October 2014

Table of Contents

Part One

  1. Category: The Essence of Composition
  2. Types and Functions
  3. Categories Great and Small
  4. Kleisli Categories
  5. Products and Coproducts
  6. Simple Algebraic Data Types
  7. Functors
  8. Functoriality
  9. Function Types
  10. Natural Transformations

Part Two

  1. Declarative Programming
  2. Limits and Colimits
  3. Free Monoids
  4. Representable Functors
  5. The Yoneda Lemma
  6. Yoneda Embedding

Part Three

  1. It’s All About Morphisms
  2. Adjunctions
  3. Free/Forgetful Adjunctions
  4. Monads: Programmer’s Definition
  5. Monads and Effects
  6. Monads Categorically
  7. Comonads
  8. F-Algebras
  9. Algebras for Monads
  10. Kan Extensions

You may also watch me teaching this material to a live audience.


For some time now I’ve been floating the idea of writing a book about category theory that would be targeted at programmers. Mind you, not computer scientists but programmers — engineers rather than scientists. I know this sounds crazy and I am properly scared. I can’t deny that there is a huge gap between science and engineering because I have worked on both sides of the divide. But I’ve always felt a very strong compulsion to explain things. I have tremendous admiration for Richard Feynman who was the master of simple explanations. I know I’m no Feynman, but I will try my best. I’m starting by publishing this preface — which is supposed to motivate the reader to learn category theory — in hopes of starting a discussion and soliciting feedback.

I will attempt, in the space of a few paragraphs, to convince you that this book is written for you, and whatever objections you might have to learning one of the most abstract branches of mathematics in your “copious spare time” are totally unfounded.

My optimism is based on several observations. First, category theory is a treasure trove of extremely useful programming ideas. Haskell programmers have been tapping this resource for a long time, and the ideas are slowly percolating into other languages, but this process is too slow. We need to speed it up.

Second, there are many different kinds of math, and they appeal to different audiences. You might be allergic to calculus or algebra, but it doesn’t mean you won’t enjoy category theory. I would go as far as to argue that category theory is the kind of math that is particularly well suited for the minds of programmers. That’s because category theory — rather than dealing with particulars — deals with structure. It deals with the kind of structure that makes programs composable.

Composition is at the very root of category theory — it’s part of the definition of the category itself. And I will argue strongly that composition is the essence of programming. We’ve been composing things forever, long before some great engineer came up with the idea of a subroutine. Some time ago the principles of structural programming revolutionized programming because they made blocks of code composable. Then came object oriented programming, which is all about composing objects. Functional programming is not only about composing functions and algebraic data structures — it makes concurrency composable — something that’s virtually impossible with other programming paradigms.

Third, I have a secret weapon, a butcher’s knife, with which I will butcher math to make it more palatable to programmers. When you’re a professional mathematician, you have to be very careful to get all your assumptions straight, qualify every statement properly, and construct all your proofs rigorously. This makes mathematical papers and books extremely hard to read for an outsider. I’m a physicist by training, and in physics we made amazing advances using informal reasoning. Mathematicians laughed at the Dirac delta function, which was made up on the spot by the great physicist P. A. M. Dirac to solve some differential equations. They stopped laughing when they discovered a completely new branch of calculus called distribution theory that formalized Dirac’s insights.

Of course when using hand-waving arguments you run the risk of saying something blatantly wrong, so I will try to make sure that there is solid mathematical theory behind informal arguments in this book. I do have a worn-out copy of Saunders Mac Lane’s Category Theory for the Working Mathematician on my nightstand.

Since this is category theory for programmers I will illustrate all major concepts using computer code. You are probably aware that functional languages are closer to math than the more popular imperative languages. They also offer more abstracting power. So a natural temptation would be to say: You must learn Haskell before the bounty of category theory becomes available to you. But that would imply that category theory has no application outside of functional programming and that’s simply not true. So I will provide a lot of C++ examples. Granted, you’ll have to overcome some ugly syntax, the patterns might not stand out from the background of verbosity, and you might be forced to do some copy and paste in lieu of higher abstraction, but that’s just the lot of a C++ programmer.

But you’re not off the hook as far as Haskell is concerned. You don’t have to become a Haskell programmer, but you need it as a language for sketching and documenting ideas to be implemented in C++. That’s exactly how I got started with Haskell. I found its terse syntax and powerful type system a great help in understanding and implementing C++ templates, data structures, and algorithms. But since I can’t expect the readers to already know Haskell, I will introduce it slowly and explain everything as I go.

If you’re an experienced programmer, you might be asking yourself: I’ve been coding for so long without worrying about category theory or functional methods, so what’s changed? Surely you can’t help but notice that there’s been a steady stream of new functional features invading imperative languages. Even Java, the bastion of object-oriented programming, let the lambdas in. C++ has recently been evolving at a frantic pace — a new standard every few years — trying to catch up with the changing world. All this activity is in preparation for a disruptive change or, as we physicist call it, a phase transition. If you keep heating water, it will eventually start boiling. We are now in the position of a frog that must decide if it should continue swimming in increasingly hot water, or start looking for some alternatives.


One of the forces that are driving the big change is the multicore revolution. The prevailing programming paradigm, object oriented programming, doesn’t buy you anything in the realm of concurrency and parallelism, and instead encourages dangerous and buggy design. Data hiding, the basic premise of object orientation, when combined with sharing and mutation, becomes a recipe for data races. The idea of combining a mutex with the data it protects is nice but, unfortunately, locks don’t compose, and lock hiding makes deadlocks more likely and harder to debug.

But even in the absence of concurrency, the growing complexity of software systems is testing the limits of scalability of the imperative paradigm. To put it simply, side effects are getting out of hand. Granted, functions that have side effects are often convenient and easy to write. Their effects can in principle be encoded in their names and in the comments. A function called SetPassword or WriteFile is obviously mutating some state and generating side effects, and we are used to dealing with that. It’s only when we start composing functions that have side effects on top of other functions that have side effects, and so on, that things start getting hairy. It’s not that side effects are inherently bad — it’s the fact that they are hidden from view that makes them impossible to manage at larger scales. Side effects don’t scale, and imperative programming is all about side effects.

Changes in hardware and the growing complexity of software are forcing us to rethink the foundations of programming. Just like the builders of Europe’s great gothic cathedrals we’ve been honing our craft to the limits of material and structure. There is an unfinished gothic cathedral in Beauvais, France, that stands witness to this deeply human struggle with limitations. It was intended to beat all previous records of height and lightness, but it suffered a series of collapses. Ad hoc measures like iron rods and wooden supports keep it from disintegrating, but obviously a lot of things went wrong. From a modern perspective, it’s a miracle that so many gothic structures had been successfully completed without the help of modern material science, computer modelling, finite element analysis, and general math and physics. I hope future generations will be as admiring of the programming skills we’ve been displaying in building complex operating systems, web servers, and the internet infrastructure. And, frankly, they should, because we’ve done all this based on very flimsy theoretical foundations. We have to fix those foundations if we want to move forward.

Ad hoc measures preventing the Beauvais cathedral from collapsing

Ad hoc measures preventing the Beauvais cathedral from collapsing

Next: Category: The Essence of Composition.

Data races lead to undefined behavior; but how bad can they really be? In my previous post I talked about benign data races and I gave several examples taken from the Windows kernel. Those examples worked because the kernel was compiled with a specific compiler for a specific processor. But in general, if you want your code to be portable, you can’t have data races, period.

You just cannot reason about something that is specifically defined to be “undefined.” So, obviously, you cannot prove the correctness of a program that has data races. However, very few people engage in proofs of correctness. In most cases the argument goes, “I can’t see how this can fail, therefore it must be correct” (maybe not in so many words). I call this “proof by lack of imagination.” If you want to become a concurrency expert, you have to constantly stretch your imagination. So let’s do a few stretches.

One of the readers of my previous post, Duarte Nunes, posted a very interesting example of a benign data race. Here’s the code:

int owner = 0;
int count = 0;
std::mutex mtx;

bool TryEnter() {
    if (owner == std::this_thread::get_id()) {
        count += 1;
        return true;

    if (mtx.try_lock()) {
        owner = std::this_thread::get_id();
        return true;
    return false;

void Exit() {
    if (count != 0) {
        count -= 1;
    owner = 0;

I highlighted in blue the parts that are executed under the lock (in a correct program, Exit will always be called after the lock has been acquired). Notice that the variable count is always accessed under the lock, so no data races there. However, the variable owner may be read outside of the lock– I highlighted that part of code in red. That’s the data race we are talking about.

Try to analyze this code and imagine what could go wrong. Notice that the compiler or the processor can’t be too malicious. The code still has to work correctly if the data race is removed, for instance if the racy read is put under the lock.

Here’s an attempt at the “proof” of correctness. First, Duarte observed that “The valid values for the owner variable are zero and the id of any thread in the process.” That sort of makes sense, doesn’t it? Now, the only way the racy read can have any effect is if the value of owner is equal to the current thread’s ID. But that’s exactly the value that could have been written only by the current thread– and under the lock.

There are two possibilities when reading owner: either we are still under that lock, in which case the read is not at all racy; or we have already released the lock. But the release of the lock happens only after the current thread zeroes owner.

Notice that this is a single-thread analysis and, within a single thread, events must be ordered (no need to discuss memory fences or acquire/release semantics at this point). A read following a write in the same thread cannot see the value that was there before the write. That would break regular single-threaded programs. Of course, other threads may have overwritten this zero with their own thread IDs, but never with the current thread ID. Or so the story goes…

Brace yourself: Here’s an example how a compiler or the processor may “optimize” the program:

void Exit() {
    if (count != 0) {
        count -= 1;
    owner = 42;
    owner = 0;

You might argue that this is insane and no compiler in its right mind would do a thing like this; but the truth is: It’s a legal program transformation. The effect of this modification is definitely not observable in the current thread. Neither is it observable by other threads in the absence of data races. Now, the unfortunate thread whose ID just happens to be 42 might observe this value and take the wrong turn. But it can only observe it through a racy read. For instance, it would never see this value if it read owner under the lock. Moreover, it would never see it if the variable owner were defined as ‘atomic’:

std::atomic<int> owner = 0

Stores and loads of atomic variables are, by default, sequentially consistent. Unfortunately sequential consistency, even on an x86, requires memory fences, which can be quite costly. It would definitely be an overkill in our example. So here’s the trick: Tell the compiler to forgo sequential consistency on a per read/write basis. For instance, here’s how you read an atomic variable without imposing any ordering constraints:


Such ‘relaxed’ operations will not introduce any memory fences– at least not on any processor I know about.

Here’s the version of the code that is exactly equivalent to the original, except that it’s correct and portable:

std::atomic<int> owner = 0;
int count = 0;
std::mutex mtx;

bool TryEnter() {
    if (owner.load(memory_order_relaxed) == std::this_thread::get_id()) {
        count += 1;
        return true;

    if (mtx.try_lock()) {
        owner.store(std::this_thread::get_id(), memory_order_relaxed);
        return true;
    return false;

void Exit() {
    if (count != 0) {
        count -= 1;
    owner.store(0, memory_order_relaxed);

So what has changed? Can’t the compiler still do the same dirty trick, and momentarily store 42 in the owner variable? No, it can’t! Since the variable is declared ‘atomic,’ the compiler can no longer assume that the write can’t be observed by other threads.

The new version has no data races: The Standard specifically states that ‘atomic’ variables don’t contribute to data races, even in their most relaxed form.

C++ Standard, (1.10.5):
[…] “Relaxed” atomic operations are not synchronization operations even though, like synchronization operations, they cannot contribute to data races.

With those changes, I believe that our “proof” of correctness may actually be turned into a more rigorous proof using the axioms of the C++ memory model (although I’m not prepared to present one). We can have our cake (correct, portable code) and eat it too (no loss of performance). And, by the way, the same trick may be used in the case of lossy counters from my previous post.

Warning: I do not recommend this style of coding, or the use of weak atomics, to anybody who is not implementing operating system kernels or concurrency libraries.


I’d like to thank Luis Ceze, Hans Boehm, and Anthony Williams for useful remarks and for verifying my assumptions about the C++ memory model.


  1. C++ Draft Standard

We have this friendly competition going on between Eric Niebler and myself. He writes some clever C++ template code, and I feel the compulsion to explain it to him in functional terms. Then I write a blog about Haskell or category theory and Eric feels a compulsion to translate it into C++.

Eric is now working on his proposal to rewrite the C++ STL in terms of ranges and I keep reinterpreting his work in terms familiar to functional programmers. Eric’s range comprehensions are a result of some of this back and forth.

Lazy ranges are such an excellent example of functional programming that it would be foolish for me to pass this opportunity to dissect them. To any functional programmer worth their salt they just scream “monad!” A monad is a higher order pattern that can be built step by step, so that’s what I’m going to do. I’ll start with the functor pattern, then add some functionality that will make it a pointed functor, then add some more to make it an applicative functor, and finally add some more to make it a monad. This gradual buildup of functionality is reminiscent of building a class hierarchy, and indeed it can be modelled as such in Haskell (although Haskell type classes are slightly different than C++ classes). This hierarchy would look something like this:

  • A monad is-an applicative functor
  • An applicative functor is-a pointed functor
  • A pointed functor is-a functor

So let’s start with a functor.


I have a pet peeve about the use of the word “functor” in C++. People keep calling function objects functors. It’s like calling Luciano Pavarotti an “operator,” because he sings operas. The word functor has a very precise meaning in mathematics — moreover, it’s the branch of mathematics that’s extremely relevant to programming. So hijacking this term to mean a function-like object causes unnecessary confusion.

A functor in functional programming is a generic template, which allows the “lifting” of functions. Let me explain what it means. A generic template takes an arbitrary type as a template argument. So a range (whether lazy or eager) is a generic template because it can be instantiated for any type. You can have a range of integers, a range of vectors, a range of ranges, and so on. (We’ll come back to ranges of ranges later when we talk about monads.)

The “lifting” of functions means this: Give me any function from some type T to some other type U and I can apply this function to a range of T and produce a range of U. You may recognize this kind of lifting in the STL algorithm std::transform, which can be used to apply a function to a container. STL containers are indeed functors. Unfortunately, their functorial nature is buried under the noise of iterators. In Eric’s range library, the lifting is done much more cleanly using view::transform. Have a look at this example:

 int total = accumulate(view::iota(1) |
                        view::transform([](int x){return x*x;}) |
                        view::take(10), 0);

Here, view::transform takes an anonymous function that squares its argument, and lifts this function to the level of ranges. The range created by view::iota(1) is piped into it from the left, and the resulting rage of squares emerges from it on the right. The (infinite) range is then truncated by take‘ing the first 10 elements.

The function view::iota(1) is a factory that produces an infinite range of consecutive integers starting from 1. (We’ll come back to range factories later.)

In this form, view::transform plays the role of a higher-order function: one that takes a function and returns a function. It almost reaches the level of terseness and elegance of Haskell, where this example would look something like this:

total = sum $ take 10 $ fmap (\x->x*x) [1..]

(Traditionally, the flow of data in Haskell is from right to left.) The (higher-order) function fmap can be thought of as a “method” of the class Functor that does the lifting in Haskell. In C++ there is no overall functor abstraction, so each functor names its lifting function differently — for ranges, it’s view::transform.

The intuition behind a functor is that it generates a family of objects that somehow encapsulate values of arbitrary types. This encapsulation can be very concrete or very abstract. For instance, a container simply contains values of a given type. A range provides access to values that may be stored in some other container. A lazy range generates values on demand. A future, which is also a functor (or will be, in C++17), describes a value that might not be currently available because it’s being evaluated in a separate thread.

All these objects have one thing in common: they provide means to manipulate the encapsulated values with functions. That’s the only requirement for a functor. It’s not required that a functor provide access to encapsulated values (which may not even exist), although most do. In fact there is a functor (really, a monad), in Haskell, that provides no way of accessing its values other than outputting them to external devices.

Pointed Functor

A pointed functor is a functor with one additional ability: it lets you lift individual values. Give me a value of any type and I will encapsulate it. In Haskell, the encapsulating function is called pure although, as we will see later, in the context of a monad it’s called return.

All containers are pointed, because you can always create a singleton container — one that contains only one value. Ranges are more interesting. You can obviously create a range from a singleton container. But you can also create a lazy range from a value using a (generic) function called view::single, which doesn’t have a backing container behind it.

There is, however, an alternative way of lifting a value to a range, and that is by repeating it indefinitely. The function that creates such infinite (lazy) ranges is called view::repeat. For instance, view::repeat(1) creates an infinite series of ones. You might be wondering what use could there be of such a repetitive range. Not much, unless you combine it with other ranges. In general, pointed functors are not very interesting other than as stepping stones towards applicative functors and monads. So let’s move on.

Applicative Functor

An applicative functor is a pointed functor with one additional ability: it lets you lift multi-argument functions. We already know how to lift a single-argument function using fmap (or transform, or whatever it’s called for a particular functor).

With multi-argument functions acting on ranges we have two different options corresponding to the two choices for pure I mentioned before: view::single and view::repeat.

The idea, taken from functional languages, is to consider what happens when you provide the fist argument to a function of multiple arguments (it’s called partial application). You don’t get back a value. Instead you get something that expects one or more remaining arguments. A thing that expects arguments is called a function (or a function object), so you get back a function of one fewer arguments. In C++ you can’t just call a function with fewer arguments than expected, or you get a compilation error, but there is a (higher-order) function in the Standard Library called std::bind that implements partial application.

This kind of transformation from a function of multiple arguments to a function of one argument that returns a function is called currying.

Let’s consider a simple example. We want to apply std::make_pair to two ranges: view::ints(10, 11) and view::ints(1, 3). To this end, let’s replace std::make_pair with the corresponding curried function of one argument returning a function of one argument:

[](int i) { return [i](int j) { return std::make_pair(i, j); };}

First, we want to apply this function to the first range. We know how to apply a function to a range: we use view::transform.

auto partial_app = view::ints(10, 11) 
                 | view::transform([](int i) { 
                      return [i](int j) { return std::make_pair(i, j); }

What’s the result of this application? Can you guess? Our curried function will be applied to each integer in the range, returning a function that pairs that integer with its argument. So we end up with a range of functions of the form:

[i](int j) { return std::make_pair(i, j); }

So far so good — we have just used the functorial property of the range. But now we have to decide how to apply a range of functions to the second range of values. And that’s the essence of the definition of an applicative functor. In Haskell the operation of applying encapsulated functions to encapsulated arguments is denoted by an infix operator <*>.

With ranges, there are two basic strategies:

  1. We can enumerate all possible combinations — in other words create the cartesian product of the range of functions with the range of values — or
  2. Match corresponding functions to corresponding values — in other words, “zip” the two ranges.

The former, when applied to view::ints(1, 3), will yield:


and the latter will yield:

{(10, 1),(11, 2)}

(when the ranges are not equal length, you stop zipping when the shorter one is exhausted).

To see that those two choices correspond to the two choices for pure, we have to look at some consistency conditions. One of them is that if you encapsulate a single-argument function in a range using pure and then apply it to a range of arguments, you should get the same result as if you simply fmapped this function directly over the range of arguments. For the record, I’ll write it here in Haskell:

pure f <*> xs == fmap f xs

This is sort of an obvious requirement: You have two different ways of applying a single-argument function to a range, they better give the same result.

Let’s try it with the view::single version of pure. When acting on a function, it will create a one-element range containing this function. The “all possible combinations” application will just apply this function to all elements of the argument range, which is exactly what view::transform would do.

Conversely, if we apply view::repeat to the function, we’ll get an infinite range that repeats this function at every position. We have to zip this range with the range of arguments in order to get the same result as view::transform. So this implementation of pure works with the zippy applicative. Notice that if the argument range is finite the result will also be finite. But this kind of application will also work with infinite ranges thanks to laziness.

So there are two legitimate implementations of the applicative functor for ranges. One uses view::single to lift values and uses the all possible combinations strategy to apply a range of functions to a range of arguments. The other uses view::repeat to lift values and the zipping application for ranges of functions and arguments. They are both acceptable and have their uses.

Now let’s go back to our original problem of applying a function of multiple arguments to a bunch of ranges. Since we are not doing it in Haskell, currying is not really a practical option.

As it turns out, the second version of applicative has been implemented by Eric as a (higher-order) function view::zip_with. This function takes a multi-argument callable object as its first argument, followed by a variadic list of ranges.

There is no corresponding implementation for the combinatorial applicative. I think the most natural interface would be an overload of view::transform (or maybe view::fmap) with the same signature as zip_with. Our example would then look like this:

view::transform(std::make_pair, view::ints(10, 11), view::ints(1, 3));

The need for this kind of interface is not as acute because, as we’ll learn next, the combinatorial applicative is supplanted by a more general monadic interface.


Monads are applicative functors with one additional functionality. There are two equivalent ways of describing this functionality. But let me first explain why this functionality is needed.

The range library comes with a bunch of range factories, such as view::iota, view::ints, or view::repeat. It’s also very likely that users of the library will want to create their own range factories. The problem is: How do you compose existing range factories to obtain new range factories?

Let me give you an example that generated a blog post exchange between me and Eric. The challenge was to generate a list of Pythagorean triples. The idea is to take a cross product of three infinite ranges of integers and select those triples that satisfy the equation x2 + y2 = z2. The cross product of ranges is reminiscent of the “all possible combinations” applicative, and indeed that’s the applicative that can be extended to a monad (the zippy one can’t).

To make this algorithm feasible, we have to organize these ranges so we can (lazily) traverse them. Let’s start with a factory that produces all integers from 1 to infinity. That’s the view::ints(1) factory. Then, for each z produced by that factory, let’s create another factory view::ints(1, z). This range will provide our xs — and it makes no sense to try xs that are bigger than zs. These values, in turn, will be used in the creation of the third factory, view::ints(x, z) that will generate our ys. At the end we’ll filter out the triples that don’t satisfy the Pythagorean equation.

Notice how we are feeding the output of one range factory into another range factory. Do you see the problem? We can’t just iterate over an infinite range. We need a way to glue the output side of one range factory to the input side of another range factory without unpacking the range. And that’s what monads are for.

Remember also that there are functors that provide no way of extracting values, or for which extraction is expensive or blocking (as is the case with futures). Being able to compose those kinds of functor factories is often essential, and again, the monad is the answer.

Now let’s pinpoint the type of functionality that would allow us to glue range factories end-to-end. Since ranges are functorial, we can use view::transform to apply a factory to a range. After all a factory is just a function. The only problem is that the result of such application is a range of ranges. So, really, all that’s needed is a lazy way of flattening nested ranges. And that’s exactly what Eric’s view::flatten does.

With this new flattening power at our disposal, here’s a possible beginning of the solution to the Pythagorean triple problem:

view::ints(1) | view::transform([](int z) { 
                view::ints(1, z) | ... } | view::flatten

However, this combination of view::transform and view::flatten is so useful that it deserves its own function. In Haskell, this function is called “bind” and is written as an infix operator >>=. (And, while we’re at it, flatten is called join.)

And guess what the combination of view::transform and view::flatten is called in the range library. This discovery struck me as I was looking at one of Eric’s examples. It’s called view::for_each. Here’s the solution to the Pythagorean triple problem using view::for_each to bind range factories:

auto triples =
  for_each(ints(1), [](int z) {
    return for_each(ints(1, z), [=](int x) {
      return for_each(ints(x, z), [=](int y) {
        return yield_if(x*x + y*y == z*z, std::make_tuple(x, y, z));

And here’s the same code in Haskell:

triples = 
  (>>=) [1..] $ \z -> 
     (>>=) [1..z] $ \x -> 
        (>>=) [x..z] $ \y -> 
           guard (x^2 + y^2 == z^2) >> return (x, y, z)

I have purposefully re-formatted Haskell code to match C++ (A more realistic rendition of it is in my post Getting Lazy with C++). Bind operators >>= are normally used in infix position but here I wanted to match them against for_each. Haskell’s return is the same as view::single, which Eric renamed to yield inside for_each. In this particular case, yield is conditional, which in Haskell is expressed using guard. The syntax for lambdas is also different, but otherwise the code matches almost perfectly.

This is an amazing and somewhat unexpected convergence. In our tweeter exchange, Eric sheepishly called his for_each code imperative. We are used to thinking of for_each as synonymous with looping, which is such an iconic imperative construct. But here, for_each is the monadic bind — the epitome of functional programming. This puppy is purely functional. It’s an expression that returns a value (a range) and has no side effects.

But what about those loops that do have side effects and don’t return a value? In Haskell, side effects are always encapsulated using monads. The equivalent of a for_each loop with side effects would return a monadic object. What we consider side effects would be encapsulated in that object. It’s not the loop that performs side effects, its that object. It’s an executable object. In the simplest case, this object contains a function that may be called with the state that is to be modified. For side effects that involve the external world, there is a special monad called the IO monad. You can produce IO objects, you can compose them using monadic bind, but you can’t execute them. Instead you return one such object that combines all the IO of your program from main and let the runtime execute it. (At least that’s the theory.)

Is this in any way relevant to an imperative programmer? After all, in C++ you can perform side effects anywhere in your code. The problem is that there are some parts of your code where side effects can kill you. In concurrent programs uncontrolled side effects lead to data races. In Software Transactional Memory (STM, which at some point may become part of C++) side effects may be re-run multiple times when a transaction is retried. There is an urgent need to control side effects and to separate pure functions from their impure brethren. Encapsulating side effects inside monads could be the ticket to extend the usefulness of pure functions inside an imperative language like C++.

To summarize: A monad is an applicative functor with an additional ability, which can be expressed either as a way of flattening a doubly encapsulated object, or as a way of applying a functor factory to an encapsulated object.

In the range library, the first method is implemented through view::flatten, and the second through view::for_each. Being an applicative functor means that a range can be manipulated using view::transform and that any value may be encapsulated using view::single or, inside for_each, using yield.

The ability to apply a range of functions to a range of arguments that is characteristic of an applicative functor falls out of the monadic functionality. For instance, the example from the previous section can be rewritten as:

for_each(ints(10, 11), [](int i) {
  return for_each(ints(1, 3), [i](int j) {
    return yield(std::make_pair(i, j));

The Mess We’re In

I don’t think the ideas I presented here are particularly difficult. What might be confusing though is the many names that are used to describe the same thing. There is a tendency in imperative (and some functional) languages to come up with cute names for identical patterns specialized to different applications. It is also believed that programmers would be scared by terms taken from mathematics. Personally, I think that’s silly. A monad by any other name would smell as sweet, but we wouldn’t be able to communicate about them as easily. Here’s a sampling of various names used in relation to concepts I talked about:

  1. Functor: fmap, transform, Select (LINQ)
  2. Pointed functor: pure, return, single, repeat, make_ready_future, yield, await
  3. Applicative functor: <*>, zip_with
  4. Monad: >>=, bind, mbind, for_each, next, then, SelectMany (LINQ)

Part of the problem is the lack of expressive power in C++ to unite such diverse phenomena as ranges and futures. Unfortunately, the absence of unifying ideas adds to the already overwhelming complexity of the language and its libraries. The functional paradigm could be a force capable of building connections between seemingly distant application areas.


I’m grateful to Eric Niebler for reviewing the draft of this blog and correcting several mistakes. The remaining mistakes are all mine.

Can a data race not be a bug? In the strictest sense I would say it’s always a bug. A correct program written in a high-level language should run the same way on every processor present, past, and future. But there is no proscription, or even a convention, about what a processor should (or shouldn’t) do when it encounters a race. This is usually described in higher-level language specs by the ominous phrase: “undefined behavior.” A data race could legitimately reprogram your BIOS, wipe out your disk, and stop the processor’s fan causing a multi-core meltdown.

Data race: Multiple threads accessing the same memory location without intervening synchronization, with at least one thread performing a write.

However, if your program is only designed to run on a particular family of processors, say the x86, you might allow certain types of data races for the sake of performance. And as your program matures, i.e., goes through many cycles of testing and debugging, the proportion of buggy races to benign races keeps decreasing. This becomes a real problem if you are using a data-race detection tool that cannot distinguish between the two. You get swamped by false positives.

Microsoft Research encountered and dealt with this problem when running their race detector called DataCollider on the Windows kernel (see Bibliography). Their program found 25 actual bugs, and almost an order of magnitude more benign data races. I’ll summarize their methodology and discuss their findings about benign data races.

Data Races in the Windows Kernel

The idea of the program is very simple. Put a hardware breakpoint on a shared memory access and wait for one of the threads to stumble upon it. This is a code breakpoint, which is triggered when the particular code location is executed. The x86 also supports another kind of a breakpoint, which is called a data breakpoint, triggered when the program accesses a specific memory location. So when a thread hits the code breakpoint, DataCollider installs a data breakpoint at the location the thread was just accessing. It then stalls the current thread and lets all other threads run. If any one of them hits the data breakpoint, it’s a race (as long as one of the accesses is a write). Consider this: If there was any synchronization (say, a lock acquisition was attempted) between the two accesses, the second thread would have been blocked from accessing that location. Since it wasn’t, it’s a classic data race.

Notice that this method might not catch all data races, but it doesn’t produce false positives. Except, of course, when the race is considered benign.

There are other interesting details of the algorithm. One is the choice of code locations for installing breakpoints. DataCollider first analyzes the program’s assembly code to create a pool of memory accesses. It discards all thread-local accesses and explicitly synchronized instructions (for instance, the ones with the LOCK prefix). It then randomly picks locations for breakpoints from this pool. Notice that rarely executed paths are as likely to be sampled as the frequently executed ones. This is important because data races, like fugitive criminals, often hide in less frequented places.

Pruning Benign Races

90% of data races caught by DataCollider in the Windows kernel were benign. For several reasons it’s hard to say how general this result is. First, the kernel had already been tested and debugged for some time, so many low-hanging concurrency bugs have been picked. Second, operating system kernels are highly optimized for a particular processor and might use all kinds of tricks to improve performance. Finally, kernels often use unusual synchronization strategies. Still, it’s interesting to see what shape benign data races take.

It turns out that half of all false positives came from lossy counters. There are many places where statistics are gathered: counting certain kinds of events, either for reporting or for performance enhancements. In those situations losing a few increments is of no relevance. However not all counters are lossy and, for instance, a data race in reference counting is a serious bug. DataCollider uses simple heuristic to detect lossy counters–they are the ones that are always incremented. A reference counter, on the other hand, is as often incremented as decremented.

Another benign race happens when one thread reads a particular bit in a bitfield while another thread updates another bit. A bit update is a read-modify-write (RMW) sequence: The thread reads the previous value of the bitfield, modifies one bit, and writes the whole bitfield back. Other bits are overwritten in the process too, but their new values are the same as the old values. A read from another thread of any of the the non-changed bits does not interfere with the write, at least not on the x86. Of course if yet another thread modified one of those bits, it would be a real bug, and it would be caught separately. The pruning of this type of race requires analysis of surrounding code (looking for the masking of other bits).

Windows kernel also has some special variables that are racy by design–current time is one such example. DataCollider has these locations hard-coded and automatically prunes them away.

There are benign races that are hard to prune automatically, and those are left for manual pruning (in fact, DataCollider reports all races, it just de-emphasizes the ones it considers benign). One of them is the double-checked locking pattern (DCLP), where a thread makes a non-synchronized read to be later re-confirmed under the lock. This pattern happens to work on the x86, although it definitely isn’t portable.

Finally, there is the interesting case of idempotent writes— two racing writes that happen to write the same value to the same location. Even though such scenarios are easy to prune, the implementers of DataCollider decided not to prune them because more often than not they led to the uncovering of concurrency bugs. Below is a table that summarizes various cases.

Benign race Differentiate from Pruned?
Lossy counter Reference counting Yes
Read and write of different bits Read and write of the whole word Yes
Deliberately racy variables Yes
Idempotent writes No


In the ideal world there would be no data races. But a concurrency bug detector must take into account the existence of benign data races. In the early stages of product testing the majority of detected races are real bugs. It’s only when chasing the most elusive of concurrency bugs that it becomes important to weed out benign races. But it’s the elusive ones that bite the hardest.


  1. John Erickson, Madanlal Musuvathi, Sebastian Burckhardt, Kirk Olynyk, Effective Data-Race Detection for the Kernel