Why OO Matters (in F#)

F# is a functional-first programming language that comes with a substantial object-oriented feature set. It is so feature-complete in fact, that almost any C# class can be ported over to F# code with little substantial alteration.

However significant, this subset of the language is seeing limited appreciation from the community, which I suspect is partly fuelled by the known criticisms of OOP and partly by a desire to be different than C#. After all, this is a functional-first language so we can just replace all our classes with functions. There is also the opinion that OOP in F# merely serves as a compatibility layer for .NET, so it’s really only there to cover those unfortunate scenarios of having to use a library that accepts interfaces.

Enabling Abstraction

One of the most important aspects of maintaining a nontrivial codebase is controlling complexity. Complexity can be contained by partitioning code into logically standalone components whose implementation details are hidden behind appropriately designed abstractions. In his excellent Solid and Functional article, Vesa Karvonen argues that selecting the correct abstraction is a hard problem, and that functional programming is no silver bullet in dealing with that. This resonates a lot with me, and I strongly encourage everyone to read the full article.

That said, Vesa is framing the article in Standard ML which supports a full-blown module system. Modules can be abstracted using signatures or they can be parameterized by other modules using functors. Modules are the predominant means of abstraction in the ML ecosystem. In Haskell, it is type classes and higher-kinded types. In F#, modules are intentionally stripped of any complicated features, effectively functioning as a mere namespacing construct.

My claim is that there are inherent limits to what can be expressed using just F# modules and functions, in terms of enabling good abstraction. Luckily, we can always make use of the next best thing, which is F# OO. The thesis of this article is that strategically admitting elements of OO in an F# codebase significantly improves quality and maintainability. While I cannot conclusively prove this within the confines of a single blog post, I will try to provide hints as to why this is.

Classes as Value Parametric Modules

It is often the case that an API exposed through a module must be context aware. Typically F# developers address this by adding extra parameters in every function:

module MyApi =
    let function1 dep1 dep2 dep3 arg1 = doStuffWith dep1 dep2 dep3 arg1
    let function2 dep1 dep2 dep3 arg2 = doStuffWith' dep1 dep2 dep3 arg2

While this does work well in simple cases, it does not scale nicely as dependencies increase. It would typically prompt the developer to group arguments in context records:

type MyApiContext = { Dep1 : Dep1 ; Dep2 : Dep2 ; Dep3 : Dep3 }

module MyApi =
    let function1 (ctx : MyApiContext) arg1 = doStuffWith ctx.Dep1 ctx.Dep2 ctx.Dep3 arg1
    let function2 (ctx : MyApiContext) arg2 = doStuffWith' ctx.Dep1 ctx.Dep2 ctx.Dep3 arg2

This complicates the implementation even more both in the definition site and in the consumption site. In practice, you either end up with one context type per component or one God context for the entire application. Even more importantly, this approach often violates encapsulation concerns, pushing the burden of gathering dependencies to the consumers of the API, every single time they do consume the API. Partial application also does little to address any of these concerns in nontrivial contexts.

Less experienced developers might be prompted to do something even worse: lift dependencies to module values.

module MyApi =
    let dep1 = File.ReadAllText "/Users/eirik/connectionstring.txt"
    let dep2 = Environment.GetEnvironmentVariable "DEP_2"
    let dep3 = Random().Next()

    let function1 arg = doStuffWith dep1 dep2 dep3 arg
    let function2 arg = doSutffWith dep1 dep2 dep3 arg

This is bad for many reasons: it makes the API reliant on global state, introduces unneeded side-effects, and pushes app configuration concerns deep inside the guts of our codebase. What’s more, module value initialization compiles to a static constructor for the entire compilation unit so the exact moment of execution is largely unpredictable. Initialization errors manifest as TypeInitializationExceptions which are difficult to debug.

Contrast the situation above with the elegance afforded by a plain old class:

type MyParametricApi(dep1, dep2, dep3) =
    member __.Function1 arg1 = doStuffWith dep1 dep2 dep3 arg1
    member __.Function2 arg2 = doStuffWith' dep1 dep2 dep3 arg2

An API object could be created once at application initialization, or as many times required depending on context. It’s also more amenable to testing. I should add that this approach is essentially just as “functional” as the approaches above, since it’s merely composing dependencies to expose a context-sensitive API. Importantly, it achieves this in a much simpler way both in the definition site and consumption site, which pays great dividends if realized in big codebases.

Expressive APIs

An important attribute of method-based APIs is that they allow for greater expressive power, in two important ways:

  1. Named/Optional parameters: unlike OCaml, whose functions support out-of-order named argument passing and omitted optional arguments, F# functions support neither. Luckily, we can do this using F# methods. I find this to be an immensely powerful tool when exposing non-trivially parameterizable functionality. A function that explicitly accepts 10 optional parameters is not acceptable; a method that accepts 10 optional arguments works like a charm.
  2. Method overloading: because function names like connect' and bind2 are simply not good enough when exposed in a public API.

More Powerful Types

The type system afforded by .NET is strictly more powerful than what can be expressed using modules and functions. For example, the interface

type Scheduler =
    abstract Run<'T> : Async<'T> -> 'T

encodes a kind of function that cannot be expressed in terms of proper F# lambdas. When combined with subtyping, it is possible to effectively encode existential types and Rank-N polymorphism. Even GADTs are possible, with minimal augmentations of the type system.

In practice, it is possible to leverage that additional power very effectively. In fact, F# makes it easy to define generic function literals using object expressions. This is also how the TypeShape library has been made possible.

Abstracting Modules

Functions are the unit of abstraction in F#, but that unit is often insufficient when abstracting APIs. This prompts developers to adopt an approach where abstract APIs are surfaced as either records or tuples of functions:

type Serializer =
        Serialize : bool -> obj -> string
        Deserialize : bool -> string -> obj

According to the F# design guidelines, use of records for building APIs is discouraged and recommends using regular interfaces instead.

I strongly agree with this recommendation for a multitude of reasons: interfaces are more powerful since they support generic methods, named arguments and optional arguments. An interface is less likely to be defined in terms of closures, making it easier to reason about when viewing from a debugger.

So the example above could be rendered as an interface like so:

type Serializer =
    abstract Serialize<'T> : preserveRefEq:bool -> value:'T -> string
    abstract Deserialize<'T> : preserveRefEq:bool -> pickle:string -> 'T

The most important aspect of this approach is readability. It is easier for a consumer of this interface to anticipate what the purpose of each argument is, in the same way that it is easier to understand a record of functions over a tuple of functions.

Representing Illegal State

A lot of proverbial ink has been spent describing how we should strive to make illegal states unrepresentable. However, I do fail to see how this could be fully realized given that the functional core of F# only consists of algebraic data types. Take for example an oft-quoted email type:

type Email = Email of string

The following values are valid instaces of type Email:

Email null
Email "John Smith"
Email "eirik@foo.bar'; DROP TABLE dbo.Users"

If we truly care about illegal states, the obvious alteration to the type above ought to be the following

type Email = private | Email of string
    member this.Address = let (Email addr) = this in addr
    static member TryParse(address : string) =
        if isValidEmail address then Some(Email address)
        else None

But really, this is a just a class encoded by a union. The implementation below is simpler:

type Email private (address : string) =
    member __.Address = address
    static member TryParse(address : string) =
        if isValidEmail address then Some(Email address)
        else None

NB the previous implementation might in fact be warranted in cases where free structural equality or comparison are needed. But for all intents and purposes, both approaches effectively subscribe to OO-style encapsulation.

OO And Purity

The relationship between OO and purity can be a frequent avenue for misconception. Occasionally someone will claim that by admitting objects we are ipso facto forsaking purity. On the contrary, I do claim that these really are orthogonal concerns. Just as a lambda is capable of producing side-effects, objects can be designed for purity. Good examples of this are Map and Set in the core library. The lambda is really just an abstract class with a single virtual method and lots of syntactic sugar. There is nothing fundamentally setting it apart from objects once you exclude the syntactic aspect.


So, is this a call to go full-blown OO in F# projects? Should we be digging up our old GoF copies? Are design patterns up there in the functional curriculum together with profunctor optics? Are inheritance and class hierarchies sexy again? No!

I am in fact proposing that there is a third way, where functional and OO components coexist, with one paradigm complementing the other. This is hardly a new idea. Quoting from the F# design guidelines:

F# is commonly called a functional-first language: object, functional and imperative paradigms are all well supported, but functional programming tends to be the first technique used. Many of the defaults of F# are set up to encourage functional programming, but programming in the other paradigms is effective and efficient, and a combination is often best of all. It is a common misconception that the functional and object programming methodologies are competing. In fact, they are generally orthogonal and largely complementary. Often, functional programming plays a stronger role “in the small” (e.g. at the implementation level of functions/method and the code contained therein) and object programming playe a bigger role “in the large” (e.g. at the structural level of classes, interfaces, and namespaces, and the organization of APIs for frameworks).

In my 6 years of working with F#, my style has gradually shifted towards embracing this approach. A few examples:

  • I typically write the implementation of a large component in the functional style behind a private module, then expose its public API as part of a standalone class. I find that method-based APIs are friendlier to consumers unfamiliar with the implementation.
  • I use records and unions for encoding internal representations and classes for encapsulating publicly visible instances. A very good example of this is the F# map implementation.
  • I rarely expose records and unions as part of a public API unless it is abundantly evident that all possible instances for the given types are valid in their context of use. This does not happen often in my experience.
  • If a module is exposed as part of a public API, care must be taken so that the number of arguments is small and behaviour can be predicted by reading the type signature of the function alone. The core List and Array modules are a good example. Avoid using modules to expose complex functionality like the Async API.

I remember reading a few years back Simon Cousins’ NOOO manifesto, which stands for Not Only Object-Oriented development. In retrospect I find this to be an excellent name for a manifesto, if only because “Not Only OO” is not the same thing as “No OO”. So here’s a proposal to revive that manifesto, perhaps with the understanding that “Not Only OO” also implies “Not Only FP” in the context of F#.

Why OO Matters (in F#)

F# and Purity

Purely functional programming according to wikipedia

designates a programming paradigm that treats all computation as the evaluation of mathematical functions. Purely functional programing may also be defined by forbidding changing state and mutable data.

The importance of the purely functional approach cannot be overstated: it eliminates entire classes of mutation-related bugs; encourages composable abstraction; pure code can be reasoned about equationally; the explicit segregation of side-effects acts as a forcing function for better abstraction. At the same time, immutable persistent data structures can be cheaply incremented while shared safely across multiple threads. Pure programs also admit compile-time optimizations that would be tricky to achieve in the presence of state.

Benefits aside, purity is an attribute of expressions and their formal semantics. A pure program would still end up executing on regular computer hardware, where it could demonstrate the following effects: high CPU utilization, excessive heap allocations, stack overflows, divizion by zero errors, etc. Thus the connection of pure code to functions of the mathematical variety should not be considered a literal one, otherwise all correct Fibonacci implementations would be equal by virtue of extensionality. Being able to reason how code —pure or otherwise— executes on the underlying hardware is critically important for the working programmer.

An ML Tradition

Transforming purely functional declarations into equivalent, optimal imperative code is a holy grail, still very much a topic of active research. In the general case though, performance can only be attained by forsaking purity. This is very much compatible with the ML philosophy, which has traditionally embraced imperative features in its language core. Quoting from Real World OCaml

Pure code is the default in OCaml, and for good reason—it’s generally easier to reason about, less error prone and more composable. But imperative code is of fundamental importance to any practical programming language, because real-world tasks require that you interact with the outside world, which is by its nature imperative. Imperative programming can also be important for performance. While pure code is quite efficient in OCaml, there are many algorithms that can only be implemented efficiently using imperative techniques.

OCaml offers a happy compromise here, making it easy and natural to program in a pure style, but also providing great support for imperative programming.

This also applies particularly well to F#, an OCaml dialect for the .NET framework. In my personal experience, performance gains when switching from purely functional to imperative code in F# can be particularly dramatic. Strategically adopting imperative implementations in performance-critical components is often vital to ensuring sustainability of a system.

Imperative Programming done right

So what does this all mean? Should we just give up on purely functional code in the interest of performance? Is F# really just a way to write C# programs in ML-style syntax?

Well, not really. I do claim though that there is a sweet spot where impure features can be utilized to combine the benefits of pure FP with a lot of the efficiency of procedural code. Defining that sweet spot is hard, however I cannot think of a more concise example than Nessos Streams, which provides a highly efficient functional streams implementation in just over 40 lines of code. Other examples could include FParsec and Hopac, whose core implementations have been written in C# out of performance considerations.

If one had to nominate a key guiding priciple for using imperative features in F# code, then surely we would single out referential transparency. Very roughly, a function is referentially transparent if it behaves like a pure function, even if its implementation might use imperative constructs. For example, the function

let incr (cell : int ref) : unit = cell := !cell + 1

is not referentially transparent because substituting a call to incr with the the result of the computation is not a behaviour-preserving transformation, in general.

We can still use imperative features to define functions that are in fact referentially transparent. A good example is implementing groupBy for lists. A purely functional groupBy could be written as follows:

let groupBy (proj : 'T -> 'Key) (ts : 'T list) : ('Key * 'T list) list =
    let folder (map : Map<'Key, 'T list>) (t : 'T) =
        let key = proj t
        let grouped = defaultArg (Map.tryFind key map) []
        Map.add key (t :: grouped) map

    |> List.fold folder Map.empty
    |> Map.toList
    |> List.map (fun (k,v) -> k, List.rev v)

This implementation may be pure, however performance-wise it is not ideal. We can still produce a more efficient, imperative equivalent of the same function:

let groupBy' (proj : 'T -> 'Key) (ts : 'T list) : ('Key * 'T list) list =
    let dict = new Dictionary<'Key, ResizeArray<'T>>()
    for t in ts do
        let key = proj t
        let mutable array = null
        if not <| dict.TryGetValue(key, &array) then
            array <- new ResizeArray<'T>()
            dict.Add(key, array)
        array.Add t

    |> Seq.map (function KeyValue(k,v) -> k, Seq.toList v)
    |> Seq.toList

The second function provides order-of-magnitude performance gains, but is also referentially transparent: it can be substituted with the pure implementation or the resulting value without affecting behaviour of the overall program. In fact, this approach very much reflects how the core library implements its own functions.


Purely functional programming in F# provides a very effective means of reasoning about code, however in the interest of efficiency it is often preferable to leverage the imperative aspects of the language. When done carefully, it can result in better performing code that sacrifices few if any of the benefits of the purely functional approach.

In fact, my claim is that this is very much the idiomatic way of using ML in general and F# in particular. Pretending that F# is a purely functional language à la Haskell is highly counterproductive, particularly when authoring production-grade projects.

F# and Purity

You’re better off using Exceptions

Exception handling is an error management paradigm that has often been met with criticism. Such criticisms typically revolve around scoping considerations, exceptions-as-control-flow abuse or even the assertion that exceptions are really just a type safe version of goto. To an extent, these seem like valid concerns but it is not within the scope of this article to address those per se.

Such concerns resonate particularly well within FP communities, often taken to the extreme: we should reject exceptions altogether, since code that throws is necessarily impure. In the F# community, this opinion is in part realized by advocating alternatives like result types and railway-oriented programming. In essence, these approaches follow the Either monad found in Haskell, but often intentionally avoiding the use of do notation/computation expressions (since that’s just interpreted exception semantics).

The TL;DR version of the approach is that we define a union type for results that looks like this:

type Result<'TSuccess, 'TError> =
    | Ok of 'TSuccess
    | Error of 'TError

then require that any function which admits the possibility of error should return a Result value. This way we are forcing the type system acknowledge error cases and makes our error handling logic more explicit and predictable. We can also use combinators like bind to encode common error handling flows.

I would like to illustrate in this article why I believe that this often causes a lot of problems — while simultaneously not achieving the advertised benefits, particularly in the context of maintaining large codebases. It also gives rise to a multitude of anti-patterns, particularly when used by people new to F#.

An issue of Runtime

Most programming languages out there carry the concept of a runtime error. Such errors sometimes stem from unexpected locations, in code otherwise considered pure. In Haskell for instance, consider what the pure expressions “2 `div` 0” or “head []” evaluate to. In the case of the CLR, runtime errors are manifested in the form of thrown exceptions, which arguably happens more often than Haskell.

Because runtime errors are difficult to anticipate, I claim that using result types as a holistic replacement for error handling in an application is a leaky abstraction. Consider the following example:

type Customer = { Id : string; Credit : decimal option }

let average (customers : Customer list) =
    match customers with
    | [] -> Error "list was empty"
    | _ ->
        |> List.averageBy (fun c -> c.Credit.Value)
        |> Ok

Notice that the above snippet now admits two possible classes of errors, string and the potential NullReferenceException admitted by careless access of the optional field in the customer record. This eventually delivers an unpleasant surprise to consumers of the function, since the type signature communicates an exception-free implementation.

An Awkward Reconciliation

It’s by such misadventure that the working F# programmer soon realizes that any function could still potentially throw. This is an awkward realization, which has to be addressed by catching as soon as possible. This of course is the purely functional equivalent of the Pokemon anti-pattern:

let avg : Result<decimal, string> =
    try average customers
    with e -> Error e.Message // gotta catch em all!


In the majority of codebases using result types that I’ve been reviewing, people typically just end up re-implementing exception semantics on an ad-hoc basis. This can result in extremely noisy code, as illustrated in the following example:

let combine x y z =
    match x, y, z with
    | Ok x, Ok y, Ok z -> Ok (x,y,z)
    | Error e, _, _ | _, Error e, _ | _, _, Error e -> Error e

or in using pre-baked combinators

let combine x y z =
    x |> bind (fun x -> y |> bind (fun y -> z |> bind (fun z -> Ok(x,y,z))))

We really shouldn’t be doing this. It is essentially the type equivalent of using a linter that demands inserting catch-alls and rethrows in every method body.

Given sufficient implementation complexity, things can get really ugly: binding on results of differing error types and other incidental complexities require substantial refactoring to get things right. Often, this will prompt developers to cut corners by doing unexpected things, like inserting a throw just to get rid of an annoying error branch or returning nested result types like

Result<Result<string, string>, string list>

Where’s my Stacktrace?

An important property of exceptions -which cannot be stressed enough- is that they are entities managed and understood by the underlying runtime, endowed with metadata critical to diagnosing bugs in complex systems. They can also be tracked and highlighted by tooling such as debuggers and profilers, providing invaluable insight when probing a large system.

By lifting all of our error handling to passing result values, we are essentially discarding all that functionality. There is nothing worse than having to deal with a production issue which comes up in the logs simply as "list was empty".

The Problem with IO

The points discussed thus far have typically focused on applications that might as well have been purely functional. As might be expected, these become even more pronounced when bringing in IO or when interfacing with third-party .NET libraries. Consider the following example:

let tryReadAllText (path : string) : string option =
    try System.IO.File.ReadAllText path |> Some
    with _ -> None

I have spotted this approach and variations in many production-grade F# codebases, and have been responsible for some of these myself back when I was younger and more naive. This adapter is motivated by the desire to have an exception-free file reader, however it is deeply flawed precisely because of that.

While it is fairly unambiguous what a Some result means, the None bit hardly conveys any information other than the obvious. Here is an official list of possible errors that can be raised by that particular call. By discarding the exception, it becomes impossible to diagnose what could have gone wrong.

Stringly-typed error handling

One could quickly point out that the snippet above can be amended so that data loss is minimized.

let readAllText (path : string) : Result<string, string> =
    try System.IO.File.ReadAllText path |> Ok
    with e -> Error e.Message

This alteration still makes error handling awkward and unsafe

match readAllText "foo.txt" with
| Error e when e.Contains "Could not find file" -> // ...

particularly when compared to how the idiomatic approach works

try File.ReadAllText path
with : ? FileNotFoundException -> // ...

Interop issues

It is often easy to forget that Result values are plain objects, with no particular bearing when used in the context of other frameworks/libraries. The following example illustrates an innocent mistake when working with TPL:

let task =
    Task.StartNew(fun () ->
        if Random.Next() % 2 = 0
        then Ok ()
        else Error ())

if task.IsFaulted then printfn "Task failed"
else printfn "Task succeeded"

Or perhaps it could be mistakenly believed that F# Async somehow recognizes result types

let job1 = async { return Ok () }
let job2 = async { return Error () }

let! result =
    [job1 ; job2]
    |> Async.Parallel
    |> Async.map Ok

match result with
| Ok _ -> printfn "All jobs completed successfully"
| Error _ -> printfn "Some of the jobs failed"


It is not in the intentions of this article to argue that result types shouldn’t be used at all. I have written DSLs and interpreters using result types. MBrace uses result types internally. In most cases though, the benefit is only maximized by writing custom result types that model a particular domain. Rather than having a general-purpose error branch that encodes infinite classes of failures, assign each class of errors to an individual branch in that result type:

type MoneyWithdrawalResult =
    | Success of amount:decimal
    | InsufficientFunds of balance:decimal
    | CardExpired of DateTime
    | UndisclosedFailure

This approach constrains error classes in a finite domain, which also allows for more effective testing of our code.

That said, I strongly believe that using result types as a general-purpose error handling mechanism for F# applications should be considered harmful. Exceptions should remain the dominant mechanism for error propagation when programming in the large. The F# language has been designed with exceptions in mind, and has achieved that goal very effectively.

You’re better off using Exceptions

Reconciling Stacktraces with Computation Expressions, Revisited.

About a year ago, I wrote an article arguing for adding stacktrace support to F# computation expressions. Thanks to work by Lincoln Atkinson and Avi Avni, the upcoming F# 4.1 will be providing this functionality.

On the occasion of the release of VS2017 RC, I decided to spend some time adapting my continuation monad implementation to the new feature. The Bind implementation now looks as follows:

member __.Bind(f : Cont<'T>, g : 'T -> Cont<'S>,
                [<CallerMemberName>]?callerName : string,
                [<CallerFilePath>]?callerFilePath : string,
                [<CallerLineNumber>]?callerLineNumber : int) : Cont<'S> =

    fun sc ec ->
        let sc' (t : 'T) =
            match (try Ok(g t) with e -> Error e) with
            | Ok g -> g sc ec
            | Error e -> ec (SymbolicException.capture e)

        let ec' (se : SymbolicException) =
            let stackMsg =
                sprintf "   at %s in %s:line %d" 

            ec (SymbolicException.append stackMsg se)

        f sc' ec'

and ReturnFrom:

member __.ReturnFrom(f : Cont<'T>,
                        [<CallerMemberName>]?callerName : string,
                        [<CallerFilePath>]?callerFilePath : string,
                        [<CallerLineNumber>]?callerLineNumber : int) : Cont<'T> =
    fun sc ec ->
        let ec' (se : SymbolicException) =
            let stackMsg =
                sprintf "   at %s in %s:line %d" 

            ec (SymbolicException.append stackMsg se)

        f sc ec'

Running the odd/even example

let rec odd (n : int) = 
    cont {
        if n = 0 then return false
            return! even (n - 1)
and even (n : int) =
    cont {
        if n = 0 then return failwith "bug!"
            return! odd (n - 1)

odd 5 |> Cont.run

Now generates the stacktrace:

System.Exception: bug!
   at Program.even@119.Invoke(Unit unitVar) in C:\Users\eirik\devel\public\cont\Program.fs:line 119
   at Program.sc'@54-1.Invoke(a t) in C:\Users\eirik\devel\public\cont\Program.fs:line 54
   at odd in C:\Users\eirik\devel\public\cont\Program.fs:line 114
   at even in C:\Users\eirik\devel\public\cont\Program.fs:line 121
   at odd in C:\Users\eirik\devel\public\cont\Program.fs:line 114
   at even in C:\Users\eirik\devel\public\cont\Program.fs:line 121
   at odd in C:\Users\eirik\devel\public\cont\Program.fs:line 114
   at Program.ContModule.ec@102-1.Invoke(SymbolicException se) in C:\Users\eirik\devel\public\cont\Program.fs:line 102
   at Program.ContModule.run[T](FSharpFunc`2 cont) in C:\Users\eirik\devel\public\cont\Program.fs:line 103
   at <StartupCode$ConsoleApplication3>.$Program.main@() in C:\Users\eirik\devel\public\cont\Program.fs:line 106

The results look pretty promising. Looking forward to apply this technique to async and cloud builders!

You can find the full implementation here.

Reconciling Stacktraces with Computation Expressions, Revisited.

TypeShape: Practical Generic Programming in F#

Last week I announced a new library, TypeShape, with claims that it provides a practical way of doing generic programming in F#. I’m following up with this blog post to elaborate why I believe this to be genuinely useful, and how it could benefit the day-to-day life of the working .NET developer.

The pain of Reflection

Almost everybody who has worked with .NET will at some point need to dabble in the murky ways of reflection. Reflection is needed in scenaria where we need to access data in an indirect fashion, or where circumvention of the type system is necessary.

For example, assume that we have defined the following static method

type Foo =
    static member Bar<'T>(?optionalParam : 'T) : unit =
        printfn "Invoked with parameter %A" optionalParam

Assume now that we would like invoke that method, with a value whose type cannot be known at compile time. In other words, we want to define a function

val invokeUntyped : obj -> unit

which takes an input of type obj and invokes the generic method using the underlying type of the object instance. How do we do this? By using reflection of course!

open System.Reflection

let invokeUntyped (value:obj) =
    // Step 1: get the underlying System.Type for the value
    let t = value.GetType()
    // Step 2: locate the required method and apply the type argument
    let methodInfo =
            .GetMethod("Bar", BindingFlags.Public ||| BindingFlags.Static)
            .MakeGenericMethod [|t|]

    // Step 3: since the parameter is optional, it must be wrapped
    let optTy = typedefof<_ option>.MakeGenericType [|t|]
    let optCtor = optTy.GetConstructor [|t|]
    let optVal = optCtor.Invoke [|value|]

    /// Step 4: invoke the method with constructed optional parameter
    methodInfo.Invoke(null, [|optVal|]) :?> unit

This is cumbersome code to implement and is highly susceptible to breakage; even minor changes to the method signature will result in runtime errors. What’s more, reflection-based implementations are known to be significantly slower than their IL counterparts.

Using TypeShape

The TypeShape library can be used to implement the same functionality, but in a significantly safer and easy-to-read fashion:

open TypeShape

let invokeUntyped' (value:obj) =
    let shape = TypeShape.Create (value.GetType())
    shape.Accept { new ITypeShapeVisitor<unit> with
        member __.Visit<'T> () = Foo.Bar(value :?> 'T)}

Let’s have a look at the code, line by line.

The first line takes the underlying type of the input value and uses that to create an object of type TypeShape. This object encapsulates essential information on the type of the object.

The second line accepts an object expression of type ITypeShapeVisitor, which in turn invokes the method Foo.Bar. The second line is an instance of what is known as the visitor pattern, a design pattern commonly found in object-oriented programming. In this case, our visitor takes no arguments other than a type variable 'T. Passing this visitor to the TypeShape instance will have it invoked using the object type as argument, hence the downcast is expected to be successful. Importantly, the invocation is performed normally, thus any disagreement in the method signature will be picked up by the compiler.

In other words, TypeShape lets us introduce type variables into scope using the relatively concise approach of F# object expressions.

Nothing magical

The implementation of TypeShape is surprisingly simple to define:

open System

type ITypeShapeVisitor<'R> =
    abstract Visit<'T> : unit -> 'R

type TypeShape() =
    abstract Type : Type
    abstract Accept : ITypeShapeVisitor<'R> -> 'R

type TypeShape<'T>() =
    inherit TypeShape()
    override __.Type = typeof<'T>
    override __.Accept v = v.Visit<'T>()

type TypeShape with
    static member Create(t : Type) =
        let tsTy = typedefof<TypeShape<_>>.MakeGenericType [|t|]
        Activator.CreateInstance tsTy :?> TypeShape

In essence, TypeShape uses a minimal amount of reflection to bootstrap typed instances, then takes advantage of the ordinary .NET type system to access type information on-demand. TypeShape instances encapsulate and bear witness to types that may not be known at compile time.

Going Further

Let’s take a look at a different application: suppose we have a tuple whose precise type cannot be known at compile time. A common example of this is the object returned by the ShapeCombination active pattern in the F# quotations module. Suppose we would like like to extract either or both of the items contained in the tuple. Here’s how it could be done using reflection:

let extractTupleElements (value:obj) =
    let t = value.GetType()
    if not t.IsGenericType || t.GetGenericTypeDefinition() <> typedefof<_ * _> then
        invalidArg "value" "not a tuple type!"
    let m_Item1 = t.GetProperty("Item1")
    let m_Item2 = t.GetProperty("Item2")
    m_Item1.GetValue(value), m_Item2.GetValue(value)

Again, the same application could be simplified using the TypeShape library:

let extractTupleElements' (value : obj) =
    match TypeShape.Create (value.GetType()) with
    | Shape.Tuple2 (s : IShapeTuple2) ->
        s.Accept {
            new ITuple2Visitor<obj * obj> with
                member __.Visit<'T, 'S>() =
                    let t,s = value :?> 'T * 'S
                    box t, box s

    | _ -> invalidArg "value" "not a tuple type!"

In this case, we use the included Shape.(|Tuple2|_|) active pattern that checks against our shape being a 2-tuple. If successful, it returns an instance of type IShapeTuple2 that accepts a different visitor, ITuple2Visitor, which introduces the tuple element types in scope.

Similarly, here’s how we can check whether an unknown F# map contains a particular key:

let mapContainsKeyUntyped (key:obj) (map:obj) =
    match TypeShape.Create(map.GetType()) with
    | Shape.FSharpMap (s : IShapeFSharpMap) ->
        s.Accept {
            new IFSharpMapVisitor<bool> with
                member __.Visit<'K,'V when 'K : comparison> () =
                    (map :?> Map<'K,'V>).ContainsKey(key :?> 'K)

    | _ -> invalidArg "map" "not an F# map!"

Generic Programming

TypeShape active patterns can be used to orchestrate what could be considered as generic programming. For instance, take this value printer generator:

let rec mkPrinter<'T> () : 'T -> string = mkPrinterUntyped typeof<'T> :?> _
and private mkPrinterUntyped (t : Type) : obj =
    match TypeShape.Create t with
    | Shape.Unit -> box(fun () -> "()")
    | Shape.Bool -> box(sprintf "%b")
    | Shape.Int32 -> box(sprintf "%d")
    | Shape.String -> box(sprintf "\"%s\"")
    | Shape.FSharpOption s ->
        s.Accept {
            new IFSharpOptionVisitor<obj> with
                member __.Visit<'T> () =
                    let tp = mkPrinter<'T>()
                    box(function None -> "None" | Some t -> sprintf "Some (%s)" (tp t))

    | Shape.Tuple2 s ->
        s.Accept {
            new ITuple2Visitor<obj> with
                member __.Visit<'T, 'S> () =
                    let tp = mkPrinter<'T>()
                    let sp = mkPrinter<'S>()
                    box(fun (t : 'T, s : 'S) -> sprintf "(%s, %s)" (tp t) (sp s))

    | Shape.FSharpList s ->
        s.Accept {
            new IFSharpListVisitor<obj> with
                member __.Visit<'T> () =
                    let tp = mkPrinter<'T>()
                    box(fun ts -> ts |> List.map tp |> String.concat "; " |> sprintf "[%s]")

    | Shape.FSharpSet s ->
        s.Accept {
            new IFSharpSetVisitor<obj> with
                member __.Visit<'T when 'T : comparison> () =
                    let tp = mkPrinter<'T>()
                    box(fun (s:Set<'T>) -> s |> Seq.map tp |> String.concat "; " |> sprintf "set [%s]")

    | _ -> failwithf "unsupported type '%O'" t

The implementation can be used to generate printers for anything within the prescribed algebra of types:

let printer = mkPrinter<(bool * string) option * Set<int * string list option>>()

More importantly, any reflection code will only be executed at generation time, meaning that generated printers execute very efficiently:

// Real: 00:00:00.561, CPU: 00:00:00.562, GC gen0: 32, gen1: 0, gen2: 0
for i = 1 to 1000 do ignore <| sprintf "%A" value
// Real: 00:00:00.010, CPU: 00:00:00.000, GC gen0: 1, gen1: 0, gen2: 0
for i = 1 to 1000 do ignore <| printer value

This technique is being utilized in libraries such as FsPickler and FSharp.AWS.DynamoDB, and is an important contributor to their performance.


If your project relies heavily on reflection, you should consider giving TypeShape a try. It could improve readability and maintainability of your code, and may in some cases lead to better performance. Check it out at https://github.com/eiriktsarpalis/TypeShape and please submit your feedback and/or bug reports!

TypeShape: Practical Generic Programming in F#

Reconciling Stacktraces with Computation Expressions

Computation expressions are an excellent F# language feature. They outline the F# way of defining monads but are immensely more flexible. Applications include Asynchronous workflows, sequence/list comprehensions and our very own cloud workflows. Overall, it’s a feature that has seen lots of use (and misuse!) by the F# community. More importantly, it can all be done at the library level without needing to modify the language itself.

For a more in-depth introduction to computation expressions, I refer you to Scott Wlaschin’s excellent series on the subject.

The Problem

Computation expressions may be great, but they’re not without problems. One of the more common annoyances is exception stacktraces; generated expressions are desugared into nested lambda invocations, which means that their corresponding stacktraces are often unreadable.

To make matters worse, implementations such as async have their own exception handling logic implemented, which often leads to stacktraces being completely erased as a side-effect of this manipulation.

Let’s make the problem a bit more concrete by presenting an example:

let rec factorial n = async {
    if n = 0 then return failwith "bug!"
        let! pd = factorial (n - 1)
        return n * pd

Async.RunSynchronously(factorial 5)

If we attempt to execute the code we will be getting the following stacktrace (as of F# 4.0):

System.Exception: bug!
   at FSI_0008.factorial@128-2.Invoke(Unit unitVar) in C:\Users\eirik\Desktop\meta2.fsx:line 128
   at Microsoft.FSharp.Control.AsyncBuilderImpl.callA@851.Invoke(AsyncParams`1 args)
--- End of stack trace from previous location where exception was thrown ---
   at Microsoft.FSharp.Control.AsyncBuilderImpl.commit[a](Result`1 res)
   at Microsoft.FSharp.Control.CancellationTokenOps.RunSynchronously[a](CancellationToken token, FSharpAsync`1 computation, FSharpOption`1 timeout)
   at Microsoft.FSharp.Control.FSharpAsync.RunSynchronously[T](FSharpAsync`1 computation, FSharpOption`1 timeout, FSharpOption`1 cancellationToken)
   at <StartupCode$FSI_0009>.$FSI_0009.main@() in C:\Users\eirik\Desktop\meta2.fsx:line 135
Stopped due to error

Notice that method invocation information is missing: factorial was called recursively 5 times before the error occured. For reference, compare this against a native implementation of the same function:

let rec factorial n =
    if n = 0 then failwith "bug!"
        n * factorial(n - 1)

factorial 5

which gives the stacktrace

System.Exception: bug!
   at FSI_0013.factorial(Int32 n) in C:\Users\eirik\Desktop\meta1.fsx:line 38
   at FSI_0013.factorial(Int32 n) in C:\Users\eirik\Desktop\meta1.fsx:line 38
   at FSI_0013.factorial(Int32 n) in C:\Users\eirik\Desktop\meta1.fsx:line 38
   at FSI_0013.factorial(Int32 n) in C:\Users\eirik\Desktop\meta1.fsx:line 38
   at FSI_0013.factorial(Int32 n) in C:\Users\eirik\Desktop\meta1.fsx:line 38
   at FSI_0013.factorial(Int32 n) in C:\Users\eirik\Desktop\meta1.fsx:line 38
   at <StartupCode$FSI_0014>.$FSI_0014.main@() in C:\Users\eirik\Desktop\meta1.fsx:line 42
Stopped due to error

At first glance, this may seem like an insignificant shortcoming. In reality though, it can be the source to a lot of confusion. Asynchronous code tends interface with system APIs that are often sources of exceptions. In many real-world applications it is impossible to determine the source of an exception simply by looking at the generated stacktrace.

We need to improve on this; the goal of this post is to identify workarounds using the current state of F# and to propose potential additions for future releases of the language.

Defining a Continuation Monad

For the purposes of this exercise, we will need to define a continuation monad on which to build our improvements on.

/// Continuation workflow that accepts a pair of
/// success and exception continuations
type Cont<'T> = ('T -> unit) -> (exn -> unit) -> unit

/// return, the monadic unit
let ret (t : 'T) : Cont<'T> = fun sc _ -> sc t

/// monadic bind combinator
let (>>=) (f : Cont<'T>) (g : 'T -> Cont<'S>) : Cont<'S> =
    fun sc ec ->
        let sc' (t : 'T) =
            match (try Choice1Of2 (g t) with e -> Choice2Of2 e) with
            | Choice1Of2 g -> g sc ec
            | Choice2Of2 e -> ec e

        f sc' ec

Let’s now define our continuation builder:

type ContBuilder() =
    member __.Zero() = ret ()
    member __.Return t = ret t
    member __.Bind(f, g) = f >>= g
    member __.Delay(f : unit -> Cont<'T>) : Cont<'T> = ret () >>= f

let cont = new ContBuilder()

and a run function for executing our continuation workflows:

let run (cont : Cont<'T>) : 'T =
    let result = ref Unchecked.defaultof<'T>
    cont (fun t -> result := t) raise

This is a toy implementation that closely mimics the implementation of F# Async. As can be expected, we can observe the same stacktrace issue as in async.

Using Symbolic Stacktraces

Before we can move on, we need to give the definition of SymbolicException. This provides a way of appending symbolic entries to the stacktrace of an exception, which can then be re-raised as a regular .NET exception. This borrows from a technique also used here.

type SymbolicException =
        Source : Exception
        Stacktrace : string list

module SymbolicException =

    open System.Reflection

    /// clones an exception to avoid mutation issues related to the stacktrace
    let private clone (e : #exn) =
        let bf = new System.Runtime.Serialization.Formatters.Binary.BinaryFormatter()
        use m = new System.IO.MemoryStream()
        bf.Serialize(m, e)
        m.Position <- 0L
        bf.Deserialize m :?> exn

    let private remoteStackTraceField =
        let getField name = typeof<System.Exception>.GetField(name, BindingFlags.Instance ||| BindingFlags.NonPublic)
        match getField "remote_stack_trace" with
        | null -> getField "_remoteStackTraceString"
        | f -> f

    /// appens a line to the symbolic stacktrace
    let append (line : string) (se : SymbolicException) =
        { se with Stacktrace = line :: se.Stacktrace }

    /// Raises exception with its appended symboic stacktrace
    let raise (se : SymbolicException) =
        let e' = clone se.Source
        let stacktrace =
            seq { yield e'.StackTrace ; yield! List.rev se.Stacktrace }
            |> String.concat Environment.NewLine

        remoteStackTraceField.SetValue(e', stacktrace + Environment.NewLine)
        raise e'

    /// Captures an exception into a SymbolicException instance
    let capture (e : exn) = { Source = clone e ; Stacktrace = [] }

Symbolic Stacktraces & Computation Expressions

Now that we have established a technique for creating symbolic stacktraces, let’s see how we can make use of this in the context of computation expressions.

First, let’s examine the structure of a typical entry in a .NET stacktrace. Assuming our previous factorial example,

let rec factorial n =
    if n = 0 then failwith "bug!"
        n * factorial(n - 1)

The stacktrace entry corresponding to the recursive call appears like so:

at FSI_0013.factorial(Int32 n) in C:\Users\eirik\Desktop\meta1.fsx:line 38

This evidently contains two pieces of information:

  1. the method signature that was invoked.
  2. Source code location of the call site. This assumes presence of a corresponding symbols file.

Let’s see how we could recover similar data using expression builders. First off, recovering the invoked method name. Consider the following modification to the continuation builder:

type ContBuilder() =
    member __.Delay(f : unit -> Cont<'T>) : Cont<'T> =
        printfn "%O" (f.GetType())
        ret () >>= f

which once executed on factorial yields the output FSI_0026+factorial@124-11. This (somewhat) recovers the method signature of the factorial method.

In similar fashion, we can recover information on the call site by recovering the type name of the continuation parameter in Bind:

type ContBuilder() =
    member __.Bind(f : Cont<'T>, g : 'T -> Cont<'S>) : Cont<'S> =
        printfn "%O" (g.GetType())
        f >>= g

This yields FSI_0058+factorial@129-30, which (somewhat) recovers the location of the call site (containing method and line number).

Updating the Continuation Builder

Let’s now use this information to define our updated continuation builder. First, we need to update our continuation workflow type:

type Cont<'T> =
        Definition : Type option
        Body : ('T -> unit) -> (SymbolicException -> unit) -> unit

It has now been changed to thread symbolic exceptions in its exception continuation and to carry the defining type for the instance, if applicable. Let’s now continue with the builder definition:

type ContBuilder() =
    let protect f x = try Choice1Of2 (f x) with e -> Choice2Of2 e
    let mkCont def bd = { Body = bd ; Definition = def }

    member __.Return(t : 'T) = mkCont None (fun sc _ -> sc t)
    member __.Zero() = __.Return()
    member __.Delay(f : unit -> Cont<'T>) : Cont<'T> =
        let def = f.GetType()
        mkCont (Some def) (fun sc ec ->
            let sc' t =
                match protect f () with
                | Choice1Of2 g -> g.Body sc ec
                | Choice2Of2 e -> ec (SymbolicException.capture e)

            __.Zero().Body sc' ec)

The definition of Delay has been changed so that instances it defines are branded with a definition type. Continuing, let’s see now how we can make use of this metadata in the implementation of Bind:

type ContBuilder() =
    member __.Bind(f : Cont<'T>, g : 'T -> Cont<'S>) : Cont<'S> =
        mkCont None (fun sc ec ->
            let sc' (t : 'T) =
                match protect g t with
                | Choice1Of2 g -> g.Body sc ec
                | Choice2Of2 e -> ec (SymbolicException.capture e)

            let ec' (se : SymbolicException) =
                match f.Definition with
                | None -> ec se
                | Some def ->
                    let callSite = g.GetType()
                    let stackMsg = sprintf "   at %O in %O" def callSite
                    ec (SymbolicException.append stackMsg se)

            f.Body sc' ec')

In this updated definition of Bind, we attach an entry to the symbolic stacktrace of the exception continuation whenever binding to a “method”, that is an expression defined using Delay.

Finally, we give the updated definition of run:

let run (cont : Cont<'T>) =
    let result = ref Unchecked.defaultof<'T>
    let sc (t : 'T) = result := t
    let ec se =
        match cont.Definition with
        | None -> SymbolicException.raise se
        | Some def ->
            let stackMsg = sprintf "   at %O in Cont.run" def
            se |> SymbolicException.append stackMsg |> SymbolicException.raise

    cont.Body sc ec

Let’s now verify that the implementation works by re-running factorial:

System.Exception: bug!
   at FSI_0058.factorial@126-29.Invoke(Unit unitVar) in C:\Users\eirik\Desktop\meta2.fsx:line 126
   at FSI_0047.ContBuilder.protect[b,c](FSharpFunc`2 f, b x) in C:\Users\eirik\Desktop\meta2.fsx:line 54
   at FSI_0058+factorial@126-29 in FSI_0058+factorial@129-30
   at FSI_0058+factorial@126-29 in FSI_0058+factorial@129-30
   at FSI_0058+factorial@126-29 in FSI_0058+factorial@129-30
   at FSI_0058+factorial@126-29 in FSI_0058+factorial@129-30
   at FSI_0058+factorial@126-29 in FSI_0058+factorial@129-30
   at FSI_0078+factorial@132-31 in Cont.run
   at FSI_0022.SymbolicExceptionModule.raise[a](SymbolicException se) in C:\Users\eirik\Desktop\meta2.fsx:line 40
   at FSI_0060.run[T](Cont`1 cont) in C:\Users\eirik\Desktop\meta2.fsx:line 101
   at <StartupCode$FSI_0061>.$FSI_0061.main@() in C:\Users\eirik\Desktop\meta2.fsx:line 124

It is left as an exercise to the reader to improve formatting of symbolic stack entries by performing further parsing of the internal type names.

Implementing ReturnFrom

Let’s now have a go at implementing ReturnFrom, the tail call keyword. By virtue of the its type signature, the call site location cannot be recovered in this context:

type ContBuilder() =
    member __.ReturnFrom (f : Cont<'T>) =
        match f.Definition with
        | None -> f
        | Some df ->
            { f with Body = fun sc ec ->
                    let ec' (se : SymbolicException) =
                        let stackMsg = sprintf "   at %O" df
                        ec (SymbolicException.append stackMsg se)

                    f.Body sc ec' }

Again, we verify by running

let rec odd (n : int) = 
    cont {
        if n = 0 then return false
            return! even (n - 1)

and even (n : int) =
    cont {
        if n = 0 then return failwith "bug!"
            return! odd (n - 1)

odd 5 |> Cont.run

which yields

System.Exception: bug!
   at FSI_0011.even@149-3.Invoke(Unit unitVar) in C:\Users\eirik\Desktop\meta2.fsx:line 149
   at FSI_0002.ContBuilder.protect[a,b](FSharpFunc`2 f, a x) in C:\Users\eirik\Desktop\meta2.fsx:line 54
   at FSI_0011+even@149-3
   at FSI_0011+odd@142-3
   at FSI_0011+even@149-3
   at FSI_0011+odd@142-3
   at FSI_0011+even@149-3
   at FSI_0011+odd@142-3 in Cont.run
   at FSI_0002.SymbolicExceptionModule.raise[a](SymbolicException se) in C:\Users\eirik\Desktop\meta2.fsx:line 40
   at FSI_0002.Cont.run[T](Cont`1 cont) in C:\Users\eirik\Desktop\meta2.fsx:line 111
   at <StartupCode$FSI_0011>.$FSI_0011.main@()
Stopped due to error


The above approach has few clear drawbacks

  1. Information relayed to symbolic stacktraces are of low quality, compared to .NET stacktraces. Ideally we would like to recover the MethodInfo that corresponds to the call, but that does not appear to be possible.
  2. The implementation makes light use of reflection at places. This may affect performance unless care is taken, and could hurt compatibility with certain platforms.
  3. Impossible to distinguish between actual method calls and inlined expressions. This can result in noisy stacktraces at places.

Future Directions

The issues discussed above beg the question: what could be added to the F# language itself to improve support for stacktraces? It’s evident that the required information is known by the compiler. So here’s a proposal:

First, let’s outline a type that encodes an entry from a stacktrace:

type Location = { File : string ; Line : int ; Column : int }
type MethodInvocationInfo = { Method : MethodInfo ; Location : Location }

Now, consider the following hypothetical computation expression implementation.

type Cont<'T> = ('T -> unit) -> (SymbolicException -> unit) -> unit

type ContBuilder() =
    let protect f x = try Choice1Of2 (f x) with e -> Choice2Of2 e
    let fmt (m : MethodInvocationInfo) = sprintf "   at %O in %s:line %d" m.MethodInfo m.Location.File m.Location.Line 
    member __.Bind(data : MethodInvocationInfo, f : Cont<'T>, g : 'T -> Cont<'S>) : Cont<'S> =
        fun sc ec ->
            let sc' (t : 'T) =
                match protect g t with
                | Choice1Of2 g -> g sc ec
                | Choice2Of2 e -> ec (SymbolicException.Capture e)

            let ec' (e : SymbolicException) = ec (e.Append (fmt data))

            f sc' ec'

    member __.ReturnFrom(data : MethodInvocationInfo, f : Cont<'T>) =
        fun sc ec ->
            let ec' (e : SymbolicException) = ec (e.Append (fmt data))
            f sc ec'

Under the proposed change, it should be possible to define overloads for Bind, Return, ReturnFrom and Yield that accept an additional method invocation metadata parameter. These would be used by the compiler only in cases where corresponding keywords involved a method invocation on the right-hand side. In other words, the line

let! x = async { return 1 }

would desugar to a call to the standard Bind overload, whereas

let f i = async { return i }
let! x = f 1

would translate to the new overload, passing metadata on the invocation of function f at compile time.


The ability to generate exception stacktraces in computation expressions is a feature sorely missed in F# computation expressions. Adopting improvements towards this direction would greatly improve the debugging experience of asynchronous workflows and other computation expressions, including mbrace. It would be great to see such improvements added to the next release of F#.

Reconciling Stacktraces with Computation Expressions

Deploying .NET code instantly using Vagabond


This is post #30 of the English 2014 F# advent calendar. Thanks to Sergey Tihon for inviting me and suggesting the topic. Be sure to check out all the other awesome posts! In this post I will be describing Vagabond, a dependency management library of mine. I will try to walk through some of the subtleties in the .NET framework that drove the library implementation, in an attempt to make it more accessible. All code and examples presented can be found in this repository.


It is often claimed in functional programming that functions are treated as values. This is a valid assumption in simple applications, permitting powerful patterns when writing code. But is this a general truth, or simply a linguistic abstraction? If functions really are values, one would expect that they exhibit all properties normally expected of values. For instance, it should be possible to serialise functions and transmit to a remote process for use. But wait, does serialising a function even make sense? In principle it should, we know for a fact that all code is essentially data. But the answer really depends on language/runtime support, serialization facilities and how lambdas are represented. Our goal here is to consider the .NET framework and F# in particular. How does F#/.NET represent lambdas? We’ll attempt to illustrate by example.

‘Thunk Server’

Our goal is to define a ‘thunk server’, an actor implementation that receives arbitrary thunks (lambdas of type unit -> 'T) which it executes replying with the result. For this exercise we will be using Thespian, a distributed actor framework for F#. The choice of actor framework is not particularly important; Thespian should be easy to grasp however, since its programming model closely follows that of the F# MailboxProcessor. We start off with the simplest imaginable conception of how a thunk server should be implemented:

open Nessos.Thespian

// Actor API: a thunk and reply channel in which to send the result
type ThunkMessage = (unit -> obj) * IReplyChannel<Choice<obj, exn>>

// actor body definition
let rec serverLoop (self : Actor<ThunkMessage>) : Async<unit> =
    async {
        // receive next message
        let! thunk, reply = self.Receive()
        // execute, catching any exception
        let result : Choice<obj, exn> =
            try thunk () |> Choice1Of2
            with e -> Choice2Of2 e
        // reply
        do! reply.Reply result
        return! serverLoop self

// initialize an actor instance in the current process
let server : Actor<ThunkMessage> = Actor.Start "thunkServer" serverLoop

We can now interface with our thunk server like so:

/// submit a thunk for evaluation to target actor ref
let evaluate (server : ActorRef<ThunkMessage>) (thunk : unit -> 'T) : 'T =
    // post to actor and wait for reply
    let result = server <!= fun replyChannel -> (fun () -> thunk () :> obj), replyChannel
    // downcast if value, raise if exception
    match result with
    | Choice1Of2 o -> o :?> 'T
    | Choice2Of2 e -> raise e

evaluate server.Ref (fun () -> 1 + 1) // returns 2

But is this implementation correct? Remember, our goal is to submit lambdas to a remote process for execution. Using the companion project we can test precisely this scenario:

// spawn a local windowed process that runs the actor
let server : ActorRef<ThunkMessage> = spawnWindow ()

Let’s attempt to submit some code to the server:

// load a third party library and submit code for evaluation
#r "ThirdPartyLibrary.dll"
evaluate server ThirdPartyLibrary.thirdPartyLambda

Evaluation fails with the following error:

System.IO.FileNotFoundException: Could not load file or assembly 'ThirdPartyLibrary, Version=, Culture=neutral, PublicKeyToken=null' or one of its dependencies. The system cannot find the file specified.
   at System.Reflection.RuntimeAssembly.GetType(RuntimeAssembly assembly, String name, Boolean throwOnError, Boolean ignoreCase, ObjectHandleOnStack type)
   at System.Reflection.RuntimeAssembly.GetType(String name, Boolean throwOnError, Boolean ignoreCase)
   at Nessos.FsPickler.ReflectionCache.loadMemberInfo(FSharpOption`1 tyConv, FSharpFunc`2 loadAssembly, FSharpFunc`2 getMethodSignature, CompositeMemberInfo mI)

Why does this happen? Because lambdas in F# are not exactly values. Rather, they are instances of compiler generated classes that inherit the abstract type FSharpFunc. Needless to say, a lambda cannot be successfully deserialised in a process unless its particular underlying class is loaded in the current application domain.

Fixing the problem

So how can we fix the problem? A popular approach is to distribute code using expression trees. Expression trees are clearly values, hence can be easily serialised and evaluated in the target machine. However, this is not a satisfactory solution: expression trees do not support type definitions; evaluation of expression trees is not always efficient; expression trees can still reference code in out-of-reach assemblies. We need to to better than this! Our experience illustrates that a proper serialisation format for lambdas must also include their dependent assemblies. So let’s try to do precisely that:

open System.IO
open System.Reflection

// Raw assembly container
type AssemblyPackage =
        FullName : string
        AssemblyImage : byte []

// transitively traverse assembly dependencies for given object
let gatherDependencies (object:obj) : AssemblyPackage list =
    let rec aux (gathered : Map<string, Assembly>) (remaining : Assembly list) =
        match remaining with
        // ignored assembly
        | a :: rest when gathered.ContainsKey a.FullName || a.GlobalAssemblyCache -> aux gathered rest
        // came across new assembly, add to state and include transitive dependencies to inputs
        | a :: rest ->
            let dependencies = a.GetReferencedAssemblies() |> Seq.map Assembly.Load |> Seq.toList
            aux (Map.add a.FullName a gathered) (dependencies @ rest)
        // traversal complete, create assembly packages
        | [] ->
            |> Seq.map (fun (KeyValue(_,a)) ->
                                { FullName = a.FullName
                                  AssemblyImage = File.ReadAllBytes a.Location })
            |> Seq.toList

    aux Map.empty [object.GetType().Assembly]

// loads raw assemblies in current application domain
let loadRawAssemblies (pkgs : AssemblyPackage list) =
    pkgs |> List.iter (fun pkg -> Assembly.Load pkg.AssemblyImage |> ignore)

The code above uses reflection to traverse the assembly dependencies for a given object. The raw binary assembly images are then read from disk to be submitted and loaded by the remote recipient. We can now revise our thunk server implementation as follows:

// our actor API is augmented with an additional message
type ThunkMessage =
    | RunThunk of (unit -> obj) * IReplyChannel<Choice<obj, exn>>
    | LoadAssemblies of AssemblyPackage list

let rec serverLoop (self : Actor<ThunkMessage>) : Async<unit> =
    async {
        let! msg = self.Receive()
        match msg with
        | RunThunk(thunk, reply) ->
            let result : Choice<obj, exn> =
                try thunk () |> Choice1Of2
                with e -> Choice2Of2 e

            do! reply.Reply result
        | LoadAssemblies assemblies ->
            loadRawAssemblies assemblies

        return! serverLoop self

/// submit a thunk for evaluation to target actor ref
let evaluate (server : ActorRef<ThunkMessage>) (thunk : unit -> 'T) =
    // traverse and upload dependencies
    server <-- LoadAssemblies (gatherDependencies thunk)
    // assembly upload complete, send thunk for execution
    let result = server <!= fun replyChannel -> RunThunk ((fun () -> thunk () :> obj), replyChannel)
    match result with
    | Choice1Of2 o -> o :?> 'T
    | Choice2Of2 e -> raise e

And we are done! We can use the companion project to verify that this indeed resolves the previously failing deployment scenario. Clearly, this should cover the case of assemblies missing from the remote process. But does it work with lambdas defined in F# interactive? Let’s try it out:

> evaluate localServer (fun () -> 1 + 1);;
System.NotSupportedException: The invoked member is not supported in a dynamic assembly.
   at System.Reflection.Emit.InternalAssemblyBuilder.get_Location()
   at ThunkServer.Naive2.aux@36-1.Invoke(KeyValuePair`2 _arg1)

Huh? What’s a dynamic assembly? And why does it appear here? Reading from MSDN:

(…) a set of managed types in the System.Reflection.Emit namespace that allow a compiler or tool to emit metadata and Microsoft intermediate language (MSIL) at run time (…)

Among such compilers is F# interactive, which uses a single dynamic assembly for emitting all code defined in a single session. This can be verified by writing

> (fun i -> i + 1).GetType().Assembly ;;
val it : System.Reflection.Assembly =
  FSI-ASSEMBLY, Version=, Culture=neutral, PublicKeyToken=null
    {IsDynamic = true;
     Location = ?;}

We have reached a dead end, it seems. Clearly, dynamic assemblies cannot be distributed to remote processes. And even if they could, we would face more problems on account of dynamic assemblies expanding over time. What is there to do?

Enter Vagabond

Vagabond is an automated dependency management framework. It is capable of converting dynamic assemblies into standalone, distributable assemblies. This is achieved using Mono.Cecil and a modification of AssemblySaver, a parser for dynamic assemblies.

Basic usage

Having worked on the above examples, the Vagabond API should come off as being natural. We can begin using the library by initialising aVagabond state object:

open Nessos.Vagabond

let vagabond : Vagabond = Vagabond.Initialize(cacheDirectory = "/tmp/vagabond")

This will be our entry point for all interactions with the library. For starters, let us compute the dependencies for a lambda defined in F# interactive:

> let deps = vagabond.ComputeObjectDependencies((fun i -> i + 1), permitCompilation = true) ;;
val deps : System.Reflection.Assembly list =
  [FSI-ASSEMBLY_17a1c27a-5c8f-4acb-ae1a-5aea74027854_1, Version=, Culture=neutral, PublicKeyToken=null]

This will return a single dependent assembly that is non-dynamic, generated by Vagabond. The numerical suffix indicates that this is the first slice produced of that particular dynamic assembly. Vagabond uses a slicing scheme to address the issue of dynamic assemblies that expand over time. To create an exportable assembly package, we simply call

let pkg : AssemblyPackage = vagabond.CreateAssemblyPackage(assembly, includeAssemblyImage = true)

Packages can be loaded to the current application domain using the same object:

let result : AssemblyLoadInfo = vagabond.LoadAssemblyPackage(pkg)

As soon as assembly dependencies have been uploaded to the recipient, communication can be established. Importantly, this must be done using the serialiser instance provided by the Vagabond object.

vagabond.Serializer : FsPicklerSerializer

This makes use of purpose-built FsPickler functionality to bridge between local dynamic assemblies and exported slices.

Putting it all together

Let’s now try to correctly implement our thunk server. We first define our updated actor body:

type ThunkMessage =
    // Evaluate thunk
    | RunThunk of (unit -> obj) * IReplyChannel<Choice<obj, exn>>
    // Query remote process on assembly load state
    | GetAssemblyLoadState of AssemblyId list * IReplyChannel<AssemblyLoadInfo list>
    // Submit assembly packages for loading at remote party
    | UploadAssemblies of AssemblyPackage list * IReplyChannel<AssemblyLoadInfo list>

let rec serverLoop (self : Actor<ThunkMessage>) : Async<unit> =
    async {
        let! msg = self.Receive()
        match msg with
        | RunThunk(thunk, reply) ->
            let result : Choice<obj, exn> =
                try thunk () |> Choice1Of2
                with e -> Choice2Of2 e

            do! reply.Reply result
        | GetAssemblyLoadState(assemblyIds, reply) ->
            // query local vagabond object on load state
            let info = vagabond.GetAssemblyLoadInfo assemblyIds
            do! reply.Reply info

        | UploadAssemblies(pkgs, reply) ->
            // load packages using local vagabond object
            let results = vagabond.LoadAssemblyPackages(pkgs)
            do! reply.Reply results

        return! serverLoop self

/// create a local thunk server instance with given name
let createServer name = Actor.Start name serverLoop |> Actor.ref

It remains to define the client side of the assembly upload logic. To make sure that assemblies are not needlessly uploaded, we could either implement our own upload protocol using the previously described Vagabond API, or we could simply use the built-in protocol and only specify the communication implementation:

/// submit a thunk for evaluation to target actor ref
let evaluate (server : ActorRef<ThunkMessage>) (thunk : unit -> 'T) =
    // receiver implementation ; only specifies how to communicate with remote party
    let receiver =
            new IRemoteAssemblyReceiver with
                member __.GetLoadedAssemblyInfo(ids: AssemblyId list) =
                    server.PostWithReply(fun reply -> GetAssemblyLoadState(ids, reply))

                member __.PushAssemblies (pkgs: AssemblyPackage list) =
                    server.PostWithReply(fun reply -> UploadAssemblies(pkgs, reply))

    // submit assemblies using the receiver implementation and the built-in upload protocol
    vagabond.SubmitObjectDependencies(receiver, thunk, permitCompilation = true)
    |> Async.RunSynchronously
    |> ignore

    // dependency upload complete, send thunk for execution
    let result = server <!= fun replyChannel -> RunThunk ((fun () -> thunk () :> obj), replyChannel)
    match result with
    | Choice1Of2 o -> o :?> 'T
    | Choice2Of2 e -> raise e

Using the companion project we can now test our implementation. The examples below all work in F# interactive.

// spawns a windowed console application that hosts a single thunk server instance
let server : ActorRef<ThunkMessage> = ThunkServer.spawnWindow()

evaluate server (fun () -> 1 + 1)
evaluate server (fun () -> printfn "Remote side-effect")
evaluate server (fun () -> do failwith "boom!")

Deploying actors

An application of particular interest is the ability to remotely deploy actor definitions straight from F# interactive, just by using our simple thunk server:

// deploy an actor body remotely using thunk server
let deployActor name (body : Actor&amp;amp;lt;'T&amp;amp;gt; -&amp;amp;gt; Async&amp;amp;lt;unit&amp;amp;gt;) : ActorRef&amp;amp;lt;'T&amp;amp;gt; =
    evaluate server (fun () -&amp;amp;gt; let actor = Actor.Start name body in actor.Ref)

Let’s try this out by implementing a simple counter actor. All type definitions and logics can be declared straight from F# interactive.

type Counter =
    | IncrementBy of int
    | GetCount of IReplyChannel<int>

let rec body count (self : Actor<Counter>) = async {
    let! msg = self.Receive()
    match msg with
    | IncrementBy i -> return! body (count + i) self
    | GetCount rc ->
        do! rc.Reply count
        return! body count self

// deploy to thunk server, receive remote actor ref
let ref : ActorRef<Counter> = deployActor "counter" (body 0)

We can now test the deployment simply by interfacing with the actor ref we have received. Side effects can be added in the actor body to verify that code indeed runs in the remote window.

ref <-- IncrementBy 1
ref <-- IncrementBy 2
ref <-- IncrementBy 3

ref <!= GetCount // 6

Further Applications

Vagabond enables on-the-fly deployment of code, be it to a collocated process, your next-door server or an Azure-hosted cluster. It is applicable not only to F# and its shell, but should work with any .NET language/REPL such as the Roslyn-based scriptcs. The MBrace framework already makes use of the library, enabling instant deployment of distributed algorithms to the cloud. We plan to incorporate Vagabond functionality with Thespian as well. Vagabond is a powerful library from which most distributed frameworks running on .NET could benefit. I hope that today’s exposition will encourage a more widespread adoption.

Deploying .NET code instantly using Vagabond