Free email newsletter: “ES.next News

2014-12-09

Meta programming with ECMAScript 6 proxies

This blog post is outdated. Please read chapter “Metaprogramming with proxies” in “Exploring ES6”.


This blog post explains the ECMAScript 6 (ES6) feature proxies. Proxies enable you to intercept and customize operations performed on objects (such as getting properties). They are a meta programming feature.

The code in this post occasionally uses other ES6 features. Consult “Using ECMAScript 6 today” for an overview of all of ES6.

Before we can get into what proxies are and why they are useful, we first need to understand what meta programming is.

Programming versus meta programming

In programming, there are levels:

  • At the base level (also called: application level), code processes user input.
  • At the meta level, code processes base level code.

Base and meta level can be diffent languages. In the following meta program, the meta programming language is JavaScript and the base programming language is Java.

    let str = 'Hello' + '!'.repeat(3);
    console.log('System.out.println("'+str+'")');

Meta programming can take different forms. In the previous example, we have printed Java code to the console. Let’s use JavaScript as both meta programming language and base programming language. The classic example for this is the eval() function, which lets you evaluate/compile JavaScript code on the fly. There are very few actual use cases for eval(). In the interaction below, we use it to evaluate the expression 5 + 2.

    > eval('5 + 2')
    7

Other JavaScript operations may not look like meta programming, but actually are, if you look closer:

    // Base level
    let obj = {
        hello() {
            console.log('Hello!');
        }
    };
    
    // Meta level
    for (let key of Object.keys(obj)) {
        console.log(key);
    }

The program is examining its own structure while running. This doesn’t look like meta programming, because the separation between programming constructs and data structures is fuzzy in JavaScript. All of the Object.* methods can be considered meta programming functionality.

Kinds of meta programming

Reflective meta programming means that a program processes itself. Kiczales et al. [2] distinguish three kinds of reflective meta programming:

  • Introspection: you have read-only access to the structure of a program.
  • Self-modification: you can change that structure.
  • Intercession: you can redefine the semantics of some language operations.

Let’s look at examples.

Example: introspection. Object.keys() performs introspection (see previous example).

Example: self-modification. The following function moveProperty moves a property from a source to a target. It performs self-modification via the bracket operator for property access, the assignment operator and the delete operator. (In production code, you’d probably use property descriptors for this task.)

    function moveProperty(source, propertyName, target) {
        target[propertyName] = source[propertyName];
        delete source[propertyName];
    }

Using moveProperty():

    > let obj1 = { prop: 'abc' };
    > let obj2 = {};
    > moveProperty(obj1, 'prop', obj2);
    
    > obj1
    {}
    > obj2
    { prop: 'abc' }

JavaScript doesn’t currently support intercession, proxies were created to fill that gap.

An overview of proxies

ECMAScript 6 proxies bring intercession to JavaScript. They work as follows. There are many operations that you can perform on an object obj. For example:

  • Getting a property prop (via obj.prop)
  • Listing enumerable own properties (via Object.keys(obj))

Proxies are special objects that allow you to provide custom implementations for some of these operations. A proxy is created with two parameters:

  • handler: For each operation, there is a corresponding handler method that – if present – performs that operation. Such a method intercepts the operation (on its way to the target) and is called a trap (a term borrowed from the domain of operating systems).
  • target: If the handler doesn’t intercept an operation then it is performed on the target. That is, it acts as a fallback for the handler. In a way, the proxy wraps the target.

In the following example, the handler intercepts the operation get (getting properties).

    let target = {};
    let handler = {
        get(target, propKey, receiver) {
            console.log('get ' + propKey);
            return 123;
        }
    };
    let proxy = new Proxy(target, handler);

When we get property foo, the handler intercepts that operation:

    > proxy.foo
    get foo
    123

The handler doesn’t implement the trap set (setting properties). Therefore, setting proxy.bar is forwarded to target and leads to target.bar being set.

    > proxy.bar = 'abc';
    > target.bar
    'abc'

Function-specific traps

If the target is a function, two additional operations can be intercepted:

  • apply: Making a function call, triggered via proxy(···), proxy.call(···), proxy.apply(···).
  • construct: Making a constructor call, triggered via new proxy(···).

The reason for only enabling these traps for function targets is simple: You wouldn’t be able to forward the operations apply and construct, otherwise.

Revocable proxies

ECMAScript 6 lets you create proxies that can be revoked (switched off):

    let {proxy, revoke} = Proxy.revocable(target, handler);

On the left hand side of the assignment operator (=), we are using destructuring to access the properties proxy and revoke of the object returned by Proxy.revocable().

After you call the function revoke for the first time, any operation you apply to proxy causes a TypeError. Subsequent calls of revoke have no further effect.

    let target = {}; // Start with an empty object
    let handler = {}; // Don’t intercept anything
    let {proxy, revoke} = Proxy.revocable(target, handler);
    
    proxy.foo = 123;
    console.log(proxy.foo); // 123
    
    revoke();
    
    console.log(proxy.foo); // TypeError: Revoked

Proxies as prototypes

A proxy proto can become the prototype of an object obj. Some operations that begin in obj may continue in proto. One such operation is get.

    let proto = new Proxy({}, {
        get(target, propertyKey, receiver) {
            console.log('GET '+propertyKey);
            return target[propertyKey];
        }
    });
    
    let obj = Object.create(proto);
    obj.bla; // Output: GET bla

The property bla can’t be found in obj, which is why the search continues in proto and the trap get is triggered there. There are more operations that affect prototypes, they are listed at the end of this post.

Forwarding operations

Operations whose traps the handler doesn’t implement are automatically forwarded to the target. Sometimes there is some task you want to perform in addition to forwarding the operation. For example, a handler that intercepts all operations and logs them, but doesn’t prevent them from reaching the target:

    let handler = {
        deleteProperty(target, propKey) {
            console.log('DELETE ' + propKey);
            return delete target[propKey];
        },
        has(target, propKey) {
            console.log('HAS ' + propKey);
            return propKey in target;
        },
        // Other traps: similar
    }

For each trap, we first log the name of the operation and then forward it by performing it manually. ECMAScript 6 has the module-like object Reflect that helps with forwarding: for each trap

    handler.trap(target, arg_1, ···, arg_n)

Reflect has a method

    Reflect.trap(target, arg_1, ···, arg_n)

If we use Reflect, the previous example looks as follows.

    let handler = {
        deleteProperty(target, propKey) {
            console.log('DELETE ' + propKey);
            return Reflect.deleteProperty(target, propKey);
        },
        has(target, propKey) {
            console.log('HAS ' + propKey);
            return Reflect.has(target, propKey);
        },
        // Other traps: similar
    }

Now what each of the traps does is so similar that we can implement the handler via a proxy:

    let handler = new Proxy({}, {
        get(target, trapName, receiver) {
            // Return the handler method named trapName
            return function (...args) {
                // Slice away target object in args[0]
                console.log(trapName.toUpperCase()+' '+args.slice(1));
                // Forward the operation
                return Reflect[trapName](...args);
            }
        }
    });

For each trap, the proxy asks for a handler method via the get operation and we give it one. That is, all of the handler methods can be implemented via the single meta method get. It was one of the goals for the proxy API to make this kind of virtualization simple.

Let’s use this proxy-based handler:

    > let target = {};
    > let proxy = new Proxy(target, handler);
    > proxy.foo = 123;
    SET foo,123,[object Object]
    > proxy.foo
    GET foo,[object Object]
    123

The following interaction confirms that the set operation was correctly forwarded to the target:

    > target.foo
    123

Use cases for proxies

This section demonstrates what proxies can be used for. That will also give you the opportunity to see the API in action.

Implementing the DOM in JavaScript

The browser Document Object Model (DOM) is usually implemented as a mix of JavaScript and C++. Implementing it in pure JavaScript is useful for:

  • Emulating a browser environment, e.g. to manipulate HTML in Node.js. jsdom is one library that does that.
  • Speeding the DOM up (switching between JavaScript and C++ costs time).

Alas, the standard DOM can do things that are not easy to replicate in JavaScript. For example, most DOM collections are live views on the current state of the DOM that change dynamically whenever the DOM changes. As a result, pure JavaScript implementations of the DOM are not very efficient. One of the reasons for adding proxies to JavaScript was to help write more efficient DOM implementations.

Accessing a restful web service

A proxy can be used to create an object on which arbitrary methods can be invoked. In the following example, the function createWebService creates one such object, service. Invoking a method on service retrieves the contents of the web service resource with the same name. Retrieval is handled via an ECMAScript 6 promise.

    let service = createWebService('http://example.com/data');
    // Read JSON data in http://example.com/data/employees
    service.employees().then(json => {
        let employees = JSON.parse(json);
        ···
    });

The following code is a quick and dirty implementation of createWebService in ECMAScript 5. Because we don’t have proxies, we need to know beforehand what methods will be invoked on service. The parameter propKeys provides us with that information, it holds an array with method names.

    function createWebService(baseUrl, propKeys) {
        let service = {};
        propKeys.forEach(function (propKey) {
            Object.defineProperty(service, propKey, {
                get: function () {
                    return httpGet(baseUrl+'/'+propKey);
                }
            });
        });
        return service;
    }

The ECMAScript 6 implementation of createWebService can use proxies and is simpler:

    function createWebService(baseUrl) {
        return new Proxy({}, {
            get(target, propKey, receiver) {
                return httpGet(baseUrl+'/'+propKey);
            }
        });
    }

Both implementations use the following function to make HTTP GET requests (how it works is explained in the 2ality blog post on promises).

    function httpGet(url) {
        return new Promise(
            (resolve, reject) => {
                let request = new XMLHttpRequest();
                Object.assign(request, {
                    onreadystatechange() {
                        if (this.status === 200) {
                            // Success
                            resolve(this.response);
                        } else {
                            // Something went wrong (404 etc.)
                            reject(new Error(this.statusText));
                        }
                    },
                    onerror() {
                        reject(new Error(
                            'XMLHttpRequest Error: '+this.statusText));
                    }
                });
                request.open('GET', url);
                request.send();    
            });
    }

Tracing property accesses

The example in this section is inspired by Brendan Eich’s talk “Proxies are Awesome”: We want to trace when a given set of properties is read or changed. To demonstrate how that works, let’s create a class for points and trace accesses to the properties of an instance.

    class Point {
        constructor(x, y) {
            this.x = x;
            this.y = y;
        }
        toString() {
            return 'Point('+this.x+','+this.y+')';
        }
    }
    // Trace accesses to properties `x` and `y`
    let p = new Point(5, 7);
    p = tracePropAccess(p, ['x', 'y']);

Getting and setting properties of p now has the following effects:

    > p.x
    GET x
    5
    > p.x = 21
    SET x=21
    21

Intriguingly, tracing also works whenever Point accesses the properties, because this now refers to the proxy, not to an instance of Point.

    > p.toString()
    GET x
    GET y
    'Point(21,7)'

In ECMAScript 5, you’d implement tracePropAccess() as follows. We replace each property with a getter and a setter that traces accesses. The setters and getters use an extra object, propData, to store the data of the properties. Note that we are destructively changing the original implementation, which means that we are meta programming.

    function tracePropAccess(obj, propKeys) {
        // Store the property data here
        let propData = Object.create(null);
        // Replace each property with a getter and a setter
        propKeys.forEach(function (propKey) {
            propData[propKey] = obj[propKey];
            Object.defineProperty(obj, propKey, {
                get: function () {
                    console.log('GET '+propKey);
                    return propData[propKey];
                },
                set: function (value) {
                    console.log('SET '+propKey+'='+value);
                    propData[propKey] = value;
                },
            });
        });
        return obj;
    }

In ECMAScript 6, we can use a simpler, proxy-based solution. We intercept property getting and setting and don’t have to change the implementation.

    function tracePropAccess(obj, propKeys) {
        let propKeySet = new Set(...propKeys);
        return new Proxy(obj, {
            get(target, propKey, receiver) {
                if (propKeySet.has(propKey)) {
                    console.log('GET '+propKey);
                }
                return Reflect.get(target, propKey, receiver);
            },
            set(target, propKey, value, receiver) {
                if (propKeySet.has(propKey)) {
                    console.log('SET '+propKey+'='+value);
                }
                return Reflect.set(target, propKey, value, receiver);
            },
        });
    }

Warning about unknown properties

When it comes to accessing properties, JavaScript is very forgiving. For example, if you try to read a property and misspell its name, you don’t get an exception, you get the result undefined. You can use proxies to get an exception in such a case. This works as follows. We make the proxy a prototype of an object.

If a property isn’t found in the object, the get trap of the proxy is triggered. If the property doesn’t even exist in the prototype chain after the proxy, it really is missing and we throw an exception. Otherwise, we return the value of the inherited property. We do so by forwarding the get operation to the target, whose prototype is the prototype of the proxy.

    let PropertyChecker = new Proxy({}, {
        get(target, propKey, receiver) {
            if (!(propKey in target)) {
                throw new ReferenceError('Unknown property: '+propKey);
            }
            return Reflect.get(target, propKey, receiver);
        }
    });

Let’s use PropertyChecker for an object that we create:

    > let obj = { __proto__: PropertyChecker, foo: 123 };
    > obj.foo  // own
    123
    > obj.fo
    ReferenceError: Unknown property: fo
    > obj.toString()  // inherited
    '[object Object]'

If we turn PropertyChecker into a constructor, we can use it for ECMAScript 6 classes via extends:

    function PropertyChecker() { }
    PropertyChecker.prototype = new Proxy(···);
    
    class Point extends PropertyChecker {
        constructor(x, y) {
            this.x = x;
            this.y = y;
        }
    }
    
    let p = new Point(5, 7);
    console.log(p.x); // 5
    console.log(p.z); // ReferenceError

If you are worried about accidentally creating properties, you have two options: You can either create a proxy that traps set. Or you can make an object obj non-extensible via Object.preventExtensions(obj), which means that JavaScript doesn’t let you add new (own) properties to obj.

Negative array indices

Some array methods let you refer to the last element via -1, to the second-to-last element via -2, etc. For example:

    > ['a', 'b', 'c'].slice(-1)
    [ 'c' ]

Alas, that doesn’t work when accessing elements via the bracket operator ([]). We can, however, use proxies to add that capability. The following function createArray() creates arrays that support negative indices. It does so by wrapping proxies around array instances. The proxies intercept the get operation that is triggered by the bracket operator.

    function createArray(...elements) {
        let handler = {
            get(target, propKey, receiver) {
                let index = Number(propKey);
                // Sloppy way of checking for negative indices
                if (index < 0) {
                    propKey = String(target.length + index);
                }
                return Reflect.get(target, propKey, receiver);
            }
        };
        // Wrap a proxy around an array
        let target = [];
        target.push(...elements);
        return new Proxy(target, handler);
    }
    let arr = createArray('a', 'b', 'c');
    console.log(arr[-1]); // c

Acknowledgement: The idea for this example comes from a blog post by hemanth.hm.

Data binding

Data binding is about syncing data between objects. One popular use case are widgets based on the MVC (Model View Controler) pattern: With data binding, the view (the widget) stays up-to-date if you change the model (the data visualized by the widget).

To implement data binding, you have to observe and react to changes made to an object. In the following code snippet, I sketch how that could work for an array.

    let array = [];
    let observedArray = new Proxy(array, {
        set(target, propertyKey, value, receiver) {
            console.log(propertyKey+'='+value);
            target[propertyKey] = value;
        }
    });
    observedArray.push('a');

Output:

    0=a
    length=1

Data binding is a complex topic. Given its popularity and concerns over proxies not being performant enough, a dedicated mechanism has been created for data binding: Object.observe(). It will probably be part of ECMAScript 7 and is already supported by Chrome.

Consult Addy Osmani’s article “Data-binding Revolutions with Object.observe()” for more information on Object.observe().

Revocable references

Revocable references work as follows: A client is not allowed to access an important resource (an object) directly, only via a reference (an intermediate object, a wrapper around the resource). Normally, every operation applied to the reference is forwarded to the resource. After the client is done, the resource is protected by revoking the reference, by switching it off. Henceforth, applying operations to the reference throws exceptions and nothing is forwarded, anymore.

In the following example, we create a revocable reference for a resource. We then read one of the resource’s properties via the reference. That works, because the reference grants us access. Next, we revoke the reference. Now the reference doesn’t let us read the property, anymore.

    let resource = { x: 11, y: 8 };
    let {reference, revoke} = createRevocableReference(resource);
    
    // Access granted
    console.log(reference.x); // 11
    
    revoke();
    
    // Access denied
    console.log(reference.x); // TypeError: Revoked

Proxies are ideally suited for implementing revocable references, because they can intercept and forward operations. This is a simple proxy-based implementation of createRevocableReference:

    function createRevocableReference(target) {
        let enabled = true;
        return {
            reference: new Proxy(target, {
                get(target, propKey, receiver) {
                    if (!enabled) {
                        throw new TypeError('Revoked');
                    }
                    return Reflect.get(target, propKey, receiver);
                },
                has(target, propKey) {
                    if (!enabled) {
                        throw new TypeError('Revoked');
                    }
                    return Reflect.has(target, propKey);
                },
                ···
            }),
            revoke() {
                enabled = false;
            },
        };
    }

The code can be simplified via the proxy-as-handler technique from the previous section. This time, the handler basically is the Reflect object. Thus, the get trap normally returns the appropriate Reflect method. If the reference has been revoked, a TypeError is thrown, instead.

    function createRevocableReference(target) {
        let enabled = true;
        let handler = new Proxy({}, {
            get(dummyTarget, trapName, receiver) {
                if (!enabled) {
                    throw new TypeError('Revoked');
                }
                return Reflect[trapName];
            }
        });
        return {
            reference: new Proxy(target, handler),
            revoke() {
                enabled = false;
            },
        };
    }

However, you don’t have to implement revocable references yourself, because ECMAScript 6 lets you create proxies that can be revoked. This time, the revoking happens in the proxy, not in the handler. All the handler has to do is forward every operation to the target. As we have seen that happens automatically if the handler doesn’t implement any traps.

    function createRevocableReference(target) {
        let handler = {}; // forward everything
        let { proxy, revoke } = Proxy.revocable(target, handler);
        return { reference: proxy, revoke };
    }
Membranes

Membranes build on the idea of revocable references: Environments that are designed to run untrusted code wrap a membrane around that code to isolate it and keep the rest of the system safe. Objects pass the membrane in two directions:

  • The code may receive objects from the outside.
  • Or it may hand objects to the outside.

In both cases, revocable references are wrapped around the objects. Objects returned by wrapped functions or methods are also wrapped.

Once the untrusted code is done, all of those references are revoked. As a result, none of its code on the outside can be executed anymore and outside objects that it has cease to work, as well. The Caja Compiler is “a tool for making third party HTML, CSS and JavaScript safe to embed in your website”. It uses membranes to achieve this task.

Other use cases

There are more use cases for proxies. For example:

  • Local placeholders that forward method invocations to remote objects. Similar: web service example.
  • Data access objects for databases: reading and writing to the object reads and writes to the database. Similar: web service example.
  • Profiling: Intercept method invocations to track how much time is spent in each method. Similar: tracing example.
  • Type checking: Nicholas Zakas has used proxies to type-check objects.

The design of the proxy API

In this section, we go deeper into how proxies work and why they work that way.

Stratification: keeping base level and meta level separate

Firefox has allowed you to do some interceptive meta programming for a while: If you define a method whose name is __noSuchMethod__, it is notified whenever a method is called that doesn’t exist. The following is an example of using __noSuchMethod__.

    let obj = {
        __noSuchMethod__: function (name, args) {
            console.log(name+': '+args);
        }
    };
    // Neither of the following two methods exist,
    // but we can make it look like they do
    obj.foo(1);    // Output: foo: 1
    obj.bar(1, 2); // Output: bar: 1,2

Thus, __noSuchMethod__ works similarly to a proxy trap. In contrast to proxies, the trap is an own or inherited method of the object whose operations we want to intercept. The problem with that approach is that base level and meta level are mixed. Base-level code may accidentally invoke or see a meta level method and there is the possibility of accidentally defining a meta level method.

Even in standard ECMAScript 5, base level and meta level are sometimes mixed. For example, the following meta programming mechanisms can fail, because they exist at the base level:

  • obj.hasOwnProperty(propKey): This call can fail if a property in the prototype chain overrides the built-in implementation. For example, it fails if obj is { hasOwnProperty: null }. Safe ways to call this method are Object.prototype.hasOwnProperty.call(obj, propKey) and its abbreviated version {}.hasOwnProperty.call(obj, propKey).
  • func.call(···), func.apply(···): For these two methods, problem and solution are the same as with hasOwnProperty.
  • obj.__proto__: In most JavaScript engines, __proto__ is a special property that lets you get and set the prototype of obj. Hence, when you use objects as dictionaries, you must be careful to avoid __proto__ as a property key.

By now, it should be obvious that making (base level) property keys special is problematic. Therefore, proxies are stratified – base level (the proxy object) and meta level (the handler object) are separate.

Virtual objects versus wrappers

Proxies are used in two roles:

  • As wrappers, they wrap their targets, they control access to them. Examples of wrappers are: revocable resources and tracing proxies.

  • As virtual objects, they are simply objects with special behavior and their targets don’t matter. An example is a proxy that forwards method calls to a remote object.

An earlier design of the proxy API conceived proxies as purely virtual objects. However, it turned out that even in that role, a target was useful, to enforce invariants (which is explained later) and as a fallback for traps that the handler doesn’t implement.

Transparent virtualization and handler encapsulation

Proxies are shielded in two ways:

  • It is impossible to determine whether an object is a proxy or not (transparent virtualization).
  • You can’t access a handler via its proxy (handler encapsulation).

Both principles give proxies considerable power for impersonating other objects. One reason for enforcing invariants (as explained later) is to keep that power in check.

If you do need a way to tell proxies apart from non-proxies, you have to implement it yourself. The following code is a module lib.js that exports two functions: one of them creates proxies, the other one determines whether an object is one of those proxies.

    // lib.js
    
    let proxies = new WeakSet();
    
    export function createProxy(obj) {
        let handler = {};
        let proxy = new Proxy(obj, handler);
        proxies.add(proxy);
        return proxy;
    }
    
    export function isProxy(obj) {
        return proxies.has(obj);
    }

This module uses the ECMAScript 6 data structure WeakSet for keeping track of proxies. WeakSet is ideally suited for this purpose, because it doesn’t prevent its elements from being garbage-collected.

The next example shows how lib.js can be used.

    // main.js
    
    import { createProxy, isProxy } from './lib.js';
    
    let p = createProxy({});
    console.log(isProxy(p)); // true
    console.log(isProxy({})); // false

The meta object protocol and proxy traps

This section examines how JavaScript is structured internally and how the set of proxy traps was chosen.

The term protocol is highly overloaded in computer science. One definition is:

A prototcol is about achieving tasks via an object, it comprises a set of methods plus a set of rules for using them.

Note that this definition is different from viewing protocols as interfaces (as, for example, Objective C does), because it includes rules.

The ECMAScript specification describes how to execute JavaScript code. It includes a protocol for handling objects. This protocol operates at a meta level and is sometimes called the meta object protocol (MOP). The JavaScript MOP consists of own internal methods that all objects have. “Internal” means that they exist only in the specification (JavaScript engines may or may not have them) and are not accessible from JavaScript. The names of internal methods are written in double square brackets.

The internal method for getting properties is called [[Get]]. If we pretend that property names with square brackets are legal, this method would roughly be implemented as follows in JavaScript.

    // Method definition
    [[Get]](propKey, receiver) {
        let desc = this.[[GetOwnProperty]](propKey);
        if (desc === undefined) {
            let parent = this.[[GetPrototypeOf]]();
            if (parent === null) return undefined;
            return parent.[[Get]](propKey, receiver); // (*)
        }
        if ('value' in desc) {
            return desc.value;
        }
        let getter = desc.get;
        if (getter === undefined) return undefined;
        return getter.[[Call]](receiver, []);
    }

The MOP methods called in this code are:

  • [[GetOwnProperty]] (trap getOwnPropertyDescriptor)
  • [[GetPrototypeOf]] (trap getPrototypeOf)
  • [[Get]] (trap get)
  • [[Call]] (trap apply)

In line (*) you can see why proxies in a prototype chain find out about get if a property isn’t found in an “earlier” object: If there is no own property whose key is propKey, the search continues in the prototype parent of this.

Fundamental versus derived operations. You can see that [[Get]] calls other MOP operations. Operations that do that are called derived. Operations that don’t depend on other operations are called fundamental.

The MOP of proxies

The meta object protocol of proxies is different from that of normal objects. For normal objects, derived operations call other operations. For proxies, each operation is either intercepted by a handler method or forwarded to the target.

What operations should be interceptable via proxies? One possibility is to only provide traps for fundamental operations. The alternative is to include some derived operations. The advantage of derived traps is that they increase performance and are more convenient: If there wasn’t a trap for get, you’d have to implement its functionality via getOwnPropertyDescriptor. One problem with derived traps is that they can lead to proxies behaving inconsistently. For example, get may return a value that is different from the value stored in the descriptor returned by getOwnPropertyDescriptor.

Selective intercession: what operations should be interceptable?

Intercession by proxies is selective: you can’t intercept every language operation. Why were some operations excluded? Let’s look at two reasons.

First, stable operations are not well suited for intercession. An operation is stable if it always produces the same results for the same arguments. If a proxy can trap a stable operation, it can become unstable and thus unreliable. Strict equality (===) is one such stable operation. It can’t be trapped and its result is computed by treating the proxy itself as just another object. Another way of maintaining stability is by applying an operation to the target instead of the proxy. As explained later, when we look at how invariants are enfored for proxies, this happens when Object.getPrototypeOf() is applied to a proxy whose target is non-extensible.

A second reason for not making more operations interceptable is that intercession means executing custom code in situations where that normally isn’t possible. The more this interleaving of code happens, the harder it is to understand and debug a program.

Traps: “get” versus “invoke”

If you want to create virtual methods via ECMAScript 6 proxies, you have to return functions from a get trap. That raises the question: why not introduce an extra trap for method invocations (e.g. invoke)? That would enable us to distinguish between:

  • Getting properties via obj.prop (trap get)
  • Invoking methods via obj.prop() (trap invoke)

There are two reasons for not doing so.

First, not all implementations distinguish between get and invoke. For example, Apple’s JavaScriptCore doesn’t.

Second, extracting a method and invoking it later via call() or apply() should have the same effect as invoking the method via dispatch. In other words, the following two variants should work equivalently. If there was an extra trap invoke then that equivalence would be harder to maintain.

    // Variant 1: call via dynamic dispatch
    let result = obj.m();
    
    // Variant 2: extract and call directly
    let m = obj.m;
    let result = m.call(obj);

Only possible with invoke. Some things can only be done if you are able to distinguish between get and invoke. Those things are therefore impossible with the current proxy API. Two examples are: auto-binding and intercepting missing methods.

First, by making a proxy the prototype of an object obj, you can automatically bind methods:

  • Retrieving the value of a method m via obj.m returns a function whose this is bound to obj.
  • obj.m() performs a method call.

Auto-binding helps with using methods as callbacks. For example, variant 2 from the previous example becomes simpler:

    let boundMethod = obj.m;
    let result = boundMethod();

Second, invoke lets a proxy emulate the previously mentioned __noSuchMethod__ mechanism that Firefox supports. The proxy would again become the prototype of an object obj. It would react differently depending on how an unknown property foo is accessed:

  • If you read that property via obj.foo, no intercession happens and undefined is returned.
  • If you make the method call obj.foo() then the proxy intercepts and, e.g., notifies a callback.

Enforcing invariants for proxies

Before we look at what invariants are and how they are enforced for proxies, let’s review how objects can be protected via non-extensibility and non-configurability.

Protecting objects

There are two ways of protecting objects:

  • non-extensibility protects objects
  • non-configurability protects properties (or rather, their attributes)

Non-extensibility. If an object is non-extensible, you can’t add properties and you can’t change its prototype:

    'use strict'; // switch on strict mode to get TypeErrors
    
    let obj = Object.preventExtensions({});
    console.log(Object.isExtensible(obj)); // false
    obj.foo = 123; // TypeError: object is not extensible
    Object.setPrototypeOf(obj, null); // TypeError: object is not extensible

Non-configurability. All the data of a property is stored in attributes. A property is like a record and attributes are like the fields of that record. Examples of attributes:

  • The attribute value holds the value of a property.
  • The boolean attribute writable controls whether a property’s value can be changed.
  • The boolean attribute configurable controls whether a property’s attributes can be changed.

Thus, if a property is both non-writable and non-configurable, it is read-only and remains that way:

    'use strict'; // switch on strict mode to get TypeErrors
    
    let obj = {};
    Object.defineProperty(obj, 'foo', {
        value: 123,
        writable: false,
        configurable: false
    });
    console.log(obj.foo); // 123
    obj.foo = 'a'; // TypeError: Cannot assign to read only property
    
    Object.defineProperty(obj, 'foo', {
        configurable: true
    }); // TypeError: Cannot redefine property

For more details on these topics (including how Object.defineProperty() works) consult the following sections in “Speaking JavaScript”:

Enforcing invariants

Traditionally, non-extensibility and non-configurability are:

  • Universal: they work for all objects.
  • Monotonic: once switched on, they can’t be switched off again.

These and other characteristics that remain unchanged in the face of language operations are called invariants. With proxies, it is easy to violate invariants, as they are not intrinsically bound by non-extensibility etc.

The proxy API prevents proxies from violating invariants by checking the parameters and results of handler methods. Non-extensibility and non-configurability are enforced by using the target object for bookkeeping. The following are a few examples of invariants (for an arbitrary object obj) and how they are enforced for proxies (an exhaustive list is given at the end of this post):

  • Invariant: Object.isExtensible(obj) must return a boolean.
    • Enforced by coercing the value returned by the handler to a boolean.
  • Invariant: Object.getOwnPropertyDescriptor(obj, ···) must return an object or undefined.
    • Enforced by throwing a TypeError if the handler doesn’t return an appropriate value.
  • Invariant: If Object.preventExtensions(obj) returns true then all future calls must return false and obj must now be non-extensible.
    • Enforced by throwing a TypeError if the handler returns true, but the target object is not extensible.
  • Invariant: Once an object has been made non-extensible, Object.isExtensible(obj) must always return false.
    • Enforced by throwing a TypeError if the result returned by the handler is not the same (after coercion) as Object.isExtensible(target).

Enforcing invariants has the following benefits:

  • Proxies work like all other objects with regard to extensibility and configurability. Therefore, universality is maintained. This is achieved without preventing proxies from virtualizing (impersonating) protected objects.
  • A protected object can’t be misrepresented by wrapping a proxy around it. Misrepresentation can be caused by bugs or by malicious code.

The following sections give examples of invariants being enforced.

Example: the prototype of a non-extensible target must be represented faithfully

In response to the getPrototypeOf trap, the proxy must return the target’s prototype if the target is non-extensible.

To demonstrate this invariant, let’s create a handler that returns a prototype that is different from the target’s prototype:

    let fakeProto = {};
    let handler = {
        getPrototypeOf(t) {
            return fakeProto;
        }
    };

Faking the prototype works if the target is extensible:

    let extensibleTarget = {};
    let ext = new Proxy(extensibleTarget, handler);
    console.log(Object.getPrototypeOf(ext) === fakeProto); // true

We do, however, get an error if we fake the prototype for a non-extensible object.

    let nonExtensibleTarget = {};
    Object.preventExtensions(nonExtensibleTarget);
    let nonExt = new Proxy(nonExtensibleTarget, handler);
    Object.getPrototypeOf(nonExt); // TypeError
Example: non-writable non-configurable target properties must be represented faithfully

If the target has a non-writable non-configurable property then the handler must return that property’s value in response to a get trap. To demonstrate this invariant, let’s create a handler that always returns the same value for properties.

    let handler = {
        get(target, propKey) {
            return 'abc';
        }
    };
    let target = Object.defineProperties(
        {}, {
            foo: {
                value: 123,
                writable: true,
                configurable: true
            },
            bar: {
                value: 456,
                writable: false,
                configurable: false
            },
        });
    let proxy = new Proxy(target, handler);

Property target.foo is not both non-writable and non-configurable, which means that the handler is allowed to pretend that it has a different value:

    > proxy.foo
    'abc'

However, property target.bar is both non-writable and non-configurable. Therefore, we can’t fake its value:

    > proxy.bar
    TypeError: Invariant check failed

Reference: the proxy API

This section serves as a quick reference for the proxy API: the global objects Proxy and Reflect.

Creating proxies

There are two ways to create proxies:

  • proxy = new Proxy(target, handler)
    Creates a new proxy object with the given target and the given handler.

  • {proxy, revoke} = Proxy.revocable(target, handler) Creates a proxy that can be revoked via the function revoke. revoke can be called multiple times, but only the first call has an effect and switches proxy off. Afterwards, any operation performed on proxy leads to a TypeError being thrown.

Handler methods

This subsection explains what traps can be implemented by handlers and what operations trigger them. Several traps return boolean values. For the traps has and isExtensible, the boolean is the result of the operation. For all other traps, the boolean indicates whether the operation succeeded or not.

Traps for all objects:

  • defineProperty(target, propKey, propDesc)boolean
    • Object.defineProperty(proxy, propKey, propDesc)
  • deleteProperty(target, propKey)boolean
    • delete proxy[propKey]
    • delete proxy.foo // propKey = 'foo'
  • enumerate(target)Iterator
    • for (x in proxy) ···
  • get(target, propKey, receiver)any
    • receiver[propKey]
    • receiver.foo // propKey = 'foo'
  • getOwnPropertyDescriptor(target, propKey)PropDesc|Undefined
    • Object.getOwnPropertyDescriptor(proxy, propKey)
  • getPrototypeOf(target)Object|Null
    • Object.getPrototypeOf(proxy)
  • has(target, propKey)boolean
    • propKey in proxy
  • isExtensible(target)boolean
    • Object.isExtensible(proxy)
  • ownKeys(target)Array<PropertyKey>
    • Object.getOwnPropertyPropertyNames(proxy) (only uses string-valued keys)
    • Object.getOwnPropertyPropertySymbols(proxy) (only uses symbol-valued keys)
    • Object.keys(proxy) (only uses enumerable string-valued keys; enumerability is checked via Object.getOwnPropertyDescriptor)
  • preventExtensions(target)boolean
    • Object.preventExtensions(proxy)
  • set(target, propKey, value, receiver)boolean
    • receiver[propKey] = value
    • receiver.foo = value // propKey = 'foo'
  • setPrototypeOf(target, proto)boolean
    • Object.setPrototypeOf(proxy, proto)

Traps for functions (available if target is a function):

  • apply(target, thisArgument, argumentsList)any
    • proxy.apply(thisArgument, argumentsList)
    • proxy.call(thisArgument, ...argumentsList)
    • proxy(...argumentsList)
  • construct(target, argumentsList)Object
    • new proxy(..argumentsList)
Fundamental operations versus derived operations

The following operations are fundamental, they don’t use other operations to do their work: apply, defineProperty, deleteProperty, getOwnPropertyDescriptor, getPrototypeOf, isExtensible, ownKeys, preventExtensions, setPrototypeOf

All other operations are derived, they can be implemented via fundamental operations. For example, for data properties, get can be implemented by iterating over the prototype chain via getPrototypeOf and calling getOwnPropertyDescriptor for each chain member until either an own property is found or the chain ends.

Invariants

Invariants are safety constraints for handlers. This subsection documents what invariants are enforced by the proxy API and how. Whenever you read “the handler must do X” below, it means that a TypeError is thrown if it doesn’t. Some invariants restrict return values, others restrict parameters. Ensuring the correct return value of a trap is ensured in two ways: Normally, an illegal value means that a TypeError is thrown. But whenever a boolean is expected, coercion is used to convert non-booleans to legal values.

This is the complete list of invariants that are enforced (source: ECMAScript 6 specification):

  • apply(target, thisArgument, argumentsList)
    • No invariants are enforced.
  • construct(target, argumentsList)
    • The result returned by the handler must be an object (not null or a primitive value).
  • defineProperty(target, propKey, propDesc)
    • If the target is not extensible then propDesc can’t create a property that the target doesn’t already have.
    • If propDesc sets the attribute configurable to false then the target must have a non-configurable own property whose key is propKey.
    • If propDesc was used to (re)define an own property for the target then that must not cause an exception. An exception is thrown if a change is forbidden by the attributes writable and configurable.
  • deleteProperty(target, propKey)
    • Non-configurable own properties of the target can’t be deleted.
  • enumerate(target)
    • The handler must return an object.
  • get(target, propKey, receiver)
    • If the target has an own, non-writable, non-configurable data property whose key is propKey then the handler must return that property’s value.
    • If the target has an own, non-configurable, getter-less accessor property then the handler must return undefined.
  • getOwnPropertyDescriptor(target, propKey)
    • The handler must return either an object or undefined.
    • Non-configurable own properties of the target can’t be reported as non-existent by the handler.
    • If the target is non-extensible then exactly the target’s own properties must be reported by the handler as existing (and none of them as missing).
    • If the handler reports a property as non-configurable then that property must be a non-configurable own property of the target.
    • If the result returned by the handler were used to (re)define an own property for the target then that must not cause an exception. An exception is thrown if the change is not allowed by the attributes writable and configurable. Therefore, the handler can’t report a non-configurable property as configurable and it can’t report a different value for a non-configurable non-writable property.
  • getPrototypeOf(target)
    • The result must be either an object or null.
    • If the target object is not extensible then the handler must return the prototype of the target object.
  • has(target, propKey)
    • A handler must not hide (report as non-existent) a non-configurable own property of the target.
    • If the target is non-extensible then no own property of the target may be hidden.
  • isExtensible(target)
    • The result returned by the handler is coerced to boolean.
    • After coercion to boolean, the value returned by the handler must be the same as target.isExtensible().
  • ownKeys(target)
    • The handler must return an object, which treated as array-like and converted into an array.
    • Each element of the result must be either a string or a symbol.
    • The result must contain the keys of all non-configurable own properties of the target.
    • If the target is not extensible then the result must contain exactly the keys of the own properties of the target (and no other values).
  • preventExtensions(target)
    • The result returned by the handler is coerced to boolean.
    • If the handler returns a truthy value (indicating a successful change) then target.isExtensible() must be false afterwards.
  • set(target, propKey, value, receiver)
    • If the target has an own, non-writable, non-configurable data property whose key is propKey then value must be the same as the value of that property (i.e., the property can’t be changed).
    • If the target has an own, non-configurable, setter-less accessor property then a TypeError is thrown (i.e., such a property can’t be set).
  • setPrototypeOf(target, proto)
    • The result returned by the handler is coerced to boolean.
    • If the target is not extensible, the prototype can’t be changed. This is enforced as follows: If the target is not extensible and the handler returns a truthy value (indicating a successful change) then proto must be the same as the prototype of the target. Otherwise, a TypeError is thrown.

The prototype chain

The following operations of normal objects perform operations on objects in the prototype chain (source: ECMAScript 6 specification). Therefore, if one of the objects in that chain is a proxy, its traps are triggered. The specification implements the operations as internal own methods (that are not visible to JavaScript code). But in this section, we pretend that they are normal methods that have the same names as the traps. The parameter target becomes the receiver of the method call.

  • target.enumerate()
    Traverses the prototype chain of target via getPrototypeOf. Per object, it retrieves the keys via ownKeys and examines whether a property is enumerable via getOwnPropertyDescriptor.
  • target.get(propertyKey, receiver)
    If target has no own property with the given key, get is invoked on the prototype of target.
  • target.has(propertyKey)
    Similarly to get, has is invoked on the prototype of target if target has no own property with the given key.
  • target.set(propertyKey, value, receiver)
    Similarly to get, set is invoked on the prototype of target if target has no own property with the given key.

All other operations only affect own properties, they have no effect on the prototype chain.

Reflect

The global object Reflect implements all interceptable operations of the JavaScript meta object protocol as methods. The names of those methods are the same as those of the handler methods, which, as we have seen, helps with forwarding operations from the handler to the target.

  • Reflect.apply(target, thisArgument, argumentsList)any
    Better version of Function.prototype.apply().
  • Reflect.construct(target, argumentsList, newTarget?)Object
    The new operator as a function.
  • Reflect.defineProperty(target, propertyKey, propDesc)boolean
    Similar to Object.defineProperty().
  • Reflect.deleteProperty(target, propertyKey)boolean
    The delete operator as a function.
  • Reflect.enumerate(target)Iterator
    Returns an iterater over all enumerable string property keys of target. In other words, the iterator returns all values that the for-in loop would iterate over.
  • Reflect.get(target, propertyKey, receiver?)any
    A function that gets properties.
  • Reflect.getOwnPropertyDescriptor(target, propertyKey)PropDesc|Undefined
    Same as Object.getOwnPropertyDescriptor().
  • Reflect.getPrototypeOf(target)Object|Null
    Same as Object.getPrototypeOf().
  • Reflect.has(target, propertyKey)boolean
    The in operator as a function.
  • Reflect.isExtensible(target)boolean
    Same as Object.isExtensible().
  • Reflect.ownKeys(target)Array<PropertyKey>
    Returns all own property keys (strings and symbols!) in an array.
  • Reflect.preventExtensions(target)boolean
    Similar to Object.preventExtensions().
  • Reflect.set(target, propertyKey, value, receiver?)boolean
    A function that sets properties.
  • Reflect.setPrototypeOf(target, proto)boolean
    The new standard way of setting the prototype of an object. The current non-standard way that works in most engines is to set the special property __proto__.

Several methods have boolean results. For has and isExtensible, they are the results of the operation. For the remaining methods, they indicate whether the operation succeeded.

Apart from forwarding operations, why is Reflect useful [4]?

  • Different return values: Reflect duplicates the following methods of Object, but its methods return booleans indicating whether the operation succeeded (where the Object methods return the object that was modified).
    • Object.defineProperty(obj, propKey, propDesc)Object
    • Object.preventExtensions(obj)Object
    • Object.setPrototypeOf(obj, proto)Object
  • Operators as functions: The following Reflect methods implement functionality that is otherwise only available via operators:
    • Reflect.construct(target, argumentsList, newTarget?)Object
    • Reflect.deleteProperty(target, propertyKey)boolean
    • Reflect.get(target, propertyKey, receiver?)any
    • Reflect.has(target, propertyKey)boolean
    • Reflect.set(target, propertyKey, value, receiver?)boolean
  • The for-in loop as an iterator: This is rarely useful, but if you need it, you can get an iterator over all enumerable (own and inherited) string property keys of an object.
    • Reflect.enumerate(target)Iterator
  • Shorter version of apply: The only safe way to invoke the built-in function method apply is via Function.prototype.apply.call(func, thisArg, args) (or similar). Reflect.apply(func, thisArg, args) is cleaner and shorter.

State of implementations

As usual, Kangax’ ES6 compatibility table is the best way of finding out how well engines support proxies. As of December 2014, Internet Explorer has the most complete support and Firefox supports some of the API (caveats: get doesn’t work properly, getPrototypeOf is not supported yet and Reflect is empty). No other browser or engine currently supports proxies.

Conclusion

This concludes our in-depth look at the proxy API. For each application, you have to take performance into consideration and – if necessary – measure. Proxies may not always be fast enough. On the other hand, performance is often not crucial and it is nice to have the meta programming power that proxies give us. As we have seen, there are numerous use cases they can help with.

Acknowledgements

Thanks go to Tom Van Cutsem: his paper [1] is the most important source of this blog post and he kindly answered questions about the proxy API that I had.

Technical Reviewers:

Further reading

  1. On the design of the ECMAScript Reflection API” by Tom Van Cutsem and Mark Miller. Technical report, 2012.
  2. The Art of the Metaobject Protocol” by Gregor Kiczales, Jim des Rivieres and Daniel G. Bobrow. Book, 1991.
  3. Putting Metaclasses to Work: A New Dimension in Object-Oriented Programming” by Ira R. Forman and Scott H. Danforth. Book, 1999.
  4. Harmony-reflect: Why should I use this library?” by Tom Van Cutsem.

2 comments:

Axel Rauschmayer said...

myArray = []; does not seem to really "clear" anything. It seems to just reassigns the variable, without any context as to the variable's previous state or "type". It is as meaningful as the (absurd) myArray = '';

Axel Rauschmayer said...

You seem to understand what sharing is about. Many people don’t.