Memoization is best technique to save on memory or CPU cycles when we deal with repeated operations. For detailed insight see: http://en.wikipedia.org/wiki/Memoization
Features
Works with any type of function arguments – no serialization is needed
If we work with large result sets, or memoize hot functions, default mode may not perform as fast as we expect. In that case it's good to run memoization in primitive mode. To provide fast access, results are saved in hash instead of an array. Generated hash ids are result of arguments to string conversion. Mind that this mode will work correctly only if stringified arguments produce unique strings.
memoized = memoize(fn, { primitive: true });memoized("/path/one");memoized("/path/one"); // Cache hit
Cache id resolution (normalization)
By default cache id for given call is resolved either by:
Direct Comparison of values passed in arguments as they are. In such case two different objects, even if their characteristics is exactly same (e.g. var a = { foo: 'bar' }, b = { foo: 'bar' }) will be treated as two different values.
Comparison of stringified values of given arguments (primitive mode), which serves well, when arguments are expected to be primitive values, or objects that stringify naturally do unique values (e.g. arrays)
Still above two methods do not serve all cases, e.g. if we want to memoize function where arguments are hash objects which we do not want to compare by instance but by its content.
Writing custom cache id normalizers
There's a normalizer option through which we can pass custom cache id normalization function.
Memoizing on dictionary (object hash) arguments)
if we want to memoize a function where argument is a hash object which we do not want to compare by instance but by its content, then we can achieve it as following:
var mfn = memoize( function (hash) { // body of memoized function }, { normalizer: function (args) { // args is arguments object as accessible in memoized function return JSON.stringify(args[0]); }, });mfn({ foo: "bar" });mfn({ foo: "bar" }); // Cache hit
If additionally we want to ensure that our logic works well with different order of properties, we need to imply an additional sorting, e.g.
const deepSortedEntries = object => Object.entries(object) .map(([key, value]) => { if (value && typeof value === "object") return [key, deepSortedEntries(value)]; return [key, value]; }) .sort();var mfn = memoize( function (hash) { // body of memoized function }, { normalizer: function (args) { // args is arguments object as accessible in memoized function return JSON.stringify(deepSortedEntries(args[
Argument resolvers
When we're expecting arguments of certain type it's good to coerce them before doing memoization. We can do that by passing additional resolvers array:
Note. If your arguments are collections (arrays or hashes) that you want to memoize by content (not by self objects), you need to cast them to strings, for it's best to just use primitive mode. Arrays have standard string representation and work with primitive mode out of a box, for hashes you need to define toString method, that will produce unique string descriptions, or rely on JSON.stringify.
Similarly if you want to memoize functions by their code representation not by their objects, you should use primitive mode.
Memoizing asynchronous functions
Promise returning functions
With promise option we indicate that we memoize a function that returns promise.
The difference from natural behavior is that in case when promise was rejected with exception,
the result is immediately removed from memoize cache, and not kept as further reusable result.
var afn = function (a, b) { return new Promise(function (res) { res(a + b); });};memoized = memoize(afn, { promise: true });memoized(3, 7);memoized(3, 7); // Cache hit
Important notice on internal promises handling
Default handling stands purely on then which has side-effect of muting eventual unhandled rejection notifications.
Alternatively we can other (explained below), by stating with promise option desired mode:
then(default). Values are resolved purely by passing callbacks to promise.then. Side effect is that eventual unhandled rejection on given promise
come with no logged warning!, and that to avoid implied error swallowing both states are resolved tick after callbacks were invoked
done Values are resolved purely by passing callback to done method. Side effect is that eventual unhandled rejection on given promise come with no logged warning!.
done:finally The only method that may work with no side-effects assuming that promise implementaion does not throw unconditionally
if no onFailure callback was passed to done, and promise error was handled by other consumer (this is not commonly implemented done behavior). Otherwise side-effect is that exception is thrown on promise rejection (highly not recommended)
Node.js callback style functions
With async option we indicate that we memoize asynchronous (Node.js style) function
Operations that result with an error are not cached.
afn = function (a, b, cb) { setTimeout(function () { cb(null, a + b); }, 200);};memoized = memoize(afn, { async: true });memoized(3, 7, function (err, res) { memoized(3, 7, function (err, res) { // Cache hit });});memoized(3, 7, function (err, res) { // Cache hit});
Memoizing methods
When we are defining a prototype, we may want to define a method that will memoize it's results in relation to each instance. A basic way to obtain that would be:
var Foo = function () { this.bar = memoize(this.bar.bind(this), { someOption: true }); // ... constructor logic};Foo.prototype.bar = function () { // ... method logic};
There's a lazy methods descriptor generator provided:
var d = require("d");var memoizeMethods = require("memoizee/methods");var Foo = function () { // ... constructor logic};Object.defineProperties( Foo.prototype, memoizeMethods({ bar: d( function () { // ... method logic }, { someOption: true } ), }));
WeakMap based configurations
In this case memoization cache is not bound to memoized function (which we may want to keep forever), but to objects for which given results were generated.
This mode works only for functions of which first argument is expected to be an object.
It can be combined with other options mentioned across documentation. However due to WeakMap specificity global clear is not possible.
var memoize = require("memoizee/weak");var memoized = memoize(function (obj) { return Object.keys(obj); });var obj = { foo: true, bar: false };memoized(obj);memoized(obj); // Cache hit
Cache handling
Manual clean up:
Delete data for particular call.
memoized.delete("foo", true);
Arguments passed to delete are treated with same rules as input arguments passed to function
Clear all cached data:
memoized.clear();
Expire cache after given period of time
With maxAge option we can ensure that cache for given call is cleared after predefined period of time (in milliseconds)
Additionally we may ask to pre-fetch in a background a value that is about to expire. Pre-fetch is invoked only if value is accessed close to its expiry date. By default it needs to be within at least 33% of maxAge timespan before expire:
Thanks @puzrin for helpful suggestions concerning this functionality
Reference counter
We can track number of references returned from cache, and manually delete them. When the last reference is cleared, the cache is purged automatically:
With max option you can limit cache size, it's backed with LRU algorithm, provided by low-level lru-queue utility.
The size relates purely to count of results we want to keep in cache, it doesn't relate to memory cost associated with cache value (but such feature is likely to be introduced with next version of memoizee).
You can register a callback to be called on each value removed from the cache:
memoized = memoize(fn, { dispose: function (value) { /*…*/ } });var foo3 = memoized("foo", 3);var bar7 = memoized("bar", 7);memoized.clear("foo", 3); // Dispose called with foo3 valuememoized.clear("bar", 7); // Dispose called with bar7 value
Benchmarks
Simple benchmark tests can be found in benchmark folder. Currently it's just plain simple calculation of fibonacci sequences. To run it you need to install other test candidates:
If you want to make sure how much you benefit from memoization or just check if memoization works as expected, loading profile module will give access to all valuable information.
Module needs to be imported before any memoization (that we want to track) is configured. Mind also that running profile module affects performance, it's best not to use it in production environment