent / niceFactor);
    Console.WriteLine("Discount (int): ${0}", discount);
}

Note that the only thing that's really different about these two methods are the types of the local variables.

Now here's the question: what will the discount be in each of these methods?

If you said that they'll both be $20, you're missing something very important.

The problem line is this one, from DivisionTest2():

int discount = maxDiscountPercent * (markupPercent / niceFactor);  

Here's the problem: because markupPercent is declared as an int (which in turn makes it an Int32), when you divide an int by another int, the result will be an int, even when we would logically expect it to be something like a double. .NET does this by truncating the result, so because 20 / 30 = 0.6666667, what you get back is 0 (and anything times 0 is 0).

In short, the discount for DivisionTest1 is the expected $20, but the discount for DivisionTest2 is $0, and the only difference between them is what types are used. That's quite a difference, no?

Test 2 - Double Addition

Now we get to see something really weird, and it involves the concept of arithmetic precision from earlier. Here's the next method:

public void DoubleAddition()  
{
    Double x = .1;
    Double result = 10 * x;
    Double result2 = x + x + x + x + x + x + x + x + x + x;

    Console.WriteLine("{0} - {1}", result, result2);
    Console.WriteLine("{0:R} - {1:R}", result, result2);
}

Just by reading this code, we expect result and result2 to be the same: multiplying .1 x 10 should equal .1 + .1 + .1 + .1 + .1 + .1 + .1 + .1 + .1 + .1.

But there's another trick here, and that's the usage of the "{O:R}" string formatter. That's called the round-trip formatter, and it tells .NET to display all parts of this number to its maximum arithmetic precision.

If we run this method, what does the output look like?

By using the round-trip formatter, we see that the multiplication result ended up being exactly 1, but the addition result was off from 1 by a miniscule (but still potentially significant) amount. Question is: why does it do this?

In most systems, a number like 0.1 cannot be accurately represented using binary. There will be some form of arithmetic precision error when using a number such as this. Generally, said arithmetic precision error is not noticeable when doing mathematical operations, but the more operations you perform, the more noticeable the error is. The reason we see the error above is because for the multiplication portion, we only performed one operation, but for the addition portion, we performed ten, and thus caused the arithmetic precision error to compound each time.

Test 3 - Decimal vs Double Performance

Now we get to see something really interesting. I'm often approached by new .NET programmers with a question like the following: why should we use decimal over double and vice-versa? This test pretty clearly spells out when and why you should use these two types.

Here's the sample code:

private int iterations = 100000000;

private void DoubleTest()  
{
    Stopwatch watch = new Stopwatch();
    watch.Start();
    Double z = 0;
    for (int i = 0; i < iterations; i++)
    {
        Double x = i;
        Double y = x * i;
        z += y;
    }
    watch.Stop();
    Console.WriteLine("Double: " + watch.ElapsedTicks);
    Console.WriteLine(z);
}

private void DecimalTest()  
{
    Stopwatch watch = new Stopwatch();
    watch.Start();
    Decimal z = 0;
    for (int i = 0; i < iterations; i++)
    {
        Decimal x = i;
        Decimal y = x * i;
        z += y;
    }
    watch.Stop();
    Console.WriteLine("Decimal: " + watch.ElapsedTicks);
    Console.WriteLine(z);
}

For each of these types, we are doing a series of operations (100 million of them) and seeing how many ticks it takes for the double operation to execute vs how many ticks it takes for the decimal operations to execute. The answer is startling:

The operations involving double take 790836 ticks, while the operations involving decimal take a whopping 16728386 ticks. In other words, the decimal operations take 21 times longer to execute than the double operations. (If you run the sample project, you'll notice that the decimal operations take visibly longer than the double ones).

But why? Why does double take so much less time than decimal?

For one thing, double uses base-2 math, while decimal uses base-10 math. Base-2 math is much quicker for computers to calculate.

Further, what double is concerned with is performance, while what decimal is concerned with is precision. When using double, you are accepting a known trade-off: you won't be super precise in your calculations, but you will get an acceptable answer quickly. Whereas with decimal, precision is built into its type: it's meant to be used for money calculations, and guess how many people would riot if those weren't accurate down to the 23rd place after the decimal point.

In short, double and decimal are two totally separate types for a reason: one is for speed, and the other is for precision. Make sure you use the appropriate one at the appropriate time.

Summary

As can be expected from such a long-lived framework, .NET has several number types to help you with your calculations, ranging from simple integers to complex currency-based values. As always, it's important to use the correct tool for the job:

  • Use double for non-integer math where the most precise answer isn't necessary.
  • Use decimal for non-integer math where precision is needed (e.g. money and currency).
  • Use int by default for any integer-based operations that can use that type, as it will be more performant than short or long.

Don't forget to check out the sample project over on GitHub!

Are there any other pitfalls or improvements we should be aware of? Feel free to sound off in the comments!

Happy Coding!

Huge thanks to Kathleen Dollard (@kathleendollard) for providing the code samples and her invaluable insight into how to effectively explain what's going on in these samples. Check out her Pluralsight course for more!

My Obviously Perfect Code: An Attention-Seeking Post Title

Bombastic opening statement about the strength of the glorious code you're about to read. Further explanation that clearly you haven't seen anything as good as this yet.

Unimportant and overlong diatribe explaining why this forthcoming code snippet is critical to the world at large, including meaningless attempt at invoking "think of the children." Bluntly state that my fellow bloggers have it all wrong, that only I have the true solution. Immediately try to roll back that statement with phrases like "they have their good ideas," but subtly acknowledge that, yes, I am better than them, thanks for asking.

Dive into a story about how, in my early years, something bad happened and compelled me to think of the world differently. Clearly I was so traumatized by this experience that everyone else should be sad for me, hence why I am relating it to you, even though you didn't need or really want me to. Weakly try to link that unnecessary story to the potential improvements the world can harness from my code and fail miserably, but continue to insist that I'm right and that you all should thank me for deigning to spend my irreplaceable time on this God-given snippet.

Attempt to drum up excitement for my impressive snippet by using words like "magical" and "time-saving" and "microservice". Finally get around to introducing the code, but not before once again flatly stating how the world should be thanking me for creating this divinely inspired piece of code.

Final introduction statement, which will invariably include the phrase "and without further ado..."

public static string Truncate(this string input)  
{
    if (input.Length > 40) return input.Substring(0, 40) + "...";
    else return input;
}

Bluntly state that yes, it is as magical as it seems. And, yes, it is also just that simple. Subtly ask for praise while maintaining that you only do this to help your fellow programmers, but really, the praise would be accepted with dignity. And repeatedly. Preferably with Kickstarter donations.

Trite closing paragraph, with sweeping but hollow affirmations that I'm the best coder who ever lived, and everyone else should recognize that. Of course, I will accept all that praise humbly and with deference to my fellow man.

Suitably dull closing signoff statement, almost certainly including the word "coding."

Post image is obviously Simon looking Smug from Flickr, used under license. I mean, come on.

Code Is Ephemeral, Concepts Are Eternal

Lots of people ask me things like "should I learn MVC or Web API first?" "HTML or Javascript?" "Angular or React?" etc. After all, there's only so many hours in the day, and what time we have to spend learning is often limited by other factors (energy, work policies, etc.) This leads to the most common question I get from junior programmers: What framework, stack, or language should I spend my precious time learning?

I always tell them the same thing: It. Does. Not. Matter.

It doesn't matter if you pick Angular, or ASP.NET, or Ruby, or Java. It does not matter if you want to build web sites, or IOS apps, or Windows programs. It does not matter if you're a fresh-out-of-school graduate or a 30-year programming veteran. All of these technologies, all of these platforms, will ultimately help you learn the same things, the same tried-and-true best practices. Just pick one!

Remember: you will be obsolescent someday. That will happen, especially in a business where you must continually stay on top of your own learning in order to do your job. You have a finite number of keystrokes left. Therefore you should spend your limited time learning whatever will stave off that obsolescence for as long as possible.

Concepts fight obsolescence. Even when ASP.NET inevitably dies, the concepts I've learned from programming in it for ten plus years will still be useful. Concepts have a longer shelf life than details, because details change. Languages are born and die, frameworks become unpopular overnight, companies go out of business, support will end. But the thoughts, the ideas, the best practices? They live forever.

Learn about SOLID. Learn KISS, DRY, and YAGNI. Learn how important naming is. Learn about proper spacing, functional vs object-oriented, composition vs. inheritance, polymorphism, etc. Learn soft skills like communication and estimation. Learn all the ideas that result in good code, rather than the details (syntax, limitations, environment, etc.) of the code itself. Mastering the ideas leads to your mind being able to warn you when you are writing bad code (as you will inevitably do).

Don't fret about the how. How you learn the concepts is irrelevant. It doesn't matter what framework you like, what stack you use, what technology you're currently in love with. Just pick one, learn that, master that, and remember some of the pain you had to deal with for the next project. Write a small project, post it to GitHub, blog about it. Get some experience with it! Experience is king, and nothing can substitute for real-world experience.

Code is ephemeral, concepts are eternal. Code is static; it will die, fall apart, degrade. It may take a long time, years or decades, but it will happen. But the concepts which programming lives off of do not die; they live on.

So again I pose the question: what should you spend your precious time learning?

The ASP.NET Web API 2 HTTP Message Lifecycle in 43 Easy Steps

Anyone who works with ASP.NET Web API should check out this poster that Microsoft created to explain the Request/Response Pipeline that Web API utilizes. It's amazing, and if you do any work in Web API you should check it out! Right now. Yes, seriously. Go ahead, I'll wait.

I love this poster, but in my opinion it doesn't do a good job of explaining the decision logic and ideas behind each step in the pipeline. Further, it doesn't explicitly tell you exactly how many things happen during this pipeline (answer: a surprisingly large number of things). In short: it's awesome, but it can be made more awesome by incorporating just a little more detail.

Here's what we're going to do in this post: using that poster, we're going to enumerate every single step involved in processing a request and receiving a response using ASP.NET Web API 2, and explain a little more about each piece of the pipeline and where we programmers can extend, change, or otherwise make more awesome this complex lifecycle. So let's get going and step through the ASP.NET Web API 2 Request Lifecycle in just 43 easy steps!

The 43 Steps

It all starts with IIS:

  1. IIS (or OWIN self-hosting) receives a request.
  2. The request is then passed to an instance of HttpServer.

  3. HttpServer is responsible for dispatching HttpRequestMessage objects.

  4. HttpRequestMessage provides strongly-typed access to the request.

  5. If one or more global instances of DelegatingHandler exist on the pipeline, the request is passed to it. The request arrives at the instances of DelegatingHandler in the order said instances were added to the pipeline.

  6. If the HttpRequestMessage passes the DelegatingHandler instances (or no such handler exists), then the request proceeds to the HttpRoutingDispatcher instance.

    • HttpRoutingDispatcher chooses which routing handler to call based on the matching route. If no such route exists (e.g. Route.Handler is null, as seen in the diagram) then the request proceeds directly to Step 10.

  7. If a Route Handler exists for the given route, the HttpRequestMessage is sent to that handler.

  8. It is possible to have instances of DelegatingHandler attached to individual routes. If such handlers exist, the request goes to them (in the order they were added to the pipeline).
  9. An instance of HttpMessageHandler then handles the request. If you provide a custom HttpMessageHandler, said handler can optionally return the request to the "main" path or to a custom end point.

  10. The request is received by an instance of HttpControllerDispatcher, which will route the request to the appropriate route as determined by the request's URL.

  11. The HttpControllerDispatcher selects the appropriate controller to route the request to.

  12. An instance of IHttpControllerSelector selects the appropriate HttpControllerDescriptor for the given HttpMessage.
  13. The IHttpControllerSelector calls an instance of IHttpControllerTypeResolver, which will finally call...
  14. ...an instance of IAssembliesResolver, which ultimately selects the appropriate controller and returns it to the HttpControllerDispatcher from Step 11.
    • NOTE: If you implement Dependency Injection, the IAssembliesResolver will be replaced by whatever container you register.
  15. Once the HttpControllerDispatcher has a reference to the appropriate controller, it calls the Create() method on an IHttpControllerActivator...
  16. ...which creates the actual controller and returns it to the Dispatcher. The dispatcher then sends the request into the Select Controller Action routine, as shown below.

  17. We now have an instance of ApiController which represents the actual controller class the request is routed to. Said instance calls the SelectAction() method on IHttpActionSelector...

  18. ...which returns an instance of HttpActionDescriptor representing the action that needs to be called.

  19. Once the pipeline has determined which action to route the request to, it executes any Authentication Filters which are inserted into the pipeline (either globally or local to the invoked action).

    • These filters allow you to authenticate requests to either individual actions, entire controllers, or globally throughout the application. Any filters which exist are executed in the order they are added to the pipeline (global filters first, then controller-level filters, then action-level filters).
  20. The request then proceeds to the [Authorization Filters] layer, where any Authorization Filters which exist are applied to the request.

    • Authorization Filters can optionally create their own response and send that back, rather than allowing the request to proceed through the pipeline. These filters are applied in the same manner as Authentication Filters (globally, controller-level, action-level). Note that Authorization Filters can only be used on the Request, not the Response, as it is assumed that if a Response exists, the user had the authorization to generate it.
  21. The request now enters the Model Binding process, which is shown in the next part of the main poster. Each parameter needed by the action can be bound to its value by one of three separate paths. Which path the binding system uses depends on where the value needed exists within the request.

  22. If data needed for an action parameter value exists in the entity body, Web API reads the body of the request; an instance of FormatterParameterBinding will invoke the appropriate formatter classes...

  23. ...which bind the values to a media type (using MediaTypeFormatter)...

  24. ...which results in a new complex type.

  25. If data needed for a parameter value exists in the URL or query string, said URL is passed into an instance of IModelBinder, which uses an IValueProvider to map values to a model (see Phil Haack's post about this topic for more info)....

  26. ...which results in a simple type.

  27. If a custom HttpParameterBinding exists, the system uses that custom binding to build the value...

  28. ...which results in any kind (simple or complex) of object being mappable (see Mike Stall's wonderful series on this topic).

  29. Now that the request is bound to a model, it is passed through any Action Filters which may exist in the pipeline (either globally or just for the action being invoked).

  30. Once the action filters are passed, the action itself is invoked, and the system waits for a response from it.

  31. If the action produces an exception AND an exception filter exists, the exception filter receives and processes the exception.

  32. If no exception occurred, the action produces an instance of HttpResponseMessage by running the Result Conversion subroutine, shown in the next screenshot.

  33. If the return type is already an HttpResponseMessage, we don't need to do any conversion, so pass the return on through.

  34. If the return type is void, .NET will return an HttpResponseMessage with the status 204 No Content.

  35. If the return type is an IHttpActionResult, call the method ExecuteAsync to create an HttpResponseMessage.

    • In any Web API method in which you use return Ok(); or return BadRequest(); or something similar, that return statement follows this process, rather than any of the other processes, since the return type of those actions is IHttpActionResult.
  36. For all other types, .NET will create an HttpResponseMessage and place the serialized value of the return in the body of that message.

  37. Once the HttpResponseMessage has been created, return it to the main pipeline.

  38. Pass the newly-created HttpResponseMessage through any AuthenticationFilters which may exist.

    • Remember that Authorization Filters cannot be used on Responses.

  39. The HttpResponseMessage flows through the HttpControllerDispatcher, which at this point probably won't do anything with it.

  40. The Response also flows through the HttpRoutingDispatcher, which again won't do anything with it.

  41. The Response now proceeds through any DelegatingHandlers that are set up to handle it. At this point, the DelegatingHandler objects can really only change the response being sent (e.g. intercept certain responses and change to the appropriate HTTP status).

  42. The final HttpResponseMessage is given to the HttpServer instance...

  43. ...which returns an Http response to the invoking client.

Tada! We've successfully walked through the entire Web API 2 request/response pipeline, and in only 43 easy steps!

Let me know if this kind of deep dive has been helpful to you, and feel free to share in the comments! Microsoft people and other experts, please chime in to let me know if I got something wrong; I intend for this post to be the definitive guide to the Web API 2 Request/Response Lifecycle, and you can't be definitive without being correct.

Happy Coding!