Tinkerbell Core

This is not the latest stable version of tink_core. See version 1.17.0 for the latest version.

To install, run:

haxelib install tink_core 1.7.3 

See using Haxelib in Haxelib documentation for more information.


Tinkerbell Core Library

Build Status Gitter

The tink_core lib contains a set of lightweight tools for robust programming.

All modules are situated in tink.core.*. Some contain more than a single type. Generally, it is advisable to import the modules of this package through using rather than import.

In addition, you can import all modules at once with using tink.CoreApi;.


<!-- START INDEX --> - Named - Any - Pair

	- [Nullness](#nullness)
- [MPair](#mpair)

<!-- END INDEX -->

Despite the rather long documentation here, tink_core does not exceed 1KLOC. And while was primarily drafted as the basis for the rest of tink, it can be used in isolation or for other libs to build on.


This is a very basic helper type, defined over its generalization:

typedef Named<V> = NamedWith<String, V>;

class NamedWith<N, V> {
  public var name(default, null):N;
  public var value(default, null):V;
  public function new(name, value) {
    this.name = name;
    this.value = value;

This just formalizes a notion of something being named.


The Any type (which has been included in the standard library in Haxe 3.4) is an alternative to Haxe's Dynamic defined like so:

abstract Any from Dynamic {
  @:to private inline function __promote<A>():A;

It is a type that is compatible with any other, in both ways. Yet you can do almost nothing with it, except promote it to other types. This is useful, because Dynamic itself behaves in unintuitive ways, because of the overloaded role it plays in the Haxe type system.

Behaviors of Dynamic that Any does not exhibit:

  1. Dynamic gives you the raw native runtime behavior - example with haxe/js:

`haxe var s = ''; var d:Dynamic = s;

trace(Std.string(s.charCodeAt(1)));//null trace(Std.string(d.charCodeAt(1)));//NaN - native runtime behavior is different! `

  1. Dynamic is erased during inference. Example:

`haxe var x = Reflect.field({ foo: [4] }, 'foo');//x is Unknown<?> if (x.length == 1)//x is now {+ length : Int }

trace(x[0]);//Compiler error: Array access is not allowed on {+ length : Int }


That error message is a bit confusing to say the least.

  1. Dynamic is weirdly related to Dynamic<T>. I won't go into details, because truth be told I myself am sometimes startled by certain nuances.

Compared to Dynamic, the Any type has a very specific meaning: it means the value could be of any type - that's it. The idea proposed here is to use Dynamic only to express the notion that a value is going to be accessed with native semantics (which is no doubt useful at times). If you want to access values in an untyped manner (which most of the time you should try to avoid), use the untyped keyword.

Notice how in the first example, the compiler will force you to choose a type.

var s = '';
var a:Any = s;

trace(Std.string(a.charCodeAt(1)));//does not compile because "Any has no field charCodeAt"
trace(Std.string((a:String).charCodeAt(1)));//null - of course

Also, if Reflect.field were to return Any, then you'd just have to type x to Array<Dynamic> to do anything with it.

As for an alternative to Dynamic<T>, Any does not offer one. But haxe.DynamicAccess<T> does!

So to be clear:

  • Any for values of a type not know at compile time`
  • haxe.DynamicAccess to access an object as a map
  • untyped to write untyped code
  • Dynamic to access the native runtime behavior

This should go toward clarifying what exactly is going on in a specific piece of code.


The Pair represents an ordered pair:

abstract Pair<A, B> {
  function new(a:A, b:B);
  var a(get, never):A;
  var b(get, never):B;

The representation is immutable and optimized for runtime performance. It can be used as a basic means of composition, although you should beware not to abuse it.

  • Good function getCredentials():Pair<User, Password>
  • Bad function getAddress():Pair<Pair<String, Int>, String>

In the latter example, there are two ways to actually convey meaning:

  • with pairs: function getAddress():Pair<Pair<Host, Port>, Path>
  • vanilla haxe: function getAddress():{ host:String, port:Int, path:String }

Advantages of the pair approach:

  1. Performance is good on all platforms.
  2. The returned value is immutable without having to declare all fields as readonly - this assumes you want immutability


As any complex data, pairs are nullable. For Pair we consider null an "empty pair", which is not sensible from a mathematical point of view, but Lisp managed to build a whole ecosystem on this convention, so it seems a fair bet.


The MPair is the mutable counterpart to Pair. Formerly optimized for speed, it has been demoted to a plain class, which should be fast enough 99% of the time.


The Lazy type is a primitive for lazy evaluation:

abstract Lazy<T> {  
  @:to function get():T;  
  function map<R>(f:T->R):Lazy<R>;
  function flatMap<R>(f:T->Lazy<R>):Lazy<R>;
  @:from static function ofFunc<T>(f:Void->T):Lazy<T>;
  @:from static private function ofConst<T>(c:T):Lazy<T>;

Notice it defines map and flatMap functions that allows you to transform one lazy value to another. It is important to understand that the final value is not computed until you call get as shown in this example:

using tink.CoreApi;

class Test {
  static function generate():Int {
    return Std.random(500);
  static function lazyInt():Lazy<Int>
    return generate;//Void->Int is automatically converted to Lazy<Int>
  static function lazyToString(o:Dynamic):Lazy<String> {
    trace("calling lazyToString");
    return Std.string.bind(o);//And Void->String becomes Lazy<String>
  static function main() {
    var l1 = lazyInt(),
        l2 = lazyInt();
    trace("before any access");
    trace(l1.get());//traces "generating" and the random value
    trace(l1.get());//traces the same value
    var l3 = l2.flatMap(lazyToString);
    trace("before printing l3");
     * The above traces:
     * 1. "generating" - because `l2` actually gets generated
     * 2. "calling lazyTotring" - because it was not needed before
     * 3. the resulting string representation of the second random int


The Outcome type is quite similar to haxe.ds.Option, but is different in that it is tailored specifically to represent the result of an operation that either fails or succeeds.

enum Outcome<Data, Failure> {

This is a way of doing error reporting without resorting to exceptions or sentinel values.

Because switching on Outcome all the time can become tedious, the module where Outcome resides also defines OutcomeTools, a set of helper functions best leveraged by simply using tink.core.Outcome.

This will give you a number of helper methods:

  • function sure<D,F>(outcome:Outcome<D,F>):D
    Will return the result in case of success or throw an exception otherwise. Therefore the exception is raised not when an operation fails, but when other code doesn't handle failure at all.

  • function toOption<D,F>(outcome:Outcome<D,F>):Option<D>
    Will transform an Outcome into an Option. Information on failure is thereby lost.

  • function toOutcome<D>(option:Option<D>, ?pos:haxe.PosInfos):Outcome<D, String>
    Converts an Option into an Outcome, where Some(value) is considered success and None is considered failure, with an error message.
  • function orUse<D,F>(outcome:Outcome<D,F>, fallback:D):D
    Attempts to get data from a successful Outcome or returns a fallback otherwise.
  • function orTry<D,F>(outcome:Outcome<D,F>, fallback:Outcome<D,F>):Outcome<D,F>
    If the outcome failed, uses fallback instead. Note that this can still be a failure.
  • function equals<D,F>(outcome:Outcome<D,F>, to:D):Bool
    Tells whether an Outcome is successful and the value is equal to to.
  • function map<A,B,F>(outcome:Outcome<A,F>, transform:A->B):Outcome<B, F>
    Returns a new Outcome, where the success (if any) is transformed with the given transformer.
  • function isSuccess<D,F>(outcome:Outcome<D,F>):Bool
    Tells whether an outcome was successful.

Here is some example code of what neko file access might look like:

enum FileAccessError {
  CannotOpen(path:String, reason:Dynamic);

class MyFS {
  static public function getContent(path:String) 
      if (sys.FileSystem.exists(path)) 
        try {
        catch (e:Dynamic) {
          Failure(CannotOpen(path, e));

class Main {
  static function main() {

    //First, let's process outcomes by hand it by hand:

      switch MyFS.getContent('path/to/file') {
        case Success(s): 
          trace('file content: ' + s);
        case Failure(f):
          switch f {
            case NoSuchFile(path): 
              trace('file not found: $path');
            case CannotOpen(path, reason): 
              trace('failed to open file $path because $reason');
    //Now, using OutcomeTools:

      trace('other file content: ' + MyFile.getContent('path/to/other_file').sure());
      //will throw an exception if the file content can not be read
      var equal = MyFile.getContent('path/to/file').map(haxe.Md5.encode).equals(otherMd5Sig);
      //will be true, if the file could be openend and its content's Md5 signature is equal to `otherMd5Sig`


The Error class is meant as a standard class for errors. It is defined like so:

typedef Error = TypedError<Dynamic>;

class TypedError<T> {
  var message(default, null):String;
  var code(default, null):ErrorCode;
  var data(default, null):T;
  var pos(default, null):Null<Pos>;
  var callStack(default, null):Stack;
  var exceptionStack(default, null):Stack;
  function new(?code:ErrorCode, message, ?pos:Pos):Void;
  static function withData(?code:ErrorCode, message:String, data:Dynamic, ?pos:Pos):Error;
  static function typed<A>(?code:ErrorCode, message:String, data:A, ?pos:Pos):TypedError<A>;
  static function catchExceptions<A>(f:Void->A, ?report:Dynamic->Error, ?pos:Pos):Outcome<A, Error>;

Most of the time you will just be dealing with Error, where data is simply Dynamic. In a very select cases you may wish to created typed errors where the type of the error's data is well defined.

The Pos type is just a typedef that will be haxe.macro.Expr.Position in the macro context and haxe.PosInfos otherwise.

There are a couple of interesting things to point out:

  • The throwSelf method will be called if you do sure on a Failure that is an Error. This is useful to not just have willy-nilly stack traces but instead have a chance to die gracefully.
  • In macro context, the throwSelf method will cause a compiler error at pos (defaults to haxe.macro.Context.currentPos() at the time of creation).
  • Outside macro context, Pos is PosInfos which happens to be a magical type, that when left to default, will contain the call site position. So when you pass around an Outcome and at some point call sure and it happens to be a Failure(someError), the stack trace will contain information on where the Error was actually constructed. Future versions may also capture the stack at the point of the error's creation.
  • If you compile with -D error_stack then the two stack fields are populated with contextual information (otherwise they are just empty arrays).
  • By default errors are constructed without data. Use withData or typed to construct them with Dynamic or specific data.
  • the catchExceptions function is a nice way to call functions that might throw like so: sys.io.File.getContent.bind('path/to/file').catchExceptions()


The ErrorCode type is mentioned a lot in the TypedError class above. Here is how it is actually defined:

@:enum abstract ErrorCode(Int) from Int to Int {
  var BadRequest = 400;
  var Unauthorized = 401;
  var PaymentRequired = 402;
  var Forbidden = 403;
  var NotFound = 404;
  var MethodNotAllowed = 405;
  var Gone = 410;
  var NotAcceptable = 406;
  var Timeout = 408;
  var Conflict = 409;
  var UnsupportedMediaType = 415;
  var OutOfRange = 416;
  var ExpectationFailed = 417;
  var I_am_a_Teapot = 418;
  var AuthenticationTimeout = 419;
  var UnprocessableEntity = 422;

  var InternalError = 500;
  var NotImplemented = 501;
  var ServiceUnavailable = 503;
  var InsufficientStorage = 507;
  var BandwidthLimitExceeded = 509;

You may notice that these numerical codes are a subset of HTTP status codes. There was no real point in inventing yet another set of error codes so instead I turned to one of the most ubiquitous standards out there. It actually covers quite a few cases, e.g. Unauthorized (you need to log in) vs. Forbidden (you are logged in but simply not allowed to do what you're trying). Of course I_am_a_Teapot absolutely had to be included.

If you wish to propose any more error codes, please do so.


Because in Haxe 3 Void can no longer have values, i.e. values of a type that always holds nothing, tink_core introduces Noise.

enum Noise { Noise; }

Technically, null is also a valid value for Noise. In any case, there is no good reason to inspect values of type noise, only to create them and ignore them.

An example where using Noise makes sense is when you have an operation that succeeds without any result to speak of:

function writeToFile(content:String):Outcome<Noise, IoError>;


Represents a value that can have either of two types:

enum Either<A,B> {

For example the following can represent a physical type in Haxe:

typedef PhysicalType<T> = Either<Class<T>, Enum<T>>`

function name(t:PhysicalType<Dynamic>) 
  return switch t {
    case Left(c): Type.getClassName(c);
    case Right(e): Type.getEnumName(e);

Historically it existed as such in tink_core but only remains as a typedef to the standard library's version of it.


At times you wish to share the same reference (and therefore changes to it) among different places. Since Haxe doesn't support old fashioned pointer arithmetics, we need to find other ways.

The Ref type does just that, but in an abstract:

abstract Ref<T> {
  var value(get, set):T;
  function toString():String;
  @:from static function to(value:T):Ref<T>;
  @:to function toPlain():T;

It is worth noting that Ref defines implicit conversion in both ways. The following code will thus compile:

var r:Ref<Int> = 4;//here `4` gets automatically wrapped into a reference
var i:Int = r;//here `r` gets automatically unwrapped

The current implementation is built over haxe.ds.Vector and should thus perform quite decently across most platforms.

Note that assigning a value to a reference will update the reference in place but rather wrap the value in a new reference.

var a:Ref<Int> = 4;
var b = a;
b.value = 3;
b = 2;
trace(a.value);//still 3, because b is now a different reference


To denote callbacks, tink_core introduces a special type:

abstract Callback<T> from T->Void {
  function invoke(data:T):Void;
  @:from static function fromNiladic<A>(f:Void->Void):Callback<A> 
  @:from static function fromMany<A>(callbacks:Array<Callback<A>>):Callback<A> 

The obvious question to ask here is why to complicate a simple concept as callbacks when we already have first class functions.

  • It brings more clarity to code. Function types use structural subtyping, i.e. the signature alone defines the type. Type matches can thus be unintentional. Also calling something a callback when that's what it really is, carries more meaning.
  • The use of abstracts allows for implicit conversions. If you want to subscribe to an event but don't really care for the data, you don't have to define an argument you're not using. You can simply do either of both:
  • myButton.onClick(function () trace('clicked'));
  • myButton.onClick(function (e) trace('clicked at (${e.x}, ${e.y})'));
  • Instead of specifically relying on a function type, we have a separate abstraction, which at some point can be used to leverage platform knowledge to provide for faster code that doesn't have suffer from the performance penalties anonymous function have on most platforms

Beyond that, one might ask what to do if you don't have any data to pass to the callback, or more than a single value. In that case, you could use these respectively:

Callback<Pair<A, B>>
Callback<{ a: A, b: B, c: C }>

This approach has two advantages:

  • For one, it greatly simplifies things. Implementations of signals only ever consume one type of callbacks, so you don't need signals for 0, 1, 2 and possibly 3 arguments.
  • Types written against this single callback type are easier to work with in a consistent matter.
  • In the last case you can add additional information in a new field without breaking code


When you register a callback to a caller, you often want to be able to undo this registration. Classically, the caller provides for this functionality by exposing two methods, one for registering and one for unregistering.

As opposed to that, tink adheres to a different approach, where the registration returns a "callback link", i.e. a value representing the link between the caller and the callback. It looks as follows:

abstract CallbackLink {
  function dissolve():Void;
  @:to function toCallback<A>():Callback<A>;
  @:from static function fromFunction(f:Void->Void):CallbackLink;

Calling dissolve will dissolve the link, as suggested by the name. Ain't no rocket science ;)

The link itself can be promoted to become a callback, so that you can in fact register it as a handler elsewhere:

button.onPress(function () {
  var stop = button.onMouseMove(function () trace('move!'));


While the Callback and CallbackLink are pretty nice in theory, on their own, they have no application. For that reasons tink_core defines a basic infrastructure to provide callback registration and link dissolving:

abstract CallbackList<T> {
  var length(get, never):Int;
  function new():Void;
  function add(cb:Callback<T>):CallbackLink;
  function invoke(data:T):Void; 
  function clear():Void;

By calling add you can thus register a callback and will obtain a link that allows undoig the registration. You can invoke all callbacks in the list with some data, or clear the list if you wish to.

Registering callbacks

Unlike with similar mechanisms, you can add the same callback multiple times and one invoke will then cause the callback to be called multiple times. You will however get distinct callback links that allow you to separately undo the registrations. While this behavior might strike you as unfamiliar, it does have advantages:

  • Adding callbacks becomes very cheap (since you don't have to check whether they are already existent)
  • Avoid trouble with all sorts of inconsistencies regarding function equality on different Haxe targets
  • Have a clear and simple behavior, that is thus highly predictable - i.e. callbacks are simply executed in the order they are registered in. If you register a new callback, you can expect all previously registered callbacks to execute before it. The same cannot be said in case of the more common approach, if the callback was registered already once, so execution order tends to be undefined.

In essence the CallbackList can be seen as a basic building block for notification mechanisms.


As the name would suggest, futures express the idea that something is going to happen in the future. Or much rather: a future represents the result of a potentially asynchronous operation, that will become available at some point in time. It allows you to register a Callback to handle the operation's result once it is available.

abstract Future<T> {
  function handle(callback:Callback<T>):CallbackLink;
  function map<A>(f:T->A, ?gather = true):Future<A>;
  function flatMap<A>(f:T->Future<A>, ?gather = true):Future<A>; 
  static function flatten<A>(f:Future<Future<A>>):Future<A>;
  function first(other:Future<T>):Future<T>;
  function merge<A, R>(other:Future<A>, how:T->A->R):Future<R>;
  @:from static function fromMany<A>(a:Array<Future<A>>):Future<Array<A>>;
  static function sync<A>(v:A):Future<A>;
  static function lazy<A>(f:Void->A):Future<A>;
  static function async<A>(f:(A->Void)->Void, ?lazy = false):Future<A>;  
  static function trigger<A>():FutureTrigger<A>;//FutureTrigger is documented below
  static function ofMany<A>(futures:Array<Future<A>>, ?gather = true):Future<Array<A>>;
  function new(f:Callback<T>->CallbackLink):Void;  

It is important to note that a future - despite its name - need not necessarily represent an operation whose result still lies in the future. Once the underlying operation has completed, the future will retain the result and if you register a callback, it will be called back immediately. There are claims that such behavior can cause problems. The problem is however caused by making assumptions about when the callback will be invoked. Do not make any such assumptions and you will be on the safe side. You can see the Future as a micro-framework, that inverts control.

Why use futures?

We can already deal with asynchrony by means of plain old callbacks. Introducing futures has two advantages:

  • Futures are values and that allows for composition
  • Futures are very generic. They need not represent an asynchronous operation, they might just as well represent a lazy one or they may even hold a value that has been available from the very start. Writing a piece of code against futures allows you to work with and even intermix these three types of evaluation strategies.

Say you have these functions, that are built on one another:

function loadFromURL(url:String, callback:String->Void):Void { 
  /* load data somehow */ 

function loadFromParameters(params: { host: String, port: Int, url:String, params:Map<String> }, callback:String->Void):Void {
  var url = buildURL(params);//wherever this comes from
  loadFromURL(url, callback);

function loadAll(params:Array<{ host:String, port:Int, url:String, params:Map<String> }, callback:Array<String>->Void):Void {
  var results = [],
    count = 0;
  for (i in 0...params.length) {
    loadFromParameters(params[i], function (data) {
      results[i] = data;
      if (--count == 0) callback(results);

Now let's see that code with futures:

function loadFromURL(url:String):Future<String> { 
  /* load the data somehow */ 

function loadFromParameters(params: { host: String, port: Int, url:String, params:Map<String> }):Future<String> {
  var url = buildURL(params);//wherever this comes from
  return loadFromURL(url);

function loadAll(params:Array<{ host:String, port:Int, url:String, params:Map<String> }):Future<Array<String>> {
  return [for (p in params)

A couple of observations:

  • Rather than passing callbacks from function to function, we return futures. We do not know, that somebody is going to want to be called back and we don't really care. We simply return a future and client code can decide whether or not it wants to handle the result of an operation.
  • Futures can be composed, because they are values. If you compare the implementations of loadAll the advantage of that should become evident. In the future based implementation, we use an array comprehension to simply comprise the individual futures resulting from loadFromParameters to an array.
    Note that the value that we constructed is Array<Future<String>> whereas what we are returning is a single Future<Array<String>>. They are not at all the same but converting the former to the latter is indeed possible and happens automagically here, because Future defines an implicit conversion rule (see fromMany) to do just that.

So rather than passing callbacks along with all calls, we simply return futures and the resulting code is much closer to how we would do this with synchronous APIs.


In the example above we can see some composition of futures at work. What's equally interesting is transformation. The basic tool is map which works very much like it does for Array. For a Array, it constructs a new Array by applying a function to all elements. For a Future it does pretty much the same thing: it constructs a new Future by applying a function to the result.

Say we want to load JSON instead of raw data in the example above. To expand on the example above, this is how we would accomplish it:

function loadJson(url:String):Future<Dynamic> 
  return loadFromUrl(url).map(haxe.Json.parse);

Now let's say that we know for a fact, that all these JSONs contain arrays of strings and we actually just want the first entry, then we would do this:

function loadJson(url:String):Future<String>
  return loadFromUrl(url).map(haxe.Json.parse).map(function (a) return a[0]);

Or let's try something else. Loading information from wikipedia.

function loadWikiDescription(article:String):Future<Null<String>> 
    loadFromUrl('http://en.wikipedia.org/wiki/$article').map(function (html:String) 
        if (html.indexOf('Wikipedia does not have an article with this exact name') != -1) null;
        else html.split('<p>').pop().split('</p>').shift();

So the above will create a future that returns null if there's no article, or the contents of the first paragraph (please do not parse HTML like that in production code).

Now let's assume that we want to load an article that is specified in a config, then we would try this:

  function (config: { article:String }):Future<Null<String>>
    return loadWikiDescription(config.article)

There is only one problem. The transformation function does not return a plain value, but rather a future. Since map with a function of type T->A transforms a Future<T> to a Future<A>, map with a function of type String->Future<Null<String>>. so the resulting future will in fact be of type Future<Future<Null<String>>> whereas we would really rather want Future<Null<String>>. One could call this a "nested" future and we want to flatten it:

  function (config: { article:String }) 
    return loadWikiDescription(config.article)

Now because this is a lot to write for a rather common situation, we have flatMap which basically just does a map and flatten:

  function (config: { article:String }) 
    return loadWikiDescription(config.article)

And the result is a plain future.


To compose futures, you have three basic options:

  1. first - take two futures of the same type and construct one that yields the result of whichever future finishes first. Example: loadFrom(source1).first(loadFrom(source2))
  2. fromMany - take an array of futures and transform it to a single future of an array of the results. In fact you've seen this in action in "Why use futures?"
  3. merge - take two futures and merge them together by means of a function. Example: loadFrom(source1).merge(loadFrom(source2), function (r1, r2) return r1 + r2)

Now you may want to use first on futures of different types. Here's how that would work:

var x:Future<X> = ...; 
var y:Future<Y> = ...; 

$type(x.map(Either.Left).first(y.map(Either.Right)));//Future<Either<X, Y>>


The keen observer may have noticed the optional gather argument for map and flatMap. This a compromise that leaks an implementation detail to give you the possibility to reduce overhead. When in doubt, leave it untouched.

Let's examine a naive implementation of map:

function map<In, Out>(future:Future<In>, transform:In->Out):Future<Out>
  return new Future(
    function (callback:Callback<Out>):CallbackLink
      return f.handle(
        function (data:In) callback.invoke(transform(data))

What this does is to create a future that deals with a callback by registering a new handle to the original future that will first transform the data and then invoke the original callback. This pretty much does what we want, with one problem: The transformation is executed for every callback.


var f = Future.sync('foo'),
  array = [];

var mapped = map(f, array.push);
mapped.handle(function (x) trace(x));//1
mapped.handle(function (x) trace(x));//2
trace(array);//[foo, foo]

Ideally the transformation should be a pure function, so you will not actually have any direct side effects. But even then it would be executed multiple times which may cause performance issues if the transformation is costly.

To avoid this, we would need a mechanism that internally stores the transformed result once it becomes available and dispatches that onto all callbacks. That's what "gathering" does.

Gathering is used by default, because by default it's the sensible thing to do. However, you may build chains such as f.map(t1).map(t2).flatMap(t3).map(t4). The intermediary futures will only ever have one handler. In that case the overhead introduced by gathering is not needed.

If the transformations are pretty straight forward but frequent and you're dealing with synchronous code, then turning off gathering can lead to noticable performance differences.

Rolling your own

While you can create Futures with the constructor, the suggested method is to use one of the static constructors.

In the example above, we've seen a loadFromUrl function. Here's a way to implement it on top of haxe.Http on async platforms with Future.async:

//First, let's have a plain old callback based function
function loadAsync(url:String, callback:String->Void) {
  var h = new haxe.Http(url);
  h.onDone = callback;
//and now, pixie dust!!!!
function loadFromUrl(url:String) 
  return Future.async(loadAsync.bind(url));

Or if we wanted to achieve the same in one step:

function loadFromUrl(url:String)
  return Future.async(function (handler:String->Void) {
    var h = new haxe.Http(url);
    h.onDone = handler;

A fair question to ask would be, how to deal with errors. Quite simply, we will use our old friend the Outcome:

function loadFromUrl(url:String):Future<Outcome<String, String>>
  return Future.async(function (handler:Outcome<String, String>->Void) {
    var h = new haxe.Http(url);
    h.onDone = function (data:String) handler(Success(data));
    h.onError = function (error:String) handler(Failure(error));

Now suppose we wanted the code to run on PHP that's synchronous. For one, the implementation above would already happen to work. But typically most APIs that you would want to deal with are purely synchronous and you'd have to deal with that. This is how:

function loadFromUrl(url:String)
  return Future.sync(haxe.Http.requestUrl(url));
function loadFromUrl(url:String)
  return Future.lazy(haxe.Http.requestUrl.bind(url));

The first version just gets the data synchronously and then "lifts" it to become a Future. The second version constructs a lazy future, i.e. the operation is executed when you register the first callback. Example:

var load = loadFromUrl('http://example.com/');//no requests have been made yet
load.handle(function (data) {});//now the request is made
load.handle(function (data) {});//the data is already available, so no request is made

Lazyness of course is something that you may want in async scenarios. For that reason Future.async has a lazy parameter that you can set to true.

Please note that lazyness is not always preferable. In the context of HTTP for example, it might make sense to have GET requests be lazy, because there's no point in loading the data if you're not going to handle it. As opposed to that, POST requests should not be lazy. You want the request to be sent of to the server, whether or not you're going to handle the response.

For any more complex scenarios, you can use FutureTrigger to complete the future by hand.


Because Future is an abstract, we can do some neat tricks with operator overloading. Before looking at those, you might want to peek at the next section to see what a Surprise is.

  1. ||
  2. If futures are of the same type, will use first
  3. If futures are of different type, will collapse the type by means of Either and then use first (as shown at the end of the "Composition" section)
  4. && - will combine a Future<A> and Future<B> to a Future<Pair<A, B>>.
  5. >>
  6. Will flatMap a Surprise<A, F> to a Surprise<B, F> with a A->Surprise<B, F>
  7. Will flatMap a Surprise<A, F> to a Surprise<B, F> with a A->Future<B>
  8. Will map a Surprise<A, F> to a Surprise<B, F> with a A->Outcome<B, F>
  9. Will map a Surprise<A, F> to a Surprise<B, F> with a A->B
  10. Will flatMap a Future<A> to a Future<B> with a A->Future<B>
  11. Will map a Future<A> to a Future<B> with a A->B

Evidently, >> is quite supercharged. Let's examine an example from above once more to see why:

  function (config: { article:String }) 
    return loadWikiDescription(config.article)

We can now write it as this:

loadJson('config.json') >> 
  function (config: { article:String }) 
    return loadWikiDescription(config.article);

Apart from shaving off a few characters, we achieved something else entirely. This piece of code is significantly more flexible. If loadJson starts returning a Surprise because the maintainer added error handling, our code remains unaffected. The overloaded >> operator will lift the transformation to the right context. The same applies for loadWikiDescription. With this syntax it no longer matters whether it returns a Surprise or just a Future or even just a plain value.

Handle the future like a boss ;)


For all those who love surprises and for all those who hate them, tink_core provides a neat way of expressing them. Simply put, a surprise is nothing but a future outcome. Literally:

typedef Surprise<D, F> = Future<Outcome<D, F>>;

This type thus represents an operation that will finish at some point in time and can end in failure. Perfect for representing asynchronous I/O and such. We've seen it in the examples above, we just didn't call it that way.


In an above section we've seen ways to construct futures on top of other APIs. However, you may need to build your own future yourself from scratch. The simplest way to do this is by using a helper class:

class FutureTrigger<T> {
  function new();
  function asFuture():Future<T>;
  function trigger(result:T):Bool;

Typically you would construct a trigger with Future.trigger() instead of new FutureTrigger(). One advantage is that the former works with import tink.core.*; and the other one requires you to import tink.core.Future; explicitly.

Here is how you would use such a trigger (as an alternative to the example above):

class Http {
  static public function requestURL(url:String):Surprise<String, String> {
    var req = new haxe.Http(url),
      f = Future.trigger();
    req.onData = function (data) f.trigger(Success(data));
    req.onError = function (error) f.trigger(Failure(error));
    return f.asFuture();

Looks pretty neat already. And it forces client code to consider failure.


Promises are a relatively late addition to tink_core and are the result of the realization that most of the time one will just use Error as the error type of a Surprise. Indeed typedef Promise<T> = Surprise<T, Error> is relatively close to what promises are, but there's a few more comfort features added to the mix:

abstract Promise<T> from Surprise<T, Error> to Surprise<T, Error> {
  function map<R>(f:Outcome<T, Error>->R):Future<R>;
  function flatMap<R>(f:Outcome<T, Error>->Future<R>):Future<R>;
  function handle(cb:Callback<Outcome<T, Error>>):CallbackLink;

  function next<R>(f:Next<T, R>):Promise<R>; 
  function recover(f:Recover<T>):Future<T>;
  @:to function noise():Promise<Noise>;

  @:from static private function ofSpecific<T, E>(s:Surprise<T, TypedError<E>>):Promise<T>;    
  @:from static private function ofFuture<T>(f:Future<T>):Promise<T>;    
  @:from static private function ofOutcome<T>(o:Outcome<T, Error>):Promise<T>;    
  @:from static private function ofError<T>(e:Error):Promise<T>;
  @:from static private function ofData<T>(d:T):Promise<T>;
  @:noUsing static function lift<T>(p:Promise<T>)

So you see that map, flatMap and handle allow you to deal with the promise as though it were an ordinary Surprise.

What's added is for one the ability to promote any Future, Outcome, Surprise, Error or plain value to a Promise. On top, we have a way to transform Promise<T> to Promise<R> with a Next<T, R> and to recover a Promise<T> to a Future<T> with a Recover<T> - let's have a look at both types:


As some point you may wish to recover from the error case of a promise and continue with a plain future. That is what Recover<T> is for:

abstract Recover<T>(Error->Future<T>) from Error->Future<T> {
  @:from static function ofSync<T>(f:Error->T):Recover<T>
    return function (e) return Future.sync(f(e));

You must provide a function that either synchronously or asynchronously turns an error into data of the expected type.


The Next<In, Out> type describes an asynchronous transformation from In to Out. In essence it is just In->Promise<Out> with a bit of abstract magic sprinkled on top.

abstract Next<In, Out>(In->Promise<Out>) from In->Promise<Out> {
  @:from static private function ofSafe<In, Out>(f:In->Outcome<Out, Error>):Next<In, Out>;
  @:from static private function ofSync<In, Out>(f:In->Future<Out>):Next<In, Out>;
  @:from static private function ofSafeSync<In, Out>(f:In->Out):Next<In, Out>;
  @:op(a * b) static private function _chain<A, B, C>(a:Next<A, B>, b:Next<B, C>):Next<A, C>;

What this achieves is that we may chain operations on promises in a similarly flexible way as >>.

function requestUrl(url:String):Promise<String> { ... };

  .next(function (s:String) 
    try return Xml.parse(s)
    catch (e:Dynamic) return new Error('Invalid XML: $s')
  .next(function (x:Xml) return
  .next(function (data:String) return Future.async(function (cb) {
    var div = document.createDivElement();
    var span = document.createSpanElement();
    span.innerHTML = 'Proceed?';
    var accept = document.createButtonElement();
    accept.innerHTML = 'Yes';
    accept.onclick = cb.bind(data);

As we see, some transformations return synchronously, others return promises, others return plain futures. The first one is particularly interesting in that in one case it returns an Xml and in the other case it returns an Error. That is because the return type of the function is actually known to be Promise<T> and both errors and data can be promoted to that.

Promise vs Surprise

Promises and futures neatly complement each other in that one means an asynchronous operation that can fail, and the other an operation that can't fail. However promises and surprises compete over the same semantics. Obviously you will want to use surprises if the error type is anything other than Error. For everything else, you will probably find Promise to be easier to deal with. It works better with type inference that >> and also leads to saner error reporting. If you get type errors with >> then it tends to results in cascades of Future<SomeInsanelyComplexTypeParameterHereThatSpansMultipleLines> should be Int.


Despite an overabundance of signal implementations, tink_core does provide its own flavour of signals. One that aims at utmost simplicity and full integration with the rest of tink. Here it is:

abstract Signal<T> {
  function new(f:Callback<T>->CallbackLink):Void;

  function handle(calback:Callback<T>):CallbackLink;
  function map<A>(f:T->A, ?gather:Bool = true):Signal<A>;
  function join(other:Signal<T>, ?gather:Bool = true):Signal<T>;
  function noise():Signal<Noise>;
  function gather():Signal<T>;
  function next():Future<T>;  
  static function trigger<A>():SignalTrigger<A>;
  static function ofClassical<A>(add:(A->Void)->Void, remove:(A->Void)->Void, ?gather = true):Signal<A>;

A Signal quite simply invokes Callbacks that are registered using handle whenever an event occurs.

There is significant similarity between Signal and Future. It's best to read up on futures if you haven't done so yet. You may also notice next, which at any time will create a future corresponding to the next occurence of the signal. So if you only want to do something on the next occurence, you would do someSignal.next().handle(function (data) {}).

Why use signals?

When compared to mechanisms like flash's EventDispatcher or the DOM's EventTarget or nodejs's EventEmitter, the critical advantage is type safety.

Say you have the following class:

class Button {
  public var pressed(default, null):Signal<MouseEvent>;
  public var clicked(default, null):Signal<MouseEvent>;
  public var released(default, null):Signal<MouseEvent>;

You know exactly which events to expect and what type they will have. Also, an interface can define signals that an implementor must thus provide. And lastly, the fact that a signal itself is a value, you can pass it around rather then the whole object owning it. Similarly to futures, this allows for composition.

Wrapping 3rd party APIs

It is quite easy to take an arbitrary API and wrap it in signals. Let's take the beloved IEventDispatcher.

function makeSignal<A:Event>(dispatcher:IEventDispatcher, type:String):Signal<A> 
  return Signal.ofClassical(

var keydown:Signal<KeyboardEvent> = makeSignal(stage, 'keydown');//yay!!!

As far as just tink.core.Signal and interoperability with other APIs is concerned, the primary goal is to make it extremely simple to represent any kind of API with signals. And as shown above, there's really not much to it.

Rolling your own

As the constructor indicates, a signal can be constructed from any function that consumes a callback and returns a link. Therefore CallbackList is pretty much a perfect fit. However, for the sake of consistency with Future, the intended usage is to create a SignalTrigger that you use internally to - well - trigger the Signal.

No way to dispatch from outside

Unlike many other signal flavors, tink's signals do not allow client code to invoke the signal. When exposing a signal that you trigger yourself, typically you will expose only a Signal while internally knowing the SignalTrigger. Here's an example of just that:

class Clock {
  public var tick(default, null):Signal<Noise>;
  public function new() {
    var s = Signal.trigger();//<-- this trigger is never passed outside the constructor
    var t = new Timer(1000);
    t.run = function () s.trigger(Noise);
    this.tick = s;

That being said, if you wish to provide means to dispatch a signal from outside, you can of course do so by exposing this functionality in whatever way you see fit.


Now let's have a look at the map and join and noise methods. The gather parameter has a very similar role to the counterpart for Future. We'll examine that later.

First of all, map comes from the functional term of mapping. The idea is to use a function that maps values of one type onto other values (possibly of another type) and give that to a more complex data structure that also deals with values of that type, creating a similar data structure, where the function has been applied to every value. In Haxe Lambda.map does this for all iterable data structures. Also Array and List have built in support for this. So if you want to understand this concept, that's probably the best place to look for an easy example.

Secondly, we have join that allows us to join two signals of the same type into one. Here's an example of what that might look like, where we assume that we have a plusButton and a minusButton on our GUI and they each have a signal called clicked:

var delta = 
    .map(function (_) return 1)
    .join(minusButton.clicked.map(function (_) return -1));


From that we have constructed a new Signal that fires 1 when the plusButton is clicked and -1, when the minusButton is clicked.

This way we can map many input events into a single application event without extraneous noise.

The noise method mentioned earlier is merely a shortcut to map any Signal to a Signal<Noise>, thus discarding any information it carries. This is useful when you want to propagate an event but not expose the original data and also have no meaningful substitute.


Similar to Future, we have "gathering" for Signal as well - for pretty much the same reasons. But it also has another role - normalizing behavior:

As we've seen any function that accepts a Callback and returns a CallbackLink is suitable to act as a Signal. But such a Signal needn't behave consistently with those built on CallbackList, i.e. invokation order might be different or duplicate registration might not be allowed or whatever. Or in some instances, the implementation might be slower or weird in some other way (e.g. ACE's EventEmitter implementation that has all sorts of unexpected behavior if you add/remove handlers for an event type, while an event of the type is dispatched).

What gathering does for signals is quite easily explained. It creates a new SignalTrigger and registers that with the original signal. Then the signal derived from the trigger is returned. Therefore behavioral oddities of the original implementation are hidden.

If we look at Signal.ofClassical as used in the example with IEventDispatcher, we've used gathering (by default). We could choose not to use it. In that case, the callbacks registration would be delegated to the dispatcher in a relatively direct fashion - with all the advantages and problems this may lead to (usually the problems outweigh any advantages).


A SignalTrigger is what permits you to build a signal that you can trigger yourself:

abstract SignalTrigger<T> {
  function new();
  function trigger(result:T):Void;
  function clear():Void
  @:to function asSignal():Signal<T>;

The "clock" example above demonstrates how to do that.

1 year ago

All libraries are free

Every month, more than thousand developers use haxelib to find, share, and reuse code — and assemble it in powerful new ways. Enjoy Haxe; It is great!

Explore Haxe

Haxe Manual

Haxe Code Cookbook

Haxe API documentation

You can try Haxe in the browser! try.haxe.org

Join us on Github!

Haxe is being developed on GitHub. Feel free to contribute or report issues to our projects.

Haxe on Github