Programming through the lens of mathematics offers new ways to solve old problems and old ways to solve new problems. Mathematics is full of surprising relations and dualities.

In this blog post, we’ll journey through an alternate reality in which an OOP language like C# becomes more abstract and expressive by adopting some functional programming concepts, rooted in a branch of mathematics called Category Theory.

Over The Rainbow

Our journey begins with a few familiar types in a probably unfamiliar categorization.

Don’t worry if you don’t understand these concepts yet, we’ll meet them again later...

IEnumerable<T> is the list monad.

IObservable<T> is the continuation monad.

Task<T> is the continuation comonad.

Lazy<T> is the singleton comonad (for lack of a better name).

These types appear to have different purposes, yet there is actually a pattern hiding in there somewhere. The pattern can be stated in structural terms as follows.

Monads have bind and unit functions.

Comonads have cobind and extract functions.

You may have noticed that Lazy<T> is incomplete. Can we define a cobind function for Lazy<T>? What would it look like? What would it do?

And what’s the relationship between monads and comonads? Is there a pattern?

Perhaps if we were to examine these monadic and comonadic types through a lens of abstraction, then we could identify the pattern and figure out how to implement cobind for Lazy<T>.

“Professor Marvel Never Guesses”

Let's start by examining the bind function for IEnumerable<T>. Here’s its signature:

public static IEnumerable<R> SelectMany<T, R>(this IEnumerable<T> source, Func<T, IEnumerable<R>> selector)

So the bind function takes an instance of the monadic type for which it's defined, in this case IEnumerable<T>, and a function that projects the element type, T, into another instance of the monadic type, IEnumerable<R>.

Its purpose may not be clear yet, so we’ll just try and implement it by heeding Erik Meijer’s advice: “let the types guide you”.

Well, we have a selector function that accepts a T, but we don’t have an instance of T yet. Since the source is a sequence, it seems that we must invoke the selector function for each T that it contains. So let's start there. We’ll do foreach over source and apply selector to each value.

public static IEnumerable<R> SelectMany<T, R>(this IEnumerable<T> source, Func<T, IEnumerable<R>> selector)
{
    foreach (var value in source)
    {
        var results = selector(value);

        // ...
    }
}

Great, we're done with the Select (a.k.a., map) part. Now let’s check the return value.

The selector returns an IEnumerable<R> for a given T, and we need to return a single IEnumerable<R>. We can’t simply return the IEnumerable<R> for the first T, because we must return them all. But how do we convert many IEnumerable<R> into a single IEnumerable<R>?

How about we iterate each of the sequences and yield all of the results?

(An Iterator Block makes this easy in C# – it spares us from having to implement IEnumerable<T> and IEnumerator<T> ourselves.)

public static IEnumerable<R> SelectMany<T, R>(this IEnumerable<T> source, Func<T, IEnumerable<R>> selector)
{
    foreach (var value in source)
        foreach (var innerValue in selector(value))
            yield return innerValue;
}

And now we're done with the Many (a.k.a., flat) part.

We've just implemented the bind function in the list monad by iterating over the source sequence, mapping each element into a nested sequence, and flattening all of the nested sequences into the output sequence. The name flatMap makes perfect sense! The name SelectMany makes sense too, although it’s not a typical naming convention.¹

The significance of this implementation may not be clear yet, but one thing is certain: it was easy to derive this implementation by following the types.

It’s A Twister! It’s A Twister!

Shouldn’t all monads have similar signatures for their SelectMany implementations?

Let's look at bind for IObservable<T>. And for comparison, IEnumerable<T> as well.

public static IObservable<R> SelectMany<T, R>(this IObservable<T> source, Func<T, IObservable<R>> selector)

public static IEnumerable<R> SelectMany<T, R>(this IEnumerable<T> source, Func<T, IEnumerable<R>> selector)

Okay, so there's clearly a pattern here. It’s the same exact signature, just with different generic types.

How can we implement bind for IObservable<T>?

Since the source is a sequence, it seems that we must invoke the selector function for each T that it contains. And there's only one way to get T's out of an observable: Subscribe

public static IObservable<R> SelectMany<T, R>(this IObservable<T> source, Func<T, IObservable<R>> selector)
    => source.Subscribe(value =>
       {
           var results = selector(value);

           // ...
       });

We're done with the Select (a.k.a., map) part. Now let’s check the return value.

The selector returns an IObservable<R> for a given T, and we need to return a single IObservable<R>. Again, we’re in a situation similar to IEnumerable<T> where we have many instances of the monadic type yet the return type requires a single instance.

To convert many IObservable<T> into a single IObservable<T>, how about we Subscribe to all of the sequences and push all of their values to the observer? It’s a similar concept to the inner foreach that flattens the IEnumerable<T>.

public static IObservable<R> SelectMany<T, R>(this IObservable<T> source, Func<T, IObservable<R>> selector)
    => source.Subscribe(value => selector(value).Subscribe( ? ));

Uh oh, from where do we get an observer to pass to Subscribe? And how do we return an IObservable<R> from this expression? Subscribe returns IDisposable, so that's not helpful.

What we need is an implementation of IObservable<T> that we can return from SelectMany, so that when Subscribe is called with an IObserver<T>, that observer is available to our expression. Then we can pass the observer to the projected observables' subscriptions.

If this seems backwards, it’s because it is. The reactive IObservable<T> interface is dual to the interactive IEnumerable<T> interface. They are structural opposites, so we must think oppositely when solving structural problems like this one.

IEnumerable<T> is synchronous because the caller of IEnumerator<T>.MoveNext blocks while the next value is computed. IObservable<T> is asynchronous because after Subscribe is called with an IObserver<T>, the caller is not necessarily blocked and may continue executing. OnNext is called when values are computed in the future.

Using the same observer to subscribe to each of the projected sequences is similar to yielding from all of the projected enumerables in our IEnumerable<T> implementation above; however, we don't have a language integrated feature like Iterator Blocks, so we’ll have to use Rx's Observable.Create function instead. The Observable.Create function takes a function that has an IObserver<T> argument, which we’ll use to make the inner subscriptions.

public static IObservable<R> SelectMany<T, R>(this IObservable<T> source, Func<T, IObservable<R>> selector)
    => Observable.Create<R>(observer => source.Subscribe(value => selector(value).Subscribe(observer)));

And now we’re done with the Many (a.k.a., flat) part.

Note: I’ve avoided addressing the following issues in the code example above for the sake of simplicity – the focus here is on the essence of the monadic operations, not the implementation details of how to properly implement operators.

1. This implementation violates one of Rx's contracts: Serialized notifications per observer (§4.2 in the Rx Design Guidelines). If any of the inner observables push notifications on different threads, then a downstream observer may receive overlapping notifications, which is very bad. I'm not including a proper implementation in this blog post, but we can imagine that synchronizing calls to OnNext within a lock would suffice.
2. This implementation fails to properly support cancellation since we’re ignoring the IDisposable object that Subscribe returns. A proper implementation must hold the disposable for each inner observable, until it terminates. Furthermore, the inner disposables must be tied to the lifetime of the outer disposable, such that if the outer disposable is disposed, then all of the inner subscriptions are cancelled as well.
3. This implementation fails to properly handle OnError and OnCompleted notifications from the source observable. Furthermore, it improperly handles OnCompleted notifications from the inner observables by completing the sequence as soon as one of the inner observables completes, rather than completing when all observables, including the source, have completed.

The implementation of bind for IObservable<T>, like bind for IEnumerable<T>, is a flattening and a mapping; however, due to the natural asynchrony of the IObservable<T> interface, bind provides an asynchronous merging rather than a synchronous flattening.

In other words, the SelectMany operator for IObservable<T> is actually Select + Merge, whereas SelectMany for IEnumerable<T> is actually Select + Concat.

The bind operator for IEnumerable<T> accepts a sequence as input. For each value, an inner sequence is projected and iterated. Due to the synchronous nature of IEnumerable<T>, only one inner sequence can be iterated at a time. The bind operator will not yield any values from subsequent projections until it has completely iterated over the current inner sequence.

Conversely, the bind operator for IObservable<T> merges all inner subscriptions together, such that when any inner subscription produces a value, it’s immediately pushed downstream.

In both implementations of bind, the inner sequences are being flattened down into a single output sequence.

We’re Not In Kansas Anymore

Language Integrated Query (LINQ) introduces the monad as a first-class citizen in C#. Query Comprehension Syntax is a syntax for monads.²

They are monads because, other than satisfying the monad laws, each monad is a generic type with a specific purpose, paired with a bind function that implements sequential composition over that type, while preserving its purpose.

We can understand how bind represents sequential composition by following this logic:

1. The bind function SelectMany(M<T> source, Func<T, M<R>> selector)has two parameters.
2. The source parameter is of the type M<T>. Think of M as the monadic type, which contains values of type T. Monadic types are generic types. (This is what is meant by “amplification” or “embellishment”.)
3. The selector function projects a value of type T into M<R>, which is the monadic type M containing values of type R.
4. M<T> represents a computation of values of type T. More precisely, M<T> represents an antecedent computation, because a value of T must be computed before it can be passed to selector.
5. The selector function represents a subsequent computation, because it accepts a value of type T and returns M<R>. And just like M<T>, M<R> represents a computation of values of type R.
6. Both M<T> and M<R> are of the monadic type M, although they may contain different types. M<R> can be used as the source parameter in a subsequent bind operation.
7. The bind operator can be chained together across may projections over the monadic type M, such that M<T> goes to M<R> goes to M<Q>, and so on. Each projection through bind is a linear step in a larger computation.
8. Therefore, the bind function sequentially composes antecedent computations with subsequent computations, while preserving the purpose of the monadic type.
9. However, the cardinality of M may be greater than 1, depending on the actual type of M. Containing more than one T means having more than one M<R> because each T will be projected into an instance of M<R>.
10. The bind operator flattens M<M<R>> into a single M<R> to ensure that its result is a single container of the monadic type M. Flattening preserves sequential composition regardless of the cardinality of M.

For example, IEnumerable<T> represents a lazily-computed, synchronous sequence of values of type T. That’s its purpose, which SelectMany must preserve. As the source for SelectMany, IEnumerable<T> represents an antecedent computation of values of type T, because each T must be computed before it can be projected. SelectMany iterates the source, invoking the selector function for each value of type T, and projecting it into IEnumerable<R>. IEnumerable<R> represents a subsequent computation, because it can be subsequently iterated to generate values of type R. To continue the computation sequentially, IEnumerable<R> can be used as the source for another call to SelectMany. Therefore, SelectMany implements sequential composition.

IObservable<T> represents a lazily-computed, asynchronous sequence of values of type T. It’s similar to IEnumerable<T>, except asynchronous. And how do we consume an asynchronous sequence? Not by foreach, but by passing in a continuation (IObserver<T>). The SelectMany operator for IObservable<T> subscribes to the source with its own IObserver<T>. Sequential composition is implemented by awaiting values pushed into the IObserver<T> and then pushing them as input into a subsequent computation, represented by IObservable<R>. As far as sequential composition goes, it doesn’t actually matter that IObservable<T> is asynchronous, because going from T to R remains sequential.

IObservable<T> is asynchronous because in order to get data out of it, you must pass in a function (a continuation) to be invoked later. It’s the so-called, “continuation passing style”, or CPS for short. Thus, IObservable<T> is the continuation monad.

The bind operator enables us to write sequential composition in a declarative style. SelectMany sequentially composes two parts of our query, while hiding the imperative- or CPS-style mechanisms that are required to read values from the monadic type. We can think of a monadic type as a container of values, and bind lets us compose them without having to explicitly specify how to extract their values. We don’t have to use foreach for enumerables or call Subscribe for observables, until we want to leave the monad. That’s why we use foreach and Subscribe only at the ends of our LINQ queries.

This pattern works with more than just sequences. There are many types of monads; e.g., the “Maybe” monad implements null propagation, as Wes Dyer describes here, while introducing monads. There’s a State monad that represents mutability in a pure functional language like Haskell, an I/O monad that represents input and output, also in Haskell, there’s the Reader monad for computing within an environment, similar to having global or static fields (you guessed it: Haskell), among others.

Monads provide a declarative programming model, but note the relation to imperative-style code. Two consecutive from clauses in a LINQ query is compiled into a single call to SelectMany. It’s like having a semicolon at the end of each line. The first from expression is evaluated, and then the second from expression is evaluated, with the value of the previous expression still in scope.

from x in Xs       //;
from y in getYs(x) //;
select (x, y)

This query looks similar to our implementation of SelectMany for IEnumerable<T>, which consisted of nested foreach loops! And in fact it works exactly the same, because this query simply compiles into a call to SelectMany.

A query is a single expression pieced together by applying a sequence of operators, rather than a series of statements that are pieced together by a sequence of semicolons. It’s the composability of the monad that enables us to write a query as a single expression, and it’s the expression syntax that makes the declarative programming style appear similar to the imperative style.

This style of programming is declarative because we can use operators like where, join and group by without sacrificing the readability of the program with all of those foreach loops that we’d need to write if we weren’t using the list monad, or all of those continuations that we’d need to write if we weren’t using the continuation monad.

Essentially, monads enable us to declare what we want the program to do rather than how the program should do it. The “how” part is hidden by the monad, and it differs between monads.

from x in Xs
where x != null
from y in getYs(x)
join z in Zs on y.P equals z.P
select (x, y, z)

Is the above an interactive query over IEnumerable<T> or a reactive query over IObservable<T>?

Or maybe it’s a hybrid? Perhaps Xs is an IObservable<T> and getYs returns an IEnumerable<T>?

Each is a valid possibility. There’s no way of knowing by looking at the query alone!

The monadic pattern is a generalization on composition with embellishments, so it’s no surprise that queries over different types can look the same; however, some operators don’t make sense in some monads. For example, Rx defines the Sample(TimeSpan) operator for IObservable<T> to sample elements by time, which only makes sense because IObservable<T> is asynchronous. Sampling by time doesn’t make sense in IEnumerable<T> because it’s synchronous; however, enumerables do have the ability to sample elements by data; i.e., the Where operator. We can also sample elements by data in IObservable<T>, thus Rx provides a similar Where operator.

Essentially, a monad consists of an embellishing type and a function, bind, that preserves the embellishment through sequential composition.

A monad must also have a unit function, which enters the monad. For example, Rx provides Observable.Return (a.k.a., just).³

Technically, a monad must also satisfy some simple mathematical laws: left identity, right identity and associativity. I won’t go into them here.

Perhaps what makes monads seem mysterious and confusing is how such a simplistic functional trick can be so powerful as to warrant an entirely new syntax in languages like Haskell and C#. Are monads confusing because they mysteriously turn complex imperative-style code into a neat declarative-style code that looks eerily similar to imperative style? I’m confused even writing that question.

A comprehension syntax isn’t technically required, it just makes monads nicer to use! C# doesn’t have keywords for every useful operator, so most of the queries we write are in the form of fluent-method call syntax, whereby we chain operators together using dot notation and pass lambda expressions as function arguments.

For example, here’s the same query again written in fluent-method call syntax:

Xs.Where(x => x != null)
  .SelectMany(x => getYs(x).Select(y => (x: x, y: y)))
  .Join(Zs, xy => xy.y.P, z => z.P, (xy, z) => (x: xy.x, y: xy.y, z: z));

It’s not great, although it’s probably how most languages use monads. Regardless, both syntaxes (comprehensions and fluent-method call) form declarative expressions. The comprehension syntax is often nicer though because it hides all of those lambdas and automatically broadens the scope of query variables; e.g., in the query comprehension written earlier, x is still in scope after join. The C# compiler accomplishes this simply by changing the type of T to a compiler-generated type that carries all of the variables in scope. We must do the same trick ourselves when writing in the fluent-method call syntax if we want to keep earlier values in scope. We typically do that by projecting anonymous types or tuples throughout our query, as I’ve done in the example above.

One last interesting thing to note about monads…

Because of its declarative nature, applying bind forms a lazily-evaluated expression, just like monads in lazily-evaluated functional programming languages. SelectMany and other operators do not cause any side effects when they are applied. They simply return a new instance of the monad, so that we may continue applying operators to build our computation. When we are ready for side effects, then we leave the monad and the computation may begin. How it begins depends on the purpose of the monadic type.

IEnumerable<T> requires code to pull each value through the composition, thus we use foreach.

IObservable<T> pushes values through the composition, once we’ve called Subscribe.

And that’s the mysterious monad. You can get deeper into the concept of monads in Category Theory, but I think that the essence of why they’re useful to us as programmers is described above.

All good, but monads are only half the story!

If I Only Had a Codomain

According to my chart at the beginning of this post, Task<T> is a comonad. Why?

If you don’t believe me, then we’ll try and disprove that claim by implementing bind for Task<T>. If we can implement bind, then it’s probably a monad, right?

We already know how the signature for bind must look based on the pattern that we’ve identified above for enumerables and observables. They each have similar signatures. To go from the enumerable bind to the observable bind we must only do a monadic type replacement. So let’s grab the bind signature for IEnumerable<T> or IObservable<T> and simply replace the monadic type with Task<T>:

public static Task<R> SelectMany<T, R>(this Task<T> source, Func<T, Task<R>> selector)

Looks good, no problems so far.

Since the source is a Task<T>, it seems that we must invoke the selector function with the T that it contains.

public static Task<R> SelectMany<T, R>(this Task<T> source, Func<T, Task<R>> selector)
    => selector(source.Result);

Okay, that compiles, but it’s totally incorrect! Reading the Result property may block, and blocking isn’t preserving the purpose of Task<T>. The purpose of Task<T> is to represent the asynchronous computation of T. The SelectMany function must allow us to sequentially compose asynchronous computations, which means that we must be able to project a Task<T> into a Task<R> without blocking to read T. Essentially, SelectMany must return a lazily-evaluated expression, just like the implementations for IEnumerable<T> and IObservabe<T>.

Great, so how do we get the T out of the Task<T> asynchronously?

For the C# devs out there, our instinct is to define SelectMany as an async method and then await source. And that may seem appropriate, but clearly await must be doing something under the covers in order to get the T from the Task<T> without blocking. How does it do it? Whatever the C# compiler is doing for await, we must be able to do ourselves…

Well, the only way to get the result asynchronously is to pass a continuation to ContinueWith (a.k.a., then). But what is this strange method anyway? Let’s take a look at its signature (one overload in particular⁴).

public static Task<R> ContinueWith<T, R>(this Task<T> source, Func<Task<T>, R> selector)

Hmm, now that method looks familiar. Is it bind? Not quite. Here’s bind again:

public static Task<R> SelectMany<T, R>(this Task<T> source, Func<T, Task<R>> selector)

The type arguments in the selector function are reversed! Well, they are sort of reversed. The T and R type arguments remain in the same positions, but the generic Task type is in the opposite position. The selector function for ContinueWith has a Task parameter, rather than returning a Task.

So what is the selector in ContinueWith supposed to do?

Instead of being given a value and returning a new, embellished value, it’s given an embellished value and it has to return a new, unembellished value. In other words, it has to extract the value from the comonadic type inside the projection! This makes sense if we think of comonads (and monads) as containers in general.

And that’s how to safely (i.e., in a non-blocking manner) extract a value from Task<T>. We must read the Result property only from within the selector (continuation) supplied to ContinueWith, in order to preserve the purpose of Task<T>.

public static Task<R> SelectMany<T, R>(this Task<T> source, Func<T, Task<R>> selector)
    => source.ContinueWith(task => selector(task.Result)).Unwrap();

Unwrap is also required, because selector returns Task<R>, yet the selector in ContinueWith requires an R. So we make R a Task<R>, thus ContinueWith returns Task<Task<R>>. SelectMany requires Task<R>, so we must apply Unwrap one time to ditch the outer Task.

Recall that bind is a flattening projection (flat map; it’s like going from IEnumerable<IEnumerable<T>> to IEnumerable<T>), whereas here we’re using ContinueWith to compose over an extending projection (Task<R> becomes Task<Task<R>>). The concepts of flattening and extending are opposites. The bind and cobind functions are opposites; therefore, ContinueWith is cobind.

Here, Unwrap is merely the flattening of our bind implementation. It’s how we convert our cobind’s extending projection into bind’s flattening projection.

The Task<T>.Result property (technically, the get_Result method) is the extract function of the Task<T> comonad. The extract function is to comonad as the unit function is to monad. The unit function gets us into the monad, extract gets us out of the comonad. They are opposites!

And now we see why Task<T> is a comonad, and not a monad. Another way of looking at: we can only implement bind in terms of cobind; i.e., we can’t implement SelectMany for Task<T> without calling ContinueWith.

Other than the comonadic laws (analogous to the monad laws), comonads are a triple consisting of a generic type, a cobind function and an extract function. Comonads are the opposite of monads, which are a triple consisting of a generic type, a bind function and a unit function.

As you can see from the code above, it’s fairly easy to convert from cobind to bind. Therefore, we can convert comonads into monads. We can also convert monads into comonads, although I’m not showing it here, but you can imagine that one indicator of a type being monadic rather than comonadic is that when implementing cobind, you’d discover that you must implement it in terms of bind. Try it for yourself.

A major reason that we don’t use Task<T> as a monad is because its comonadic form lends itself naturally to a really useful compiler trick: coroutines. A coroutine is syntactic sugar that allows us to write imperative-style code for continuations, hiding the messy details of the underlying state machine that the compiler must generate. In C#, it’s known as async methods. The await keyword essentially compiles into a call to ContinueWith.

Although we can’t use this compiler trick with monads directly, we can convert a monad into a comonad, and that’s exactly what Rx’s ToTask extension method does. But even better than that, Rx offers an extension method that implements C#’s awaitable pattern for IObservable<T>, thus we can directly await an IObservable<T> to safely (i.e., in a non-blocking manner) extract the final value of the sequence. This works well for singleton observables, especially for aggregations like Sum and ToList.

Because C# has great support for async methods, it’s often recommended to use Task<T> rather than IObservable<T> when your computation only has a single result; however, it appears that for a future release of Rx, Bart De Smet has implemented the new awaitable pattern in C# 7. We’ll be able to define async methods that directly return an observable rather than a Task<T>. This is certainly not a replacement for Task<T> in all async methods, but it may come in handy. To return an observable from an async method, we’ll have to use a new interface, ITaskObservable<T>, which derives from IObservable<T>. It’s just a small grammatical difference, perhaps. One neat thing about ITaskObservable<T> is that it has singleton semantics; i.e., it only calls OnNext once, similar to the fact that Task<T> can only contain a single Result.

Along The Yellow Brick Road…

At the beginning of our journey we had asked the following questions:

“Can we create a cobind function for Lazy<T>? What would it look like? What would it do?”

Along the way we’ve met IEnumerable<T>, IObservable<T> and Task<T>. Now we know how to proceed.

We’ll take the signature for Task<T>’s cobind function and replace all Task<T> with Lazy<T>.

public static Lazy<R> ContinueWith<T, R>(this Lazy<T> source, Func<Lazy<T>, R> selector)

We have a Lazy<T> and a function that requires a Lazy<T>, so should we apply it? (#WWEMD⁵)

public static Lazy<R> ContinueWith<T, R>(this Lazy<T> source, Func<Lazy<T>, R> selector)
    => selector(source);

Sure, but what about the return value now? We have an R but we require a Lazy<R>.

That’s simple. Let’s make the whole expression lazy by wrapping it in a new Lazy<R>. That makes sense because we must preserve the purpose of Lazy<T>, which is lazy initialization, right?

public static Lazy<R> ContinueWith<T, R>(this Lazy<T> source, Func<Lazy<T>, R> selector)
    => new Lazy<R>(() => selector(source));

When reading Lazy<R>.Value, the selector will first extract the value of Lazy<T>, and with it compute R. In effect, Lazy<T> isn’t evaluated until we attempt to evaluate Lazy<R>, because Lazy<R> depends on Lazy<T>.

Thus, laziness is preserved through composition.

How might our code look when using this ContinueWith implementation for Lazy<T>?

Let’s first define a simple function named Create to take advantage of type inference. It’s just a bit more declarative than having to construct a Lazy<T> ourselves and specify T explicitly.

public static class Lazy
{
    public static Lazy<T> Create<T>(Func<T> valueFactory)
        => new Lazy<T>(valueFactory);
}

Here’s a simple program that composes two lazy computations:

var i = Lazy.Create(() => 
{
    Console.WriteLine("Initializing ‘i’");
    return 42;
});

var d = i.ContinueWith(lazy => lazy.Value * 2);

Notice that the construction of Lazy<T> assigned to d is hidden by the cobind function. This is the declarative style of programming, and it appears the same in comonads as in monads.

Using i and d as defined above, what does the following console program print?

Console.WriteLine(i.IsValueCreated);
Console.WriteLine(d.IsValueCreated);
 
Console.WriteLine(d.Value);

Console.WriteLine(i.IsValueCreated);
Console.WriteLine(d.IsValueCreated);

Console.WriteLine(i.Value);

Scroll down to see the output…

.

.

.

.

.

.

False
False
Initializing ‘i’
84
True
True
42

Notice that i.IsValueCreated returns False at first, even though d is defined in terms of i. Invoking the ContinueWith function doesn’t cause any side effects, just like how applying SelectMany doesn’t cause any side effects for monads. We can build up a computation with cobind and evaluate it later; i.e., lazy evaluation.

When the program reads d.Value, then i.Value is initialized. Afterwards, i.IsValueCreated returns True, even though the program hasn’t read i.Value directly yet. That’s because d reads i.Value in its continuation, so evaluating d requires i to be evaluated as well. Evaluating the outermost Lazy<T> propagates the act of evaluation up to the source.

Finally, when the program reads i.Value at the end, notice that it’s not reinitialized. Lazy<T> caches the result of its factory function.

We’re Off To See The Wizard

Where does laziness come from? Is it something in the water?

No, it’s something in the function.

A function provides an identifiable name for a sequence of statements in the imperative style. Imperative style programming languages are eager, in that when a program encounters a statement, it’s evaluated immediately. That’s what imperative means. Therefore, laziness is achieved simply by wrapping statements in a function. A function may be evaluated at any time by invoking it, by name or delegate.

A statement evaluates eagerly.

A statement in a function evaluates lazily.

A lambda expression represents an anonymous function with a body consisting of a single statement. We bind a lambda expression to a named variable or parameter, and that becomes its name for the purpose of invoking it. A lambda expression isn’t evaluated at the site at which it’s defined, but at the site at which it’s invoked. Therefore, lambda expressions are lazily-evaluated. They are functions.

In a pure functional language like Haskell, a function consists of expressions, and expressions are interchangeable with values.⁶ Replacing an expression with its value cannot be done safely in imperative-style languages because an expression may have side effects, and you’d expect those side effects to occur each time that the expression is evaluated. Evaluating a Haskell program is similar to reducing a mathematical expression into its simplest form, yet expressions in Haskell are, perhaps surprisingly, lazily evaluated. Evaluation is deferred until an expression’s value is actually needed by the program. And once an expression has been evaluated, its value may replace every instance of that expression in the program. Perhaps expressions in Haskell are similar to Lazy<T> in C#?

So how is Lazy<T> related to a function?

Well, the Lazy<T> constructor turns a parameterless function into a type

public Lazy(Func<T> valueFactory)

that has a property

public T Value { get; }

which is technically a parameterless function

public T get_Value()

that has the same exact signature as valueFactory, the Func<T> that was used to construct the Lazy<T> instance in the first place.

So can we shake off the Lazy<T> type and implement Func<T> as a comonad?

Let’s try it. We’ll start with the signature of ContinueWith by simply replacing all Lazy<T> with Func<T>.

public static Func<R> ContinueWith<T, R>(this Func<T> source, Func<Func<T>, R> selector)

That’s wonderfully weird and magical looking. And for our next trick, the types will guide us…

We have a Func<T>, and we have a function that takes a Func<T> as its argument.

Note: A function with a function parameter is known as a higher order function. In this case, we’ve got two higher order functions: ContinueWith and selector.

Functions are given data as arguments.

Higher order functions are given functions as data.

Let’s pass our source function to selector.

public static Func<R> ContinueWith<T, R>(this Func<T> source, Func<Func<T>, R> selector)
    => selector(source);

Just like ContinueWith for Lazy<T>, we notice that our return types don’t match. The selector function returns R, but we need a Func<R>.

We must wrap the entire expression in another function!

public static Func<R> ContinueWith<T, R>(this Func<T> source, Func<Func<T>, R> selector)
    => () => selector(source);

Well, that was easy. It looks the same as the implementation for Lazy<T>, but without a Lazy<R> wrapper. The outer function provides the laziness.

For comparison, here’s the Lazy<T> definition from above:

public static Lazy<R> ContinueWith<T, R>(this Lazy<T> source, Func<Lazy<T>, R> selector)
    => new Lazy<R>(() => selector(source));

Is the Func<T> comonad the same as the Lazy<T> comonad?

No, because Func<T> will always compute a value when you invoke it, whereas Lazy<T> only computes its value once. Furthermore, Func<T> may return a different value each time that it’s invoked (in a side-effecting language like C#, this is common.)

Therefore, Func<T> is lazy, but it may also cause side effects each time that you invoke it, whereas Lazy<T> may only cause side effects once.

Should we use the Func<T> comonad at all?

Probably not. It offers no real advantage compared to composing lambda expressions, although it does hide the closure (for better or worse). Perhaps Func<T> is the closure comonad? A blog post for another day.

Meet The Lazy Wizard

The laziness of Lazy<T> is related to the idea that applying co/bind forms a lazily-evaluated expression.

We can declare computations over IEnumerable<T> and IObservable<T> that are lazy; i.e., there are no computational side effects until we iterate or subscribe, respectively. And likewise for Task<T>, we can compose together many ContinueWith calls without causing any side effects; that is, until the Task<T> completes. There’s obviously a race condition here, but the point is that ContinueWith doesn’t block. ContinueWith forms a lazily-evaluated computation, although it’s possible that Task<T>.Result was already computed before ContinueWith was applied, which may result in the continuation executing synchronously.⁷

And so it goes for Lazy<T>. The purpose of Lazy<T> is simply lazy evaluation. That’s what ContinueWith needs to preserve across compositions; although, it’s possible that the lazy value may have already been computed by the time the program attempts to extract it. Just like Task<T>.Result, once it’s computed, we can keep reading Lazy<T>.Value over and over again without recomputing it. We can even build new computations on top. In the case of Lazy<T>, a new computation over a pre-computed value would simply reuse the pre-computed value as the source for the new computation, without having to reevaluate the original source; i.e., memoization of a parameterless function. This is illustrated in the code example above with the explicit call to i.Value at the end of the program: the side effect of writing to the console upon initializing i.Value had already occurred earlier in the program, and it’s not repeated again at the end.

All of these co/monadic types are lazy, but they also differ in behavior when it comes to laziness. The laziness of Lazy<T> and Task<T> isn’t permanent, it’s a one-time use. The laziness of IObservable<T> and IEnumerable<T> is potentially repeatable.

To put it another way: Lazy<T> and Task<T> cache their value to avoid duplicating side effects, whereas IObservable<T> and IEnumerable<T> may cause side effects each time that they are “invoked”. I smell a pattern.

Pulling Back The Curtain

Lurking in the shadows, as usual, are side effects. Being able to control side effects is a major benefit of having these co/monadic constructs in C#.

The point of Lazy<T> is to ensure that any side effects only occur once, when computing a single value of T. That’s why Lazy<T> is often used to represent a static singleton in C#. If computing the value causes side effects, then we don’t want to compute the value until it’s actually needed by the program. And once it’s been computed, we don’t want to repeat the computation. Of course, there are other reasons to have singletons, but using a static readonly field without Lazy<T> is generally good enough for most singletons. It’s only when we want to defer side effects that we use Lazy<T> instead.

A synchronous sequence of values is similar to Lazy<T> but with a larger cardinality: Lazy<T> only has a single Value, whereas IEnumerable<T> has many. But there’s another significant difference: IEnumerable<T> doesn’t necessarily cache its results. Each time that you iterate, you may get different values. Furthermore, an enumerable query is a computation that may have its own side effects regardless of whether the source IEnumerable<T> caches its results or not, but those side effects aren’t realized until you iterate it.

Task<T> represents an asynchronous function that eventually computes a Result. We use Task<T> to avoid duplicating the side effects of an asynchronous computation. Once it completes, subsequent reads of the Result property will always return the same value, thus side effects won’t occur again. If we actually want to repeat the side effects, then we must invoke the async function again, which will return a new instance of Task<T>.

An asynchronous sequence of values is similar to Task<T> but with a larger cardinality: Task<T> only has a single Result, whereas IObservable<T> has many. But there’s another significant difference: IObservable<T> doesn’t necessarily cache its results. Each time that you Subscribe, you may get different notifications. Furthermore, an observable query is a computation that may have its own side effects regardless of whether the source IObservable<T> caches its notifications or not, but those side effects aren’t realized until you Subscribe.

Side effects are related to temperature:

Cold means that side effects may occur.

Hot means that side effects do not occur.

So how does the laziness of co/monads relate to temperature?

Let’s consider the properties of the types of the co/monads separately from their co/bind implementations.

The Func<T> type may return a different value each time that it’s invoked, and that’s the cold programming model. Therefore, functions are cold.

The Func<T> comonad described above defers side effects until the outer-most function is invoked. That’s also the cold programming model, and essentially its purpose. The Func<T> comonad composes functions lazily. But it’s also pointless, given that Func<T> is already cold. The Func<T> comonad simply adds an unnecessary layer of cold indirection.

The Lazy<T> type is initially cold, because it defers side effects until its Value property is read, yet it becomes hot for subsequent reads of the Value property. In other words, the first time that Value is read it may cause side effects, but subsequent reads definitely will not. (There’s also a race condition if Value may be read by multiple threads simultaneously, but that's similar to Task<T>.Result, so let’s address the race condition in our analysis of Task<T> that follows.)

The Lazy<T> comonad is initially cold, because cobind defers side effects until the outer-most Value property is read, yet it becomes hot for subsequent reads of the outer-most Value property.

The Task<T> type is initially cold, because it can be constructed by passing in a function and it defers side effects until its Start method is called; however, it can only be executed once. Calling Start again throws an exception. Furthermore, the Task<T>.Result property is always hot; therefore, Task<T> transitions to hot when Start is called.

However, it’s rare to acquire a Task<T> that hasn’t already been started. Typically we use Task.Run, Task.Factory.StartNew, or invoke some BCL or user-defined function that returns a Task<T>. In these cases, Start has already been invoked, thus the computation may have already begun even before the caller acquires a reference to the Task<T>. The caller can’t invoke Start again, and reading the Result property never causes side effects, thus our typical usage of Task<T> is hot, yet the Task<T> type itself is technically cold. In other words, notice that we typically acquire an instance of Task<T> by invoking a function that returns it. Functions are cold, as stated previously. Therefore, each time that we invoke a function that performs an asynchronous computation, it may return a new instance of Task<T> to represent its Result, which is always hot. Invoking the function is essentially like invoking the Start method yourself. In that case, Task<T> is hot because it has already been started by the function that returned it, yet the Task<T> type itself is technically cold because it must be started at some point.

Alternatively, there’s another way to get an instance of Task<T> (a.k.a., “future”): use a TaskCompletionSource<T> (a.k.a., “promise”). TaskCompletionSource<T> replaces the need to call Start by essentially making the Result property assignable. The Task<T> that it returns is hot because it’s constructed without a function (via an internal constructor), thus it has no computation to be deferred. A consumer of the Task<T> is unable to call Start, thus it’s unable to cause side effects, and so this usage of Task<T> is hot. This doesn’t change the fact that the Task<T> type itself is cold, at least as far as its public API is concerned. The promise mechanism is useful because it decouples the two features of a future: It separates its ability to defer a computation from its ability to compute asynchronously; i.e., TaskCompletionSource<T> makes Task<T> hot without having to begin a computation by calling Start. Instead, the computation occurs externally and terminates when Result is assigned (or the task faults, or it’s canceled – but we’ll ignore these states for the purposes of this analysis).

The Task<T> comonad is initially cold, because cobind defers side effects until the Task<T> completes. It’s cold because the continuation is a function, and functions are cold. However, it becomes hot once the Task<T> completes. If you apply ContinueWith after the Task<T> has completed, then it may execute the continuation synchronously. There’s a race condition.

Therefore, the Lazy<T> and Task<T> types have the same temperature and similar behavior; i.e., they are both initially cold and then transition to hot by caching their value, which avoids duplicating side effects. Lazy<T>.Value blocks until its computation has completed; likewise, Task<T>.Result blocks until its computation has completed. The difference between them is that Lazy<T> computes its Value synchronously, and Task<T> computes its Result asynchronously. This means that to extract the Value from Lazy<T>, a program must block. To extract the Result from Task<T>, a program can do other things while Result is being computed. Since the only purpose of Lazy<T> is to defer its computation of Value, it doesn’t make sense to have a method like Run; it would simply block until Value has been computed, thus defeating the purpose of Lazy<T>. Since the purpose of Task<T> is asynchrony, in addition to deferring its computation of Result, it’s actually useful to hide its coldness behind a function that returns a hot Task<T>.

As comonads, both Lazy<T> and Task<T> are initially cold, yet they become hot once their value has been computed. In the case of Lazy<T>, cobind is synchronous (it blocks until Value is computed), whereas for Task<T>, cobind is asynchronous (it doesn’t block, although it may execute the continuation synchronously when Result has already been computed and the ExecuteSynchronously flag was set). In either case, the transition from cold to hot means that there’s a race condition.

Therefore, comonads and their comonadic types are cold, which is the definition of deferred execution. In addition, comonads that transition from cold to hot have a race condition.

The IEnumerable<T> type is either cold, as in C# Iterator Blocks, or it’s hot, as in pre-computed lists and dictionaries.

The IEnumerable<T> monad is cold, because bind defers side effects until the enumerable that it returns is iterated. This is true even if the IEnumerable<T> that is the source of the computation is hot. Although no side effects will occur when iterating a hot enumerable such as a List<T>, the computation itself (a.k.a., the query) may cause side effects when it’s iterated, and it can only cause side effects when it’s iterated.

The IObservable<T> type is either cold, as in Rx generators like Return, Range, Generate, Timer, Interval and Create, or it’s hot, as in Subject<T>, events and hot Task<T> conversions.

The IObservable<T> monad is cold, because bind defers side effects until an observer subscribes to the observable that it returns. This is true even if the IObservable<T> that is the source of the computation is hot. Although no side effects will occur when subscribing to a hot observable such as a Subject<T>, the computation itself (a.k.a., the query) may cause side effects when an observer subscribes, and it can only cause side effects when an observer subscribes. Furthermore, due to the asynchrony of IObservable<T>, some interesting and useful combinations of cold and hot temperatures are also possible. (See Hot and Cold Observables for details.)

Therefore, the IEnumerable<T> and IObservable<T> types can either be cold or hot, and they have similar behaviors; i.e., they both represent sequences, which may or may not cause side effects when computing their elements. The difference between them is asynchrony. IEnumerable<T> computes its elements synchronously. IObservable<T> computes its elements asynchronously. This means that to pull the next value from an IEnumerable<T>, a program must block. To be pushed values from an IObservable<T>, a program calls Subscribe and can immediately do other things. In either case, iterating an IEnumerable<T> or subscribing to an IObservable<T> doesn’t necessarily cause side effects. It depends on the implementation; e.g., List<T> does not cause side effects, so it’s hot. An Iterator Block may cause side effects, to it’s cold. Observable.FromEventPattern returns an observable that wraps an event, which is hot. Observable.Range generates notifications each time that you subscribe, so it’s cold.

As monads, both IEnumerable<T> and IObservable<T> are initially cold. Applying bind on an IEnumerable<T> or IObservable<T> won’t cause any side effects until it’s iterated or subscribed, respectively. However, depending upon the implementation, they may or may not transition to hot.

Therefore, monads are cold regardless of the temperature of their types, which is the definition of deferred execution. In addition, monads that transition from cold to hot have a race condition.

Monads and comonads have identical temperatures with regard to composition – they are both cold. Therefore, consecutively applying co/monadic operators is free of side effects. Co/bind implements deferred execution, such that side effects won’t occur until you leave the co/monad.

There’s No Place Like Home

We use monads and comonads to control side effects. It’s the purpose of the property of laziness in co/bind. Co/monads allow us to write queries declaratively, without causing side effects. We control when side effects may occur by reading/invoking/iterating/subscribing at a later time.

Lazy<T> is a comonad with the purpose of causing side effects only once, attached to a blocking computation. The Func<T> comonad is even closer to the metal in that it represents repeatable side effects; although, perhaps it’s not very useful as a comonad.

Task<T> is a comonad with the purpose of causing side effects only once, attached to an asynchronous computation. However, the laziness of Task<T> is subtle. Due to a race condition that is common in the Task<T> implementation of cobind, we must either live with side effects possibly occurring immediately (although asynchronously in general), or we can build our computation on a promise (TaskCompletionSource) and glue it to the future (Task<T>) when the program is ready for side effects. One really nice way of doing that is to define a coroutine; a.k.a., an async method in C#. An async method let’s you await a Task within the imperative style of programming, meaning that typical control flow constructs are allowed without having to resort to implementing a complex state machine over CPS (because the compiler does it for you). Upon the first await, if the Task didn’t complete synchronously, then a promise is returned to the caller of the async method and the remainder of the computation from that point is referenced in a continuation that is attached to the Task via cobind. Side effects of the downstream computation can be controlled easily by simply holding the Task rather than applying await. The program can apply await at any time in the future to continue the computation with its Result. The race condition, in this case, is irrelevant.

IEnumerable<T> is a monad with the purpose of causing repeatable side effects, attached to a blocking computation.

IObservable<T> is a monad with the purpose of causing repeatable side effects, attached to an asynchronous computation.

Even though a sequence may be hot or cold, applying the bind operator (or any operator) itself doesn’t cause any side effects, until the sequence is iterated/subscribed. Furthermore, an operator may cause side effects each time that its return value is iterated/subscribed. The monadic bind function is about repeatable side effects, regardless of the temperature of its monadic type.

In conclusion:

Comonads represent lazy computations of single values, are initially cold, and transition to hot when the computation completes.

Monads represent lazy computations over zero or more values, depending on the type. They are initially cold, and may transition from cold to hot, depending on the type.

And why is any of this important?

Because a language like C# is full of side effects. Being able to introduce side effects so easily is often what makes C# really useful, but it’s also what makes many C# programs really confusing and buggy.

Controlling side effects in a declarative manner makes programs much easier to reason about, and to prove correctness. Although, some parts of a program are just much easier to write using imperative side effects.

Therefore, the careful combination of declarative and imperative styles within the same program is often the best choice. Functions generally help with the former, as in the procedural programming style. Monads also help with the former, computing over sequences in a declarative style. Comonads (via coroutines) generally help with the latter, when single-valued asynchronous computations are required.

These techniques are becoming even more important as modern programs increasingly depend on concurrency.

References

• The original Rx team’s videos and Erik Meijer’s papers – just search for this stuff online, it’s all great.
• Bartosz Milewski’s Blog and his YouTube channel on Category Theory
• Wes Dyer’s Blog – The Marvels of Monads
• Eric Lippert’s Blog – Monads

Notes

1. The name of SelectMany is similar to the name of the Select operator, which is the same as the SELECT operator in SQL. Likewise, the Where operator in LINQ is named the same as the WHERE operator in SQL, and so on. When LINQ was invented, apparently the designers had chosen SQL-like names for the operators to make developers more comfortable. Most developers were used to SQL, but didn't have a clue about monads. And this naming convention works of course because SQL is a monadic comprehension syntax!
2. The C# compiler actually looks for a special signature of a method named SelectMany rather than the one shown above. It must have an intermediate selector parameter in addition to the result selector. It's just a performance optimization and so I won’t elaborate on it here.
3. Rx actually provides several generators, including Range, Generate, Timer, Interval and Create. There are many ways to enter the continuation monad. Likewise, there are many ways to enter the list monad, IEnumerable<T>, because there are many types that implement IEnumerable<T>. And C# offers another option: Iterator Blocks.
4. The FCL implementation of ContinueWith is defined as an instance method. Here, I’ve rewritten it as an extension method and renamed its parameters, for parity with the SelectMany method. Other than those insignificant changes, it remains identical to the definition in the FCL.
5. What Would Erik Meijer Do?   https://twitter.com/headinthebox/status/543919077752729600
6. https://wiki.haskell.org/Referential_transparency
7. The TPL doesn’t invoke continuations synchronously unless you opt-in, and even then it’s not guaranteed. For the most part, the TPL is designed to introduce concurrency. This is in contrast to the design of Rx, which is free-threaded and rarely introduces concurrency. Operators introduce concurrency when you explicitly specify a concurrency-introducing scheduler, although operators that involve some kind of timing may default to a concurrency-introducing scheduler if you don’t pass in one; e.g., Throttle, Sample, Buffer, etc. This is because Rx is primarily about taming concurrency, not introducing it.

The T in IObservable<T> is more powerful than you may realize.

Structural T

T carries data, its most obvious purpose.

T can be any type, such as Int32, String or MyAmazingObject.

T can be void, as in System.Reactive.Unit.

T can be anonymous, as in:  new { Value = "T", Awesome = true }

T can be a choice, as in Either<TLeft, TRight>, which also seems to fix so-called "glitches".

T can be a group of T, as in IObservable<IGroupedObservable<K,T>>; e.g., GroupBy.

T might be there or it might not, as in Maybe<T>.

Meta T

T carries notification.  Observables notify observers by pushing T.

T can be notification reified, as in Notification<T>.

T can be collection change notifications, as in CollectionNotification<T>.

T can be directed mutations, as in CollectionModification<T>.

Temporal T

T carries time.  Observables push T asynchronously (though sequentially-§4.2).  Even pushing synchronously takes time though.

T can be a point in time; implicitly, the time at which it's observed, as in Amb or CombineLatest.

T can be time reified, as in Timestamped<T>, TimeInterval<T>, Recorded<T> or simply DateTimeOffset.

Temporal Window T

T carries duration.  A window is a duration of time that has a start and maybe an end.

T can start a window, as in Join, GroupJoin, Window or Buffer.

T can end a window, as in Join, GroupJoin, Window(closingSelector), Buffer(closingSelector), etc.

T can be a window, as in IObservable<IObservable<T>>.

T can be the end of time, as in that which precedes it, interpreted relative to the time of subscription (e.g., Timer), relative to a previous T (e.g., Interval or Throttle), or relative to some out-of-band time (e.g., Sample).

Channel T

T can be a channel or callback.  Observers invoke T to affect an observable.

T can be a delegate, as in Action or Func<TResult>.

T can be a subject, as in IObservable<ISubject<T>>.  This establishes a full duplex, asynchronous communication channel.  The observable is a singleton and both ends subscribe to the subject; e.g., a chat application. (See also this blog post.)

Future T

T can be the promise of a future T, as in IObservable<Task<T>>; although, it's somewhat redundant since IObservable<T> is also a promise of a future T and it has a wider cardinality [0,∞].  Task's cardinality is strictly [0,1].

Conclusion

T carries data, notification, time and duration.  It can also have communication channel or promise semantics.  Best of all, T carries any additional semantics that you choose.

So the next time that you find yourself stuck trying to figure out which Rx operator is the right one to solve your particular problem, pause and give T the attention that it deserves.  You may just find that the answer to your problem lies within how you make use of T.

So, have I missed any categorically interesting uses for T?  How do you use T?

That is the question!  There appears to be some confusion on the web about whether or not Subject<T> should be used, ever.  Tis a question oft asked by developers new to Rx after reading expert recommendations to avoid subjects, due to common misuse, yet subjects seem to persist an air of mystery even for seasoned reactive developers.

First, a primer on Subjects.  Take a moment to read that.

Done?  Great.  Let's continue...

As you may already know, Erik Meijer doesn't like subjects.  End of discussion, right?  Wrong!  Erik doesn't like subjects to be used in place of purely functional alternatives.¹

As Erik puts it:

[Subjects] are the "mutable variables" of the Rx world and in most cases you do not need them.

So let's enumerate the cases where we should and shouldn't use subjects.

We'll start by examining two interesting properties of subjects:

1. They implement IObservable<T> and IObserver<T> simultaneously, which makes them suitable for both output and input, respectively.
2. They are hot.  (See also: Hot and Cold Observables)

Thus to answer the question of whether we should use a subject when defining an observable (output) we must consider:

1. the input; i.e., the source of the OnNext, OnCompleted and OnError notifications.
2. the desired temperature of the generated observable; e.g., hot or cold.

To determine whether or not to use a subject, start by asking yourself the following questions:

1. what is the source of the notifications?  I.e., where do the values of T come from?  Who is responsible for calling OnNext?
2. must observers share subscription side-effects or not?  I.e., does calling Subscribe execute some kind of initialization code each time it's called or does initialization occur once and its state is shared by every call to Subscribe?

Let's answer all possible combinations of these 2 questions.  I'll provide more details and examples later in this post to explain why these answers make sense.

Question 1:  What is the source of the notifications?

The source can be one of two kinds: Let's call them external and local.  An external source is any observable or event that already exists outside of your code.  A local source is when you generate an observable from your code, kind of like defining and raising a custom .NET event.  More details on these distinctions later in this post.

1. If the source is external, then do not use a subject.
2. If the source is local, then do use a subject.

Question 2:  Must observers share subscription side-effects or not?

1. If not, then do not use a subject.  (Define a cold observable.)
2. If yes, then do use a subject.  (Define a hot observable.)

Let's enumerate all possible combinations: 

1. The source is external and I want a cold observable.
2. The source is external and I want a hot observable.
3. The source is local and I want a cold observable.
4. The source is local and I want a hot observable. 

As you may have noticed, there are conflicts.

• What if you have an external source and you do want to share subscription side-effects?
• What if you have a local source and you do not want to share subscription side-effects?
• What if the temperature of an external source differs from the desired temperature?

I believe these conflicts cause additional confusion among developers.  They seem to blur the line when it's appropriate or inappropriate to use subjects, though in fact the line is actually quite clear, as we'll soon see.

Let's resolve these conflicts now.

The source is external and I want a cold observable.

1. If the source is cold, then use it as is.  You can also wrap it with Defer to add subscription side-effects.
2. If the source is hot, then wrap it with Defer to add subscription side-effects.

The source is external and I want a hot observable.

1. If the source is cold, then Publish it.  Alternatively, you can use PublishLast or Replay, if you need their additional behavior.
2. If the source is hot, then use it as is.

The source is local and I want a cold observable.

1. Define your observable with existing Rx generator methods; e.g., Return, Range, Timer, Interval, Generate or Create.

The source is local and I want a hot observable.

1. This scenario is a potential candidate for defining your observable as Subject<T>, or even BehaviorSubject, AsyncSubject or ReplaySubject, if you need their behavior.  More details on why it's merely a "potential candidate" later in this post.

Note that Publish, PublishLast and Replay use subjects internally (via Multicast).  This isn't a coincidence.  Subscribing a subject to a cold observable broadcasts its notifications to multiple observers, thus making it hot.

So it's clear there are only two scenarios where it's correct to use subjects:

1. The source is external and cold, and I want a hot observable.
2. The source is local and I want a hot observable.

Let's simplify those statements.  The two scenarios where it's correct to use subjects are:

1. Converting a cold observable into a hot observable.
2. Generating a hot observable imperatively and statefully, without any direct external source.

Although, only in the latter scenario is it correct to use subjects explicitly.  As mentioned previously, Rx defines various operators like Publish for use in the former scenario.  Thus, only 1 scenario remains.

And so now we've answered our original questions:

When should I use a subject?  To generate a hot observable imperatively and statefully, without any direct external source.
When shouldn't I use a subject?  Everywhere else.

Perhaps that answer is a bit vague.  What do I mean by imperative and stateful?  What about local vs. external?  And why did I mention before that our final scenario is only a "potential candidate" for using a subject?

If (true == interested_in_the_answers_to_these_questions) { continue; } else { goto Conclusion; }

External Sources

Let's redefine external as quite simply a source of notifications that you do not have to generate yourself.  It comes from outside the scope of the observable that you're defining, such as a dependent object or a parameter.  It could be any number of things; e.g., an event on another object, an event on a base class, a timer callback, a Task<T> that represents the asynchronous result of reading from a file or sending a web request, an IEnumerable<T>, etc.

All of these examples have something in common: they can easily be converted into IObservable<T> using various From* and To* operators in Rx.  For example, Task<T> is converted by the ToObservable operator.  So if your external source is not already an IObservable<T>, then simply covert it using one of Rx's existing conversion operators.

Once you have an IObservable<T>, you need to determine whether it's hot or cold and whether you need it to be hot or cold.  If there's a conflict, as defined above, then use either Defer or Publish to change the temperature accordingly.  For this, there's no need to use subjects in your code.²

Finally, use LINQ to query the observable.  Again, you shouldn't use subjects here either.

The result is that you'll have a succinct observable query, controlling side-effects and meeting your specification in a declarative manner.  Subjects are simply the wrong tool for this job.  It's like hammering in a nail with a sledgehammer.  Subjects do much more than what is needed.  (Though as you'll see later in this post, in some ways subjects actually do less than what you may think, potentially introducing nasty threading bugs.)

Local / Imperative / Stateful Sources

I'm defining local as a source where you imperatively push notifications toward the outside of a stateful scope.  Another way of looking at it: a local source is external to other scopes that can't access its implementation.

Sorry if that explanation is confusing.  Let's clear it up.  Consider the following abstract requirements:

1. You must define a hot observable.
2. The code that generates the observable's notifications must be encapsulated by some type that you define.
3. Your code generates notifications when callers invoke certain members of your type or when it performs certain operations.
4. The observable must be exposed as a public property.

Now let's apply these requirements to what we've already learned earlier: 

1. Since we're defining a hot observable, Subject<T> is certainly a candidate...
2. We've established a scope for what is external: any source defined by another type and referenced by ours is external to our type.
3. We must write code to push notifications (input); therefore, our source is local.  We have no external source to convert or to generate notifications for us.
4. Our code must be able to reference something in order to push notifications, thus our type must be stateful.  We'll store a reference to "something" in a field.
5. The field will be returned by a public property (output).
6. Thus we need a type that represents both input and output, and is hot.

How should we model these requirements?

Scroll down for the answer:
.
.
.
.
.
.
.
.

Answer: Define a .NET event and expose it as an observable public property using FromEventPattern!

Ha, I bet you didn't see that one coming, did you?

Yes, alternatively we could use Subject<T>.  The conditions are perfect.  If we don't actually need or want a .NET event, then we should use a subject; however, if we already have an event defined, then there's no need to duplicate state and "raise" two similar events.  The conversion approach is probably better in that case.

Let's compare the similarity of subjects to .NET events in general:

1. Events are also hot.
2. An event's delegate shares the same scope (encapsulation) within our type as our proposed observable.
3. We also write code to raise events (input).
4. Our code must be able to reference the delegate in order to raise events, thus our type must be stateful.  We store a reference to the delegate in a field.
5. The field will be returned by a public event (output).
6. A multicast delegate represents both input and output, and is hot.

Subject<T> is kind of like Rx's implementation of an observable "event".  It broadcasts notifications to multiple observers, like raising an event for multiple event handlers.  It is the stateful component of Rx as it can be stored in a field or a variable and mutated (invoked) imperatively.  It is hot, so subscribing does not cause any side-effects (and subscribers may have already missed notifications, which is a common symptom of events and hot observables alike.)

It's correct to use Subject<T> to replace local .NET events, though .NET events are still useful sometimes due to their metadata and first-class language/tooling support.  It all depends on context, so consider this carefully before replacing all of your .NET events with subjects.  You may want to examine each event on a case-by-case basis.  And keep in mind that you can always expose an event as an observable via FromEventPattern, or perhaps it's best to leave that decision up to the consumers of your class and simply expose a .NET event.

Subjects actually have an additional related use case that events cannot cover: If we shrink our scope from type to method and apply the same logic as before, then we'll discover that subjects may still apply though events do not.

Let's specify similar method-scoped requirements:

1. You must define a hot observable.
2. The code that generates the observable's notifications must be within a method that you define.
3. Your code generates notifications when it performs certain asynchronous operations within the method.
4. The observable must be exposed as the return value of the method.

And again, let's apply these requirements to what we've already learned earlier: 

1. Since we're defining a hot observable, Subject<T> is certainly a candidate...
2. We've established a scope for what is external: any source passed into our method and any fields on the method's defining type are external to our method.
3. We must write code to push notifications (input); therefore, our source is local.  We have no external source to convert or to generate notifications for us.
4. Our code must be able to reference something in order to push notifications, thus our method must be stateful.  We'll store a reference to "something" in a local variable.
5. The variable will be returned from the method (output).
6. Thus we need a type that represents both input and output, and is hot.

How should we model these requirements?

Scroll down for the answer:
.
.
.
.
.
.
.
.

Answer 1: Write an anonymous event (a.k.a., anonymous method, delegate or lambda expression) and convert it using FromEvent!

Ha, did I trick you again?

You thought for sure this time I'd recommend using Subject<T>.  You'd also be correct, we could use Subject<T>.  The conditions may be perfect, though as you'll see in a moment there are better alternatives, even better than this one.

Alright, that answer was actually a joke, I don't recommend that particular approach at all.  It was meant to get you thinking about alternatives to subjects in places where you may think they're absolutely necessary.

But out of curiosity, how would that have worked anyway?

An event is backed by a delegate; it's the delegate that has all of those important properties (e.g., input, output, hot) that we discussed earlier.  For example, we could start by declaring a delegate as our observer.  By assigning the delegate to a local variable we can invoke it imperatively like a method, which is similar to calling OnNext on an observer.  We can use Task.Factory.StartNew to execute an async computation that invokes the delegate.  Finally, we can convert our delegate into an observable via the FromEvent method, which gives us our return value.

NOTE:  I don't recommend this approach, as mentioned previously.  For one thing, it doesn't support OnError or OnCompleted.  It doesn't support cancellation either, but given that we're generating a hot observable perhaps it doesn't really matter.  It's also a bit strange anyway.

using System;
using System.Reactive.Linq;
using System.Threading.Tasks;

namespace Rx.Labs
{
    static class ToUseSubjectOrNotToUseSubject
    {
        public static void Main()
        {
            IObservable<string> xs = Generate();

            using (xs.Take(1).Subscribe(Console.WriteLine))
            using (xs.Take(2).Subscribe(Console.WriteLine))
            using (xs.Take(3).Subscribe(Console.WriteLine))
            using (xs.Take(4).Subscribe(Console.WriteLine))
            {
                Console.ReadLine();
            }
        }

        static IObservable<string> Generate()
        {
            Action<string> observer = _ => { };    /* We could use null, but then at every invocation 
                                                    * we'd have to copy to a local and check for null.
                                                    */

            Task.Factory.StartNew(() =>
                {
                    Console.WriteLine("Invocation side-effect.");

                    Task.Delay(TimeSpan.FromSeconds(1)).Wait();  // Simulate work

                    observer("Generating");
                    observer("a");
                    observer("hot");
                    observer("observable.");
                });

            return Observable.FromEvent<string>(
                eh => observer += eh,
                eh => observer -= eh);
        }
    }
}

Output:

Invocation side-effect.
Generating
Generating
Generating
Generating
a
a
a
hot
hot
observable.

Answer 2 (Recommended):  Define an observable using an existing Rx generator method, then apply Publish with RefCount.

Rx provides various generators that make it easy to create an observable from thin air; e.g., Return, Range, Timer, Interval, Generate, and also Create.  If none of the former generators apply, then fallback to Create.  Create is the Swiss Army knife of observable generation.  I suspect that it also has similar performance characteristics to Subject<T>.  (See the Thread Safety section below for details.)

A particularly useful approach is to write an async iterator using a special overload of Create:

Rx 2.0 (Preferred):    Create<TResult>(Func<IObserver<TResult>, Task>)
Rx 1.1 Experimental:   Create<TResult>(Func<IObserver<TResult>, IEnumerable<IObservable<object>>>)

Generators are better than subjects because they allow you to remain functional, though Create does seem quite similar to using a subject.  Create lets you take more of a declarative approach, which is hopefully easier to understand and maintain than code that uses a subject.  Note that a primary benefit of Create is that it supports cancelation and auto-detachment, though it's not very useful when you need a hot observable.

Generators aren't the whole story either.  They return a cold observable, but our specification was to return a hot observable.

That's where Publish is useful, as described earlier in this post.  Publish converts a cold observable into a hot observable; however, it returns IConnectableObservable<T>, which requires a call to Connect.  You can return IConnectableObservable<T> from your method, just make sure that you change the return type of your method to match so that callers have the opportunity to call Connect themselves.

Alternatively, you can call Publish then RefCount to get an IObservable<T> that is shared among multiple consecutive observers.  Note that this isn't truly a hot observable - it's more like a warm observable.  RefCount makes a single subscription to the underlying observable while there's at least one observer of your query.  When your query has no more observers, changing the reference count to 0, the underlying subscription is disposed.  If another observer subscribes to your query later, moving the reference count from 0 to 1 again, then RefCount makes a new subscription to the underlying observable, causing subscription side-effects to occur again.

Note that it's bad practice to call Connect yourself in this scenario.  If you're tempted to do so, perhaps because RefCount doesn't meet your needs, then you should strongly consider returning a cold observable instead and leave it up to callers to decide whether Publish should be used at all.

Also beware to avoid closures when using Create.³  Closing over a local variable that you intend to mutate may cause unexpected behavior since every observer will share the same variable.  This may seem fine when generating a hot observable, and perhaps it is, though RefCount generates a warm observable, as mentioned previously.

Here's a correct example that uses Create, Publish and RefCount instead of a subject: 

using System;
using System.Reactive.Concurrency;
using System.Reactive.Linq;
using System.Threading.Tasks;

namespace Rx.Labs
{
    static class ToUseSubjectOrNotToUseSubject
    {
        public static void Main()
        {
            IObservable<string> xs = Generate();

            using (xs.Take(1).Subscribe(Console.WriteLine))
            using (xs.Take(2).Subscribe(Console.WriteLine))
            using (xs.Take(3).Subscribe(Console.WriteLine))
            using (xs.Take(4).Subscribe(Console.WriteLine))
            {
                Console.ReadLine();
            }
        }

        static IObservable<string> Generate()
        {
            return Observable.Create<string>(observer =>
                TaskPoolScheduler.Default.Schedule(() =>
                {
                    Console.WriteLine("Subscription side-effect.");

                    Task.Delay(TimeSpan.FromSeconds(1)).Wait();  // Simulate work

                    observer.OnNext("Generating");
                    observer.OnNext("a");
                    observer.OnNext("hot");
                    observer.OnNext("observable.");

                    observer.OnCompleted();
                }))
                .Publish()
                .RefCount();
        }
    }
}

Output:

Subscription side-effect.
Generating
Generating
Generating
Generating
a
a
a
hot
hot
observable.

Thread Safety

Subject<T>, Observable.Create and ObserverBase<T> do not make the same guarantees that are made by Rx generator methods, for performance reasons.  They allow callers to violate some of Rx's contracts for IObserver<T>, though they do ensure some contracts automatically regardless.

The contracts to which I'm referring are defined in the Rx Design Guidelines document.

Section 4.2 basically states that notifications must be pushed serially for Rx operators to behave properly.  But more generally, this should be a safe assumption everywhere.  When you subscribe to an observable, you don't expect to observe concurrent notifications.  Imagine coding subscriptions against this assumption but actually receiving concurrent notifications at runtime.  Take a look at some of your existing code.  What could happen?  For example, if you subscribe on a UI thread then you may get a threading context exception.  If you subscribe on some other thread instead, then maybe you'll get some strange threading bugs or maybe you'll get an exception, if you're lucky.

All Rx generator methods, excluding Create, ensure that this contract is satisfied.  Subject<T> and ObserverBase<T> do not.  For instance, calling OnNext concurrently on a Subject<T> will generate concurrent notifications.⁴

Section 4.1 states the Rx grammar as follows:

OnNext* (OnCompleted | OnError)?

All Rx operators and generator methods ensure that this contract is satisfied.  ObservableBase<T>, ObserverBase<T>, Create and Subject<T> actually ensure this contract as well, but only for serial notifications.  If a caller doesn't satisfy the §4.2 contract, then it seems that these APIs do not automatically ensure the §4.1 contract is satisfied either, due to lock-free race conditions.  For instance, it seems to me that calling a subject's OnNext and OnError methods simultaneously on different threads may cause an observer to receive OnError followed by OnNext.  But if you call OnError followed by OnNext on the same thread, then the call to OnNext is correctly ignored.

Rx offers Synchronize to satisfy these contracts for a given subject.  It offers this Synchronize method to do the same for observers, in general.  And of course, just because you can violate the contracts doesn't mean that you necessarily will.  But you do have to be aware of them when using these APIs, especially considering the kinds of bugs that may arise.  Threading bugs originating from within Rx operators, due to the misuse of a subject somewhere within a query, can be very difficult to diagnose.

So when you find yourself invoking OnNext, OnError or OnCompleted on a subject or within Create, make sure that you're satisfying Rx's contracts.

The takeaway is that when a subject isn't necessary due to the presence of an external source or an appropriate generator method, it's best not to use a subject in the first place so you can avoid having to invoke it imperatively and risk introducing subtle threading bugs in your queries.

Correct Examples

As shown in this post, there's only 1 particular set of conditions in which it's absolutely appropriate to use subjects.  Here are some examples.

Reactive Property

A common example of when it's correct to use a subject is to represent a sequence of property changes for a particular property, exposing it with AsObservable to hide its identity.  It's similar to implementing the PropertyChanged event, but only raising it for a particular property.  Instead of raising an event from within the property's setter, you'd call subject.OnNext(value).

Subject<T> works fine, though more commonly BehaviorSubject<T> is used instead because it stores the latest value of the property and pushes it immediately to new observers.  If your program is highly reactive, then you may find that you don't even need to keep a backing field for the property since BehaviorSubject<T> encapsulates it.  To set the value, call subject.OnNext(value).  Unfortunately, there's no way to extract the value imperatively (at least not until Rx exposes a Current property), so you may find that a backing field is still necessary unless your program is highly reactive.  Subscribing to the subject is the natural way of extracting the current value and all subsequent changes.

using System;
using System.Reactive.Linq;
using System.Reactive.Subjects;

namespace Rx.Labs
{
    class SubjectAsProperty
    {
        public static void Main()
        {
            var foo = new Foo();

            Console.WriteLine("Normal property usage:");

            Console.WriteLine(foo.Number);
            foo.Number += 8;
            Console.WriteLine(foo.Number);

            Console.WriteLine("Reactive property usage:");

            using (foo.NumberObservable.Subscribe(n => Console.WriteLine("Observer A: {0}", n)))
            {
                foo.Number = 1;
                foo.Number = 2;

                using (foo.NumberObservable.Subscribe(n => Console.WriteLine("Observer B: {0}", n)))
                {
                    foo.Number = 3;
                }

                using (foo.NumberObservable.Subscribe(n => Console.WriteLine("Observer C: {0}", n)))
                {
                    foo.Number = 4;
                }
            }

            Console.WriteLine(foo.Number);
            Console.ReadKey();
        }
    }

    class Foo
    {
        public int Number
        {
            get
            {
                return number.FirstAsync().Wait();
            }
            set
            {
                number.OnNext(value);
            }
        }

        public IObservable<int> NumberObservable
        {
            get
            {
                return number.AsObservable();
            }
        }

        private readonly BehaviorSubject<int> number = new BehaviorSubject<int>(42);
    }
}

Output:

Normal property usage:
42
50
Reactive property usage:
Observer A: 50
Observer A: 1
Observer A: 2
Observer B: 2
Observer A: 3
Observer B: 3
Observer C: 3
Observer A: 4
Observer C: 4
4

Cross-Cutting APIs

It's common to use a subject to represent cross-cutting concerns in a reactive manner.  For example, imagine exposing a property on a purely memory-bound logging class, returning an observable that pushes a notification whenever something is logged, such as entering and exiting methods (possibly via AOP).  This could be useful for real-time performance analysis, reactive heuristic-based adjustments, post-mortem debugging and even real-time debugging.  You could have one observer writing notifications to disk and one writing them to a GUI console, while another adjusts algorithms at runtime based on performance heuristics.  ReplaySubject<T> could be used to buffer notifications in memory and replay them to new observers, if necessary.

Another example would be a sequence of alerts exposed as an observable property on an Alert class.  Various parts of the program could send alerts concurrently by calling Alert.Unimportant(...) or Alert.RequiresUserInteraction(...), for example.  One observer could send emails while another displays toast notifications.

using System;
using System.Reactive.Linq;
using System.Reactive.Subjects;

namespace Rx.Labs
{
    class AlertLab
    {
        public static void Main()
        {
            var toast = from alert in Alert.All.OfType<ToastAlert>()
                        select alert.Message;

            var email = Alert.All.OfType<EmailAlert>();

            using (toast.Subscribe(a => Console.WriteLine("Important: {0}", a)))
            using (email.Subscribe(a => Console.WriteLine("Email \"{0}\" with \"{1}\".", a.To, a.Message)))
            {
                new ToastAlert("Hello Console!").Send();
                new EmailAlert("Hello Fred!", "fred@example.com").Send();
                new ToastAlert("Goodbye!").Send();
            }
        }
    }

    public abstract class Alert
    {
        public static IObservable<Alert> All
        {
            get
            {
                return alerts.AsObservable();
            }
        }

        public string Message
        {
            get
            {
                return message;
            }
        }

        private static readonly ISubject<Alert, Alert> alerts = Subject.Synchronize(new Subject<Alert>());
        private readonly string message;

        protected Alert(string message)
        {
            this.message = message;
        }

        public void Send()
        {
            alerts.OnNext(this);
        }
    }

    public class ToastAlert : Alert
    {
        public ToastAlert(string message)
            : base(message)
        {
        }
    }

    public class EmailAlert : Alert
    {
        public string To { get; private set; }

        public EmailAlert(string message, string to)
            : base(message)
        {
            this.To = to;
        }
    }
}

Output:

Important: Hello Console!
Email "fred@example.com" with "Hello Fred!".
Important: Goodbye!

Conclusion

If you've got anything that can be converted into an observable, such as a Task or Task<T>, an event, or an enumerable, then convert it using a From* or To* operator such as these or this, or if you've already got an observable, then use it.  Now consider temperature conflicts, change the temperature if necessary via Defer or Publish, and finally query it.  Don't use subjects.

Else if you don't have an observable or anything that can be converted into one:

If you require a cold observable, then use an existing Rx generator method; e.g., Return, Range, Timer, Interval, Generate or Create.  Don't use subjects.

Else if you require a hot observable:

If the scope of your observable is a method (i.e., it's entirely encapsulated by a method), then use an existing Rx generator method.  You can also apply Publish and RefCount, but consider relaxing the conditions to simply return a cold observable.  Don't use subjects.

Else if the scope of your observable is a type (e.g., it's exposed as a public property and backed by a field):

If you need to define a similar event or a similar event already exists, then convert the event into an observable like the first case.  Don't use subjects.

Else if you don't need to define a similar event and no similar event already exists, then use a subject.

In summary:

When should I use a subject?

When all of the following are true:

• you don't have an observable or anything that can be converted into one.
• you require a hot observable.
• the scope of your observable is a type.
• you don't need to define a similar event and no similar event already exists.

Why should I use a subject in that case?

Because you've got no choice!

1. At least that's my interpretation of Erik's comments.  I find it to be really good advice in practice when working with Rx.
2. Perhaps in very advanced and rare scenarios you may find a custom ISubject<T> implementation to be useful, in which case it's safe to use Multicast rather than Publish.
3. Which happens to be an easy-to-miss bug when implementing custom operators in Rx.  If you need to return a cold observable, as generally operators do, and you must use a closure, then wrap the closure and the call to Create in a call to Defer to ensure that the closure isn't shared across multiple subscriptions.
4. In some advanced and rare scenarios you may want to elide this contract and permit concurrent notifications through a subject, for performance reasons, though you should do so only if you can guarantee that none of Rx's query operators will be applied to the subject and all subscribers are thread-safe.