Computation Expressions

Computation expressions in F# provide a convenient syntax for writing computations that can be sequenced and combined using control flow constructs and bindings. Depending on the kind of computation expression, they can be thought of as a way to express monads, monoids, monad transformers, and applicative functors. However, unlike other languages (such as do-notation in Haskell), they are not tied to a single abstraction, and do not rely on macros or other forms of metaprogramming to accomplish a convenient and context-sensitive syntax.

Overview

Computations can take many forms. The most common form of computation is single-threaded execution, which is easy to understand and modify. However, not all forms of computation are as straightforward as single-threaded execution. Some examples include:

• Non-deterministic computations
• Asynchronous computations
• Effectful computations
• Generative computations

More generally, there are context-sensitive computations that you must perform in certain parts of an application. Writing context-sensitive code can be challenging, as it is easy to "leak" computations outside of a given context without abstractions to prevent you from doing so. These abstractions are often challenging to write by yourself, which is why F# has a generalized way to do so called computation expressions.

Computation expressions offer a uniform syntax and abstraction model for encoding context-sensitive computations.

Every computation expression is backed by a builder type. The builder type defines the operations that are available for the computation expression. See Creating a New Type of Computation Expression, which shows how to create a custom computation expression.

Syntax overview

All computation expressions have the following form:

builder-expr { cexper }

In this form, builder-expr is the name of a builder type that defines the computation expression, and cexper is the expression body of the computation expression. For example, async computation expression code can look like this:

let fetchAndDownload url =
    async {
        let! data = downloadData url

        let processedData = processData data

        return processedData
    }

There is a special, additional syntax available within a computation expression, as shown in the previous example. The following expression forms are possible with computation expressions:

expr { let! ... }
expr { and! ... }
expr { do! ... }
expr { yield ... }
expr { yield! ... }
expr { return ... }
expr { return! ... }
expr { match! ... }

Each of these keywords, and other standard F# keywords are only available in a computation expression if they have been defined in the backing builder type. The only exception to this is match!, which is itself syntactic sugar for the use of let! followed by a pattern match on the result.

The builder type is an object that defines special methods that govern the way the fragments of the computation expression are combined; that is, its methods control how the computation expression behaves. Another way to describe a builder class is to say that it enables you to customize the operation of many F# constructs, such as loops and bindings.

let!

The let! keyword binds the result of a call to another computation expression to a name:

let doThingsAsync url =
    async {
        let! data = getDataAsync url
        ...
    }

If you bind the call to a computation expression with let, you will not get the result of the computation expression. Instead, you will have bound the value of the unrealized call to that computation expression. Use let! to bind to the result.

let! is defined by the Bind(x, f) member on the builder type.

and!

The and! keyword allows you to bind the results of multiple computation expression calls more efficiently. This keyword enables applicative computation expressions, which provide a different computational model from the standard monadic approach.

let doThingsAsync url =
    async {
        let! data = getDataAsync url
        and! moreData = getMoreDataAsync anotherUrl
        and! evenMoreData = getEvenMoreDataAsync someUrl
        ...
    }

Using a series of let! ... let! ... executes the computations sequentially, even if they are independent. In contrast, let! ... and! ... indicates that the computations are independent, allowing applicative combination. This independence allows computation expression authors to:

• Execute computations more efficiently.
• Can run computations in parallel.
• Accumulate results without unnecessary sequential dependencies.

The restriction is that computations combined with and! cannot depend on the results of previously bound values within the same let!/and! chain. This trade-off enables the performance benefits.

and! is defined primarily by the MergeSources(x1, x2) member on the builder type.

Optionally, MergeSourcesN(x1, x2 ..., xN) can be defined to reduce the number of tupling nodes, and BindN(x1, x2 ..., xN, f), or BindNReturn(x1, x2, ..., xN, f) can be defined to bind computation expression results efficiently without tupling nodes.

For more information on applicative computation expressions, see Applicative Computation Expressions in F# 5 and F# RFC FS-1063.

do!

The do! keyword is for calling a computation expression that returns a unit-like type (defined by the Zero member on the builder):

let doThingsAsync data url =
    async {
        do! submitData data url
        ...
    }

For the async workflow, this type is Async<unit>. For other computation expressions, the type is likely to be CExpType<unit>.

do! is defined by the Bind(x, f) member on the builder type, where f produces a unit.

yield

The yield keyword is for returning a value from the computation expression so that it can be consumed as an IEnumerable<T>:

let squares =
    seq {
        for i in 1..10 do
            yield i * i
    }

for sq in squares do
    printfn $"%d{sq}"

In most cases, it can be omitted by callers. The most common way to omit yield is with the -> operator:

let squares =
    seq {
        for i in 1..10 -> i * i
    }

for sq in squares do
    printfn $"%d{sq}"

For more complex expressions that might yield many different values, and perhaps conditionally, simply omitting the keyword can do:

let weekdays includeWeekend =
    seq {
        "Monday"
        "Tuesday"
        "Wednesday"
        "Thursday"
        "Friday"
        if includeWeekend then
            "Saturday"
            "Sunday"
    }

As with the yield keyword in C#, each element in the computation expression is yielded back as it is iterated.

yield is defined by the Yield(x) member on the builder type, where x is the item to yield back.

yield!

The yield! keyword is for flattening a collection of values from a computation expression:

let squares =
    seq {
        for i in 1..3 -> i * i
    }

let cubes =
    seq {
        for i in 1..3 -> i * i * i
    }

let squaresAndCubes =
    seq {
        yield! squares
        yield! cubes
    }

printfn $"{squaresAndCubes}"  // Prints - 1; 4; 9; 1; 8; 27

When evaluated, the computation expression called by yield! will have its items yielded back one-by-one, flattening the result.

yield! is defined by the YieldFrom(x) member on the builder type, where x is a collection of values.

Unlike yield, yield! must be explicitly specified. Its behavior isn't implicit in computation expressions.

return

The return keyword wraps a value in the type corresponding to the computation expression. Aside from computation expressions using yield, it is used to "complete" a computation expression:

let req = // 'req' is of type 'Async<data>'
    async {
        let! data = fetch url
        return data
    }

// 'result' is of type 'data'
let result = Async.RunSynchronously req

return is defined by the Return(x) member on the builder type, where x is the item to wrap. For let! ... return usage, BindReturn(x, f) can be used for improved performance.

return!

The return! keyword realizes the value of a computation expression and wraps that result in the type corresponding to the computation expression:

let req = // 'req' is of type 'Async<data>'
    async {
        return! fetch url
    }

// 'result' is of type 'data'
let result = Async.RunSynchronously req

return! is defined by the ReturnFrom(x) member on the builder type, where x is another computation expression.

match!

The match! keyword allows you to inline a call to another computation expression and pattern match on its result:

let doThingsAsync url =
    async {
        match! callService url with
        | Some data -> ...
        | None -> ...
    }

When calling a computation expression with match!, it will realize the result of the call like let!. This is often used when calling a computation expression where the result is an optional.

Built-in computation expressions

The F# core library defines four built-in computation expressions: Sequence Expressions, Async expressions, Task expressions, and Query Expressions.

Creating a New Type of Computation Expression

You can define the characteristics of your own computation expressions by creating a builder class and defining certain special methods on the class. The builder class can optionally define the methods as listed in the following table.

The following table describes methods that can be used in a workflow builder class.

Many of the methods in a builder class use and return an M<'T> construct, which is typically a separately defined type that characterizes the kind of computations being combined, for example, Async<'T> for async expressions and Seq<'T> for sequence workflows. The signatures of these methods enable them to be combined and nested with each other, so that the workflow object returned from one construct can be passed to the next.

Many functions use the result of Delay as an argument: Run, While, TryWith, TryFinally, and Combine. The Delayed<'T> type is the return type of Delay and consequently the parameter to these functions. Delayed<'T> can be an arbitrary type that does not need to be related to M<'T>; commonly M<'T> or (unit -> M<'T>) are used. The default implementation is M<'T>. For a more in-depth look, see Understanding the type constraints.

The compiler, when it parses a computation expression, translates the expression into a series of nested function calls by using the methods in the preceding table and the code in the computation expression. The nested expression is of the following form:

builder.Run(builder.Delay(fun () -> {{ cexpr }}))

In the above code, the calls to Run and Delay are omitted if they are not defined in the computation expression builder class. The body of the computation expression, here denoted as {{ cexpr }}, is translated into further calls to the methods of the builder class. This process is defined recursively according to the translations in the following table. Code within double brackets {{ ... }} remains to be translated, expr represents an F# expression and cexpr represents a computation expression.

In the previous table, other-expr describes an expression that is not otherwise listed in the table. A builder class does not need to implement all of the methods and support all of the translations listed in the previous table. Those constructs that are not implemented are not available in computation expressions of that type. For example, if you do not want to support the use keyword in your computation expressions, you can omit the definition of Use in your builder class.

The following code example shows a computation expression that encapsulates a computation as a series of steps that can be evaluated one step at a time. A discriminated union type, OkOrException, encodes the error state of the expression as evaluated so far. This code demonstrates several typical patterns that you can use in your computation expressions, such as boilerplate implementations of some of the builder methods.

/// Represents computations that can be run step by step
type Eventually<'T> =
    | Done of 'T
    | NotYetDone of (unit -> Eventually<'T>)

module Eventually =

    /// Bind a computation using 'func'.
    let rec bind func expr =
        match expr with
        | Done value -> func value
        | NotYetDone work -> NotYetDone (fun () -> bind func (work()))

    /// Return the final value
    let result value = Done value

    /// The catch for the computations. Stitch try/with throughout
    /// the computation, and return the overall result as an OkOrException.
    let rec catch expr =
        match expr with
        | Done value -> result (Ok value)
        | NotYetDone work ->
            NotYetDone (fun () ->
                let res = try Ok(work()) with | exn -> Error exn
                match res with
                | Ok cont -> catch cont // note, a tailcall
                | Error exn -> result (Error exn))

    /// The delay operator.
    let delay func = NotYetDone (fun () -> func())

    /// The stepping action for the computations.
    let step expr =
        match expr with
        | Done _ -> expr
        | NotYetDone func -> func ()

    /// The tryFinally operator.
    /// This is boilerplate in terms of "result", "catch", and "bind".
    let tryFinally expr compensation =
        catch (expr)
        |> bind (fun res ->
            compensation();
            match res with
            | Ok value -> result value
            | Error exn -> raise exn)

    /// The tryWith operator.
    /// This is boilerplate in terms of "result", "catch", and "bind".
    let tryWith exn handler =
        catch exn
        |> bind (function Ok value -> result value | Error exn -> handler exn)

    /// The whileLoop operator.
    /// This is boilerplate in terms of "result" and "bind".
    let rec whileLoop pred body =
        if pred() then body |> bind (fun _ -> whileLoop pred body)
        else result ()

    /// The sequential composition operator.
    /// This is boilerplate in terms of "result" and "bind".
    let combine expr1 expr2 =
        expr1 |> bind (fun () -> expr2)

    /// The using operator.
    /// This is boilerplate in terms of "tryFinally" and "Dispose".
    let using (resource: #System.IDisposable) func =
        tryFinally (func resource) (fun () -> resource.Dispose())

    /// The forLoop operator.
    /// This is boilerplate in terms of "catch", "result", and "bind".
    let forLoop (collection:seq<_>) func =
        let ie = collection.GetEnumerator()
        tryFinally
            (whileLoop
                (fun () -> ie.MoveNext())
                (delay (fun () -> let value = ie.Current in func value)))
            (fun () -> ie.Dispose())

/// The builder class.
type EventuallyBuilder() =
    member x.Bind(comp, func) = Eventually.bind func comp
    member x.Return(value) = Eventually.result value
    member x.ReturnFrom(value) = value
    member x.Combine(expr1, expr2) = Eventually.combine expr1 expr2
    member x.Delay(func) = Eventually.delay func
    member x.Zero() = Eventually.result ()
    member x.TryWith(expr, handler) = Eventually.tryWith expr handler
    member x.TryFinally(expr, compensation) = Eventually.tryFinally expr compensation
    member x.For(coll:seq<_>, func) = Eventually.forLoop coll func
    member x.Using(resource, expr) = Eventually.using resource expr

let eventually = new EventuallyBuilder()

let comp =
    eventually {
        for x in 1..2 do
            printfn $" x = %d{x}"
        return 3 + 4
    }

/// Try the remaining lines in F# interactive to see how this
/// computation expression works in practice.
let step x = Eventually.step x

// returns "NotYetDone <closure>"
comp |> step

// prints "x = 1"
// returns "NotYetDone <closure>"
comp |> step |> step

// prints "x = 1"
// prints "x = 2"
// returns "Done 7"
comp |> step |> step |> step |> step

A computation expression has an underlying type, which the expression returns. The underlying type may represent a computed result or a delayed computation that can be performed, or it may provide a way to iterate through some type of collection. In the previous example, the underlying type was Eventually<_>. For a sequence expression, the underlying type is System.Collections.Generic.IEnumerable<T>. For a query expression, the underlying type is System.Linq.IQueryable. For an async expression, the underlying type is Async. The Async object represents the work to be performed to compute the result. For example, you call Async.RunSynchronously to execute a computation and return the result.

Custom Operations

You can define a custom operation on a computation expression and use a custom operation as an operator in a computation expression. For example, you can include a query operator in a query expression. When you define a custom operation, you must define the Yield and For methods in the computation expression. To define a custom operation, put it in a builder class for the computation expression, and then apply the CustomOperationAttribute. This attribute takes a string as an argument, which is the name to be used in a custom operation. This name comes into scope at the start of the opening curly brace of the computation expression. Therefore, you shouldn't use identifiers that have the same name as a custom operation in this block. For example, avoid the use of identifiers such as all or last in query expressions.

Extending existing Builders with new Custom Operations

If you already have a builder class, its custom operations can be extended from outside of this builder class. Extensions must be declared in modules. Namespaces cannot contain extension members except in the same file and the same namespace declaration group where the type is defined.

The following example shows the extensions of the existing FSharp.Linq.QueryBuilder class.

open System
open FSharp.Linq

type QueryBuilder with

    [<CustomOperation>]
    member _.any (source: QuerySource<'T, 'Q>, predicate) =
        System.Linq.Enumerable.Any (source.Source, Func<_,_>(predicate))

    [<CustomOperation("singleSafe")>] // you can specify your own operation name in the constructor
    member _.singleOrDefault (source: QuerySource<'T, 'Q>, predicate) =
        System.Linq.Enumerable.SingleOrDefault (source.Source, Func<_,_>(predicate))

Custom operations can be overloaded. For more information, see F# RFC FS-1056 - Allow overloads of custom keywords in computation expressions.

Compiling computation expressions efficiently

F# computation expressions that suspend execution can be compiled to highly efficient state machines through careful use of a low-level feature called resumable code. Resumable code is documented in F# RFC FS-1087 and used for Task Expressions.

F# computation expressions that are synchronous (that is, they don't suspend execution) can alternatively be compiled to efficient state machines by using inline functions including the InlineIfLambda attribute. Examples are given in F# RFC FS-1098.

List expressions, array expressions, and sequence expressions are given special treatment by the F# compiler to ensure generation of high-performance code.

Tail operations in custom builders

For a custom builder, an explicit tail return! can use ReturnFromFinal and an explicit tail yield! can use YieldFromFinal, when the corresponding optional member exists. The ordinary ReturnFrom or YieldFrom remains the fallback.

Starting with F# 11, a final do! binds the operation and completes with unit instead of selecting a final-forwarding member. Normally it calls Bind then Return(()), or Zero() without Return. A Zero marked with DefaultValue takes precedence over Return. Binding optimizations such as BindReturn can still apply.

If your builder relied on the earlier behavior, supply binding and unit-completion methods. Use explicit return! or yield! only when you intend to forward the computation.

The distinction is observable in this builder:

type TraceBuilder() =
    member _.Return(value) =
        System.Console.WriteLine("Return")
        value
    member _.Zero() = ()
    member _.Bind(value, continuation) =
        System.Console.WriteLine("Bind")
        continuation value
    member _.ReturnFrom(value) =
        System.Console.WriteLine("ReturnFrom")
        value
    member _.ReturnFromFinal(value) =
        System.Console.WriteLine("ReturnFromFinal")
        value

let trace = TraceBuilder()
let forwarded = trace { return! 42 }
let completed = trace { do! () }

The output is:

ReturnFromFinal
Bind
Return

A let! still requires a following body. For example, use async { let! value = work in return value }, not the incomplete async { let! value = work }.

Rethrow from a computation expression

To log an error and preserve its original exception identity and stack, use reraise () in a computation-expression with handler:

open System
open System.IO
open System.Threading.Tasks

let readReport (logFailure: exn -> Task) path =
    task {
        try
            return! File.ReadAllTextAsync(path)
        with :? IOException as error ->
            do! logFailure error
            return reraise ()
    }

After the logger completes, reraise propagates the caught exception with its original throw-site information. If logging fails, its failure can replace the original. The nearest enclosing handler selects the exception; raise error doesn't preserve the stack in the same way.

Custom builders must infer the handler input as exn. The compiler uses ExceptionDispatchInfo, so dispatch frames can appear in the preserved stack.

See also

• F# Language Reference
• Sequences
• Async expressions
• Task expressions
• Query Expressions
• Series on Computation Expressions from F# for Fun and Profit