Effective F#, tips and tricks

Architecture

Use Onion architecture
- Dependencies go inwards. That is, the Core domain doesn't know about outside layers
Use pipeline model to implement workflows/use-cases/stories
- Business logic makes decisions
- IO does storage with minimal logic
- Keep Business logic and IO separate
- Keep IO at edges
Testing
- Only need to unit test the business logic
- IO is tested as part of integration tests

High-level design

It's good to switch between top-down (decomposition) and bottom up coding.

Top down: Decide on the inputs and outputs first, then implement based on that. Then break that down into smaller functions than can be composed in a pipeline.
Bottom up: write functions the way they should be, as independent things, then fit them together using "adapters" such as bind, map etc
To parameterise behavior, pass in a function!

Grouping code:

Types that change together should live together (e.g. in same module)
Functions that change together should live together (e.g. in same module)

Make illegal states unrepresentable:

use state machines to avoid using option for data which is not used in a particular state
avoid bools
when values change together, group them together (e.g a option and b option becomes (a*b) option)

Setting up a project

Compiler settings: warn unused "-warnon:1182"

Modules and namespaces

A namespace is just like C#
A "module" is like a C# class with only static methods

If defining types only (eg domain), use a namespace. If defining functions, they MUST be in a module (otherwise you get a compiler error)

For functions closely associated with a type (eg create/value), use a module with the same name as the type.

NOTES:

You can't use namespaces in F# interactive
You can use "rec" for recursive modules so that the definitions don't have to be in order.

TIP: Open few modules:

Use qualified names so that readers know what you are referencing. Don't force people to remember things! Eg List.map rather than open List and then using unqualified map later in the file.
You can force this with the [<RequireQualifiedAccess>] attribute on a module!
Also, open modules as close as possible to where they are used (e.g. in a submodule rather than the top of the file)

Signature files

have fsi file extension.
must be above the corresponding fs file.

|             | FSI              | FS          |
| top of file | namespace/module | same as FSI |
| types       | public           | public types must be defined the same |
|             |                  | private types OK |
| functions   | use "val"        | use "let"        |

Example signature file

/// Signature file
module MyDomain.Dto.Converter

open MyDomain

type DtoError =
    | BadName

module MyEntity =
    val fromDomain : MyEntity -> MyEntityDto
    val toDomain : MyEntityDto -> Result<MyEntity,DtoError>

Example implementation file

module MyDomain.Dto.Converter

open MyDomain

type DtoError =
    | BadName

type MyPrivateType = string

module MyEntity =
    let fromDomain (x:MyEntity) :MyEntityDto =
        failwith "not implemented"

    let toDomain (x:MyEntityDto) :Result<MyEntity,DtoError> =
        failwith "not implemented"

Organising modules - general

F# is order sensitive -- modules only know about modules above them. This maps nicely to the Onion architecture

core domain is at top
IO near bottom
Main app (composition root/Startup) is always last

Example module structure:

Common libraries
Domain types
Workflows/use-cases
DTOs
DB/Infrastructure
API interfaces
Composition Root that combines everything and exposes the API.

You can use dots in the filename to make the namespace clear:

Common.PrimitiveTypes
Common.Types
MyWorkflow.Types
MyWorkflow.Impl

For example: https://github.com/swlaschin/DomainModelingMadeFunctional/blob/master/src/OrderTaking/OrderTaking.fsproj

Working with multiple assemblies

If you want to swap out or hide infrastructure, you can use multiple assembles. For example:

MyDomain.Core assembly

Common lib
Domain types
Workflows (with infrastructure dependencies injected in later)
API

MyDomain.Azure assembly

DB/Infrastructure for Azure

MyDomain.AWS assembly

DB/Infrastructure for AWS

MyDomain.Web assembly (top-level assembly)

Composition Root that combines everything and exposes the API

Code formatting

Reading happens much more than writing, so make it readble first. Make the INTENT understandable During a code review if reader and writer disagree, reader is always right!

Make blocks obvious style - have only one indent per scope change.

// good
let xxxx xx xx =
  xxxx xx xx
  xxx xx 
  xxxx xxx
  xxxx xxx
  xxxx xx xxxx

// badly indented blocks
let xxxx xx xx =
  xxxx xx xx
    xxx xx 
   xxxx xxx
        xxxx xxx
     xxxx xx xxxx

Function style

Two styles that I use:

"stanza" style. Code is divided into related groups like stanzas in poetry.
"recipe" style. Helper functions are defined (the "ingredients") and then combined in a pipeline.

Here's a "stanza" style function:

// "stanza" style
let myFunction xxx = 
  // part 1
  do this 
  andthen this 
  andthen something else 
  
  // part 2
  start a new thing 
  and some more

  // part 3
  ...

I believe a long function is perfectly fine if everything belongs together -- don't split it up for the sake of it

Here's a "recipe" style function:

// "recipe" style
let myFunction xxx = 
  // assemble ingredients
  let helperA = ....
  let helperB = ...
  let helperC = ...
  
  // then combine the ingredients
  start
  |> helperA
  |> helperB
  |> helperC

If you have a "big recipe" when the "ingredients" are more than one liners, then consider creating a private helper module to put the "ingredients" in.

module private Ingredients =
  // define ingredients
  let helperA = 
    ...
  let helperB = 
    ...
  let helperC = 
    ...

// bring module into scope
open Ingredients

// public  
let myFunction xxx xxx = 
  // combine the ingredients defined above
  start
  |> helperA
  |> helperB
  |> helperC

If you're using signature files and only one workflow per file, the "private" module is the whole fs file, so this approach is not needed.

Format to make diffs more readable

Diffs are an important part of reading so make diffs easy to understand. Changing one thing should cause only one line to change!

// 1 or 2 fields
type MyRecord = { fieldA:string; fieldB:int }

// 3 or more fields
type MyRecord = {  
   fieldA : string
   fieldB : string
   fieldC : string
   fieldD : string
   }

Use same style for constructing records

let myRecord = {  
   fieldA = "..."
   fieldB = "..."
   fieldC = "..."
   fieldD = "..."
   }

// enum-style choices
type Colour = Red | Blue | Green

// single case choices
type CustomerId = CustomerId of int

// complex choices   
type Payment = 
   | Cash
   | Check of CheckNumber
   | Card of CardInfo

Formatting match expressions

Line up the vertical bars under the match:

let xxx p = 
  match p with
  | CaseA z -> handler...
  | CaseB z -> handler...
  | CaseC z -> handler...

When the "handlers" for each case become longer than a one-liner, put each handler in a new block underneath:

let xxx p = 
  match p with
  | CaseA z -> 
	handler...
  | CaseB z -> 
	handler...
  | CaseD z -> 
	handler...

Formatting inline lambdas

Short ones can go on one line

aCollection
|> List.map (fun x -> x + 1)

Longer ones may need to start a new block

aCollection
|> List.map (fun x -> 
   xxxxx xxxx
   xxxx xxx
   )

Do NOT have "hanging" lambdas. If the top line changes, the indentation of the entire block will change, breaking the "diff" rule above.

// example of BAD indenting. 
// If List.map is changed to List.choose, say, the entire block changes which messes up the diff
aCollection
|> List.map (fun x -> 
             xxxxx xxxx
             xxxx xxx
             )

Formatting Generic types

F# allows generic types to be used as a suffix. Eg 'a list or List<'a>.

Which one to use? Answer: Use suffix form for built-in types such as int list, int seq, int option, int[]. For all others, use C# style form (e.g. Result<T,U>).

Coding tips

Use exhaustive pattern matching

Always be explicit and match every pattern
Avoid using _ as a wildcard pattern. Exception needed for matching strings and ints, where you often have to.
If you do this, the compiler will warn you when new cases are added. If you don't do this, you will never know!

Avoid boilerplate

Often there are two similar bits of code but slightly different. In this case, it's easy to ignore subtle differences.

FIX: parameterize with a function parameter! This forces differences to be made explicit:

Example 1: iterating (aka fold) https://youtu.be/E8I19uA-wGY?t=1100
Example 2: continuations https://youtu.be/E8I19uA-wGY?t=1966

Make common errors obvious

Return a Result or Option rather than an exception.

But don't use Result for everything!

There is a difference between "domain errors" and "panics". Panics do not need to be exposed in the domain. Throw an exception and catch at a top-level function.
Exceptions can be better than Result if the scope is clear and very local (e.g. drilling down in a tree and throwing to exit in the middle of iteration)

If you DO throw exceptions as part of the API, make it clear in the name of the function:

List.tryHead // ok
List.headExn // ok
List.head    // bad (sadly this is how the standard library does it

Similarly

List.find      // bad unless you KNOW that the item exists
List.tryFind   // good

Complex type hacks

Don't do it!

F# has Statically Resolved Type Parameters (SRTP) which can be used to do polymorphism, monads, etc. Don't!

KISS - simple is better than complex!

Mutability

Mutable values are OK if they are local and no one sees them. But don't go overboard! Purity is one goal but it's not the only goal.

Operators

Avoid importing operators from other modules. It will be hard for a reader to know where they came from. Instead define operators at the top of module they're used in (or even just in the function they are used in).

module xxxxx 

// define at top
let ( <!> ) = Validation.map
let ( <*> ) = Validation.apply

// use
ctor <!> firstParam <*> secondParam

OO code vs FP code

It's OK to use OO style code if behavior is the most important thing -- that is, you want polymorphism.

e.g. in x.ToString() we don't care what x is. To define methods see https://fsharpforfunandprofit.com/posts/type-extensions/

But this can mess with type inference :( You might well need to use type annotations more often. See https://fsharpforfunandprofit.com/posts/type-extensions/#methods-dont-play-well-with-type-inference

Similarly, it's OK to use interfaces to define groups of functions that can have multiple implementations. See https://fsharpforfunandprofit.com/posts/interfaces/

Documentation on doing OO in F# is here: https://fsharpforfunandprofit.com/series/object-oriented-programming-in-fsharp.html

Working with Dictionaries

module DictionaryExample =
    // needed to reference IDictionary
    open System.Collections.Generic

    // create a dictionary from a list of pairs
    let myDict = [ (1,"a"); (2,"b") ] |> dict

    /// get a value from a dictionary, with dictionary as LAST parameter
    let tryGetValue key (dict:IDictionary<_,_>) =
        match dict.TryGetValue(key) with
        | true, value -> Some value
        | false,_ -> None

    // this style is used when piping the dictionary into a key check
    myDict |> tryGetValue 1
    myDict |> tryGetValue 42


    /// get a value from a dictionary, with dictionary as FIRST parameter
    let tryGetValue2 (dict:IDictionary<_,_>) key =
        match dict.TryGetValue(key) with
        | true, value -> Some value
        | false,_ -> None

    // this style is used when you want to bake in the dictionary
    // in a helper function and then pass the keys in later
    let lookup = tryGetValue2 myDict
    lookup 1
    lookup 42

Working with the Result type

If using Result, include this Result.fs file at the top of your project.

I am fine with duplicating this file in many projects to avoid dependencies on nuget or library DLLs.

To chain results in series ("monads")

If the first value is not a Result:

myValue
|> pointsFunctionA
|> Result.bind pointsFunctionB

If the first value is a Result:

myResult
|> Result.bind pointsFunctionA
|> Result.bind pointsFunctionB

Result computation expressions instead of bind

You can use result computation expressions instead of Result.bind:

let finalResult = result {
    let! x = myResult
    let! y = pointsFunctionA x
    let! z = pointsFunctionB y
    return z
    }

To combine results in parallel ("applicatives")

Use the Validation type in Result.fs. It's the same type as Result but with a list of errors.

If need to do validation (multiple errors), the validation code has the same pattern

Define a ctor for the type
Create all the values (which return Results)
Option 1: Use the applicative style to construct the object.
- First param has <!> in front
- Subsequent params have <*> in front
Option 2: Lift the ctor using lift2, lift3, lift4` etc., depending on how many parameters the ctor has.
When validation is complete and you are returning to normal code, consider mapping the list of validation errors to an error case so you have a normal Result again.

// Option 1: Applicative style

// define at top of file
let ( <!> ) = Validation.map
let ( <*> ) = Validation.apply

type MyError = 
   | ValidationError of string list
   | DbError of string 
   | etc

let createMyObject param1 param2 param3 = 
   { Field1 = param1; etc ... }
let param1OrError =  ...validate and return a Result
let param2OrError =  ...validate and return a Result
let param3OrError =  ...validate and return a Result
let myObjectOrError = createMyObject <!> param1OrError <*> param2OrError <*> param3OrError

// convert to a Result<_,MyError> without a list of errors
myObjectOrError |> Result.mapError ValidationError

// Option 2: Using lift

type MyError =  ...

let createMyObject param1 param2 param3 = 
   { Field1 = param1; etc ... }
let param1OrError =  ...validate and return a Result
let param2OrError =  ...validate and return a Result
let param3OrError =  ...validate and return a Result
let myObjectOrError = (Validation.lift3 createMyObject) param1OrError param2OrError param3OrError

// convert to a Result<_,MyError> without a list of errors
myObjectOrError |> Result.mapError ValidationError

Validating lists of items

To validate a list, validate each item using List.map and then use Result.sequence to put the Result type on the outside.

let itemsOrError =
    itemDtos
    // convert to F# list
    |> List.ofArray
    // convert each DTO item to a domain object
    |> List.map (ItemDto.toDomain)
    // flip from List<Result<>> to Result<List<>>
    |> Result.sequence

Constrained types

Define a type with a private constructor and then a helper module (in same scope) with same name. Helper module should have create and value functions

type EmailAddress = private EmailAddress of string

module EmailAddress =
  let create str :Result<EmailAddress,_> =  // or Option
    if String.IsNullOrEmpty(str) then
      Error ...
    else if (* check validity *) then
       Ok (EmailAddress str)
    else
       Error ...

  let value (EmailAddress str) = str

DTOs and Validation

Define them in their own module. For each DTO, define an associated module with functions "toDomain" and "fromDomain"

namespace MyDomain.Dto

type MyDto = {
    Something: string
    }

/// could be in different module if you want to hide the implementation
module MyDto = 

   // may fail if DTO has bad data
   let toDomain (dto:MyDto) :Validation<MyDomain,ValidationError) =

   // always succeeds   
   let fromDomain (domainObj:MyDomain) :MyDto =

If using validation (multiple errors), the toDomain code has the "applicative" pattern described above.

Define a ctor for the DTO
Create all the values (which return Results)
Use the applicative style to construct the DTO
Map the list of validation errors to an error case

let toDomainObj dto = 
  let ctor = ...
  let firstParam = ... construct from dto.First 
  let secondParam = ... construct from dto.Second
  let thirdParam = ...
  let domainObjR = ctor <!> firstParam <*> secondParam <*> thirdParam
  domainObjR |> Result.mapError ValidationError

Interacting with C#

Avoid exposing F# types if possible
NOTE: If Fsharp.Core.dll is missing, use nuget to add package

Exposing an API to C#

Expose API in a .NET friendly manner. Add a module called either Api or Api.Csharp
Understand tuple-style vs curried functions. Expose tuple-style only.

module Api =
  // correct
  let DoSomething(x,y,x) = 

  // incorrect
  let DoSomething x y x =

Use C#-compatible collections not F# list
- F# list -- not available in C#
- F# int seq -- same as IEnumerable<int> in C#
- F# ResizeArray<int> -- same as List<int> in C#
- F# int[] -- same as Array<int> in C#
Have a CSharpHelper module with useful functions

module List =
    // IEnumerable to F# list
    let EnumToList enum = enum |> List.ofSeq
    // IEnumerable from F# list
    let EnumFromList list = list |> List.toSeq

Use Func<> instead of F# functions
NOTE: F# Async needs to be converted to C# Task
NOTE: F# float == C# double
NOTE: C# Enums are not the same as choice/union types
Don't expose tuples in Api

Exposing choice types to C#

Create a Match function for each choice type you expose, with a Func for each case. Here's the one for Result:

module Result =
    let Match(result, onOk:Func<_,_>, onError:Func<_,_>) =
        match result with
        | Ok x -> onOk.Invoke(x)
        | Error e -> onError.Invoke(e)

Common compiler errors

Inconsistent type. Fix: add type annotations to find where the error is.
Value restriction. Fix: add a parameter.
See also https://fsharpforfunandprofit.com/troubleshooting-fsharp/

Common errors in F# Interactive

Can't use namespaces in interactive
Need to explitly refer to other files using #load or #r
Type mismatch even though you know its the same!
- caused by recompiling one bit without recompiling the other bit
- FIX: recompile the whole thing

dpraimeyuu/effective-fsharp.md