@silverwind/nano-memoize v3.0.17
Faster than fast, smaller than micro ... nano-memoizer.
Introduction
Version 3.x.x of nano-memoize was modified to use newer versions of JavaScript built-in classes and take advantage of current v8 loop optimizations. As a result, the minified/brotli size of 3.0.4 at 487 bytes is 30% smaller and is slightly faster that v2.x.x and v1.x.x.
Tests show 'nano-memoize' is overall the smallest and fastest openly available JavaScript memoizer for single and multiple argument functions accepting primitives and objects. However, I have found that benchmarks can vary dramatically from O/S to O/S or Node version to Node version, so I could be wrong. These tests were run on a Windows 10 Pro 64bit 2.8ghz i7 machine with 16GB RAM and Node v18.13.0. Garbage collection was forced between each sample run to minimize its impact on results.
The speed tests are below.
For single primitive argument functions
nano-memoizeis typically 2% faster thanfast-memoizeandmicro-emoize. However, all three are almost always within the margins of error of the other.For single object argument functions
nano-memoizeis consistently at least 15% faster than its closest competitorfast-memoize.For multiple primitive argument functions
nano-memoizeis typically 2% faster thanfast-memoizeandmicro-emoize. However, all three are frequently within the margin of error of the other.For multiple object argument functions
micro-memoizeis always at least 5% faster than its closest competitorsmoizeandnano-memoizewhich alternate in second place typically within each other's margin of error.For custom equality functions 'nano-memoize
andmicro-memoize vary in speed depending on the equality function used.nano-memoizeis always first by at least 15% usinghash-itandmicro-memoizeis always second usinghash-itorlodash.For a real world simulation where only 20% of the function calls are memoized and they take and return mixed argument types,
nanomemoizeis by far fastest. It appears this is becausenanomemoizehandles the passing of mutiple values returned as a string better than others.
The planetheidea/moize library (which claims to be the fastest on average) does not include nano-memoize for comparison and the repository is not accepting comments or a pull request for some technical reason. The repository has been forked and its own benchmarking has been updated and run to confirm the results below.
Starting cycles for functions with a single primitive parameter...
| Name | Ops / sec | Relative margin of error | Sample size |
|---|---|---|---|
| nano-memoize | 150,014,960 | ± 1.26% | 88 |
| fast-memoize | 148,920,057 | ± 1.37% | 88 |
| micro-memoize | 148,405,982 | ± 1.46% | 90 |
| iMemoized | 117,844,380 | ± 1.65% | 88 |
| moize | 96,796,008 | ± 1.80% | 85 |
| underscore | 54,815,804 | ± 1.28% | 87 |
| lodash | 54,617,110 | ± 1.30% | 87 |
| lru-memoize | 41,426,130 | ± 1.16% | 87 |
| memoizee | 31,747,590 | ± 1.52% | 85 |
| addy-osmani | 14,848,551 | ± 1.76% | 82 |
| memoizerific | 12,835,863 | ± 1.84% | 85 |
Starting cycles for functions with a single object parameter...
| Name | Ops / sec | Relative margin of error | Sample size |
|---|---|---|---|
| nano-memoize | 149,402,848 | ± 1.20% | 90 |
| fast-memoize | 121,117,510 | ± 2.41% | 86 |
| iMemoized | 98,480,257 | ± 3.36% | 85 |
| micro-memoize | 97,781,728 | ± 1.36% | 87 |
| moize | 94,370,945 | ± 4.58% | 84 |
| lodash | 41,298,824 | ± 2.60% | 84 |
| lru-memoize | 39,315,541 | ± 1.55% | 86 |
| underscore | 33,803,044 | ± 1.66% | 87 |
| memoizee | 27,298,946 | ± 2.40% | 86 |
| addy-osmani | 16,003,489 | ± 1.72% | 87 |
| memoizerific | 12,502,443 | ± 1.33% | 86 |
Starting cycles for functions with multiple parameters that contain only primitives...
| Name | Ops / sec | Relative margin of error | Sample size |
|---|---|---|---|
| nano-memoize | 57,333,928 | ± 1.83% | 85 |
| micro-memoize | 51,539,993 | ± 2.07% | 83 |
| moize | 49,918,411 | ± 3.30% | 81 |
| lru-memoize | 29,594,676 | ± 2.83% | 79 |
| memoizee | 17,895,742 | ± 2.08% | 87 |
| iMemoized | 12,263,731 | ± 1.99% | 85 |
| memoizerific | 7,804,227 | ± 1.81% | 84 |
| addy-osmani | 4,190,503 | ± 2.03% | 84 |
| fast-memoize | 1,101,952 | ± 2.87% | 81 |
Starting cycles for functions with multiple parameters that contain objects...
| Name | Ops / sec | Relative margin of error | Sample size |
|---|---|---|---|
| micro-memoize | 51,551,926 | ± 2.10% | 82 |
| nano-memoize | 40,243,156 | ± 3.75% | 82 |
| moize | 37,285,146 | ± 16.13% | 65 |
| lru-memoize | 27,946,905 | ± 3.41% | 79 |
| memoizee | 17,209,457 | ± 1.96% | 84 |
| memoizerific | 7,609,760 | ± 4.26% | 76 |
| addy-osmani | 1,212,045 | ± 1.85% | 86 |
| fast-memoize | 563,610 | ± 26.34% | 69 |
Starting cycles for alternative cache types...
| Name | Ops / sec | Relative margin of error | Sample size |
|---|---|---|---|
| nanomemoize deep equals (hash-it isEqual) | 107,990,728 | ± 2.59% | 83 |
| nanomemoize deep equals (lodash isEqual) | 96,543,576 | ± 2.20% | 84 |
| moize deep equals (lodash isEqual) | 88,305,997 | ± 2.55% | 82 |
| micro-memoize deep equals (fast-equals) | 86,511,616 | ± 1.67% | 85 |
| moize deep equals (fast-deep-equal) | 85,948,355 | ± 1.55% | 79 |
| micro-memoize deep equals (lodash isEqual) | 85,231,542 | ± 1.83% | 84 |
| moize deep equals (fast-equals) | 84,844,833 | ± 1.77% | 85 |
| moize deep equals (hash-it isEqual) | 76,605,158 | ± 1.82% | 83 |
| micro-memoize deep equals (hash-it isEqual) | 73,619,713 | ± 2.26% | 82 |
| nanomemoize fast equals (fast-equals deep) | 64,710,177 | ± 1.59% | 79 |
| micro-memoize deep equals (fast-deep-equal) | 62,658,012 | ± 2.95% | 78 |
| nanomemoize fast equals (fast-deep-equals) | 42,443,623 | ± 1.91% | 82 |
Starting real world simulation...
| Name | Ops / sec | Relative margin of error | Sample size |
|---|---|---|---|
| nanomemoizedFunctions | 3,409,618 | ± 0.99% | 94 |
| microMemoizedFunctions | 1,845,306 | ± 0.76% | 92 |
| moizeMemoizedFunctions | 1,842,258 | ± 0.82% | 90 |
| fastMemoizedFunctions | 638,857 | ± 1.30% | 94 |
If you want similar performance for intersection, union or Cartesian product also see:
- https://github.com/anywhichway/intersector
- https://github.com/anywhichway/unionizor
https://github.com/anywhichway/cxproduct
For a complete high performance solution to Cartesian product and set operations for Arrays and Sets with a standardized API, plus the addition of the standard map/reduce/find operations to Set see:
Usage
npm install nano-memoize
API
The API is a subset of the moize API.
const memoized = nanomemoize(sum(a,b) => a + b);
memoized(1,2); // 3
memoized(1,2); // pulled from cachenanomemoize(function,options) returns function
The shape of options is:
{
// only use the provided maxArgs for cache look-up, useful for ignoring final callback arguments
maxArgs: number,
// call last argument of memoized multi-args functions after a number of milliseconds via a timeout after the
// cached result has been returned, perhaps to ensure that callbacks are invoked, does not cache the timemout result
// e.g. nanomemoize(function(a,b,callback) { var result = a + b; callback(result); return result; },{maxArgs:2,callTimeout:0});
callTimeout: number,
// number of milliseconds to cache a result, set to `Infinity` or `-1` to never create timers or expire
maxAge: number,
// the serializer/key generator to use for single argument functions (optional, not recommended)
// must be able to serialize objects and functions, by default a Map is used internally without serializing
serializer: function,
// the equals function to use for multi-argument functions (optional, try to avoid) e.g. deepEquals for objects
equals: function,
// forces the use of multi-argument paradigm, auto set if function has a spread argument or uses `arguments` in its body.
vargs: boolean
}The returned function will also have these methods:
.clear() clears the cache for the function.
.keys() returns an array of arrays with each array being the arguments provided on a call of the function.
.values() returns an array of arrays with each array being the results of a function call with the same index position as the keys.
Release History (reverse chronological order)
2023-04-08 v3.0.10 Enhanced real world simulation.
2023-04-07 v3.0.9 Added real world simulation. Removed .parcel-cache from deployment.
2023-02-22 v3.0.8 Documentation updates.
2023-02-15 v3.0.7 Documentation updates.
2023-02-15 v3.0.6 Documentation updates.
2023-02-15 v3.0.5 Documentation updates.
2023-02-04 v3.0.4 A code walkthrough revealed an opportunity to remove unused code from v2.x.x.
2023-02-02 v3.0.3 Added unit test for maxAge. Adjusted varArg unit tests for more accuracy. Slight optimizations to multi argument memoized functions. Slight improvement to cache clearing that may reduce GC. Updated license file for copyright period. Updated docs on callTimeout for clarity.
2023-02-01 v3.0.2 Fixed https://github.com/anywhichway/nano-memoize/issues/52 with custom equals functions not consistently working. fast-equals or lodash.isEqual now work. Slight performance degradation, but still generally the fastest.
2022-01-29 v3.0.1 Fixed build issue where root index.js was not getting updated.
2023-01-28 v3.0.0 Slight size optimization. 25% speed improvement. Moved to module format. There is a known issue with providing fast-equals or lodash.isEqual as an optional comparison function. Unit tests pass, but the functions fail under load. The hash-it object equivalence function does work. A formerly undocumented method .keyValues() has been deprecated since it is no longer relevant with the new optimizations.
2022-12-08 v2.0.0 Removed callTimeout from TypeScript typings since it was not implemented and there are no plans to implement. Bumped version to 2.0.0 since this may break some users.
2022-10-20 v1.3.1 Bumping version for https://github.com/anywhichway/nano-memoize/pull/41
2022-03-30 v1.3.0 Dropped support for dist and browser directories.
2022-03-15 v1.2.2 Bumped minor version for TS typings and some package dependency updates.
2020-11-02 v1.2.1 Added main: "index.js" reference in package.json.
2020-06-18 v1.2.0 Enhanced multi-arg handling so that arg lengths must match precisely (thanks @amazingmarvin), unless maxArgs is passed as an option.
Also added callTimeout to go ahead and call underlying function to ensure callbacks are invoked if necessary.
2020-05-26 v1.1.11 Fixed Issue 17Fixed https://github.com/anywhichway/nano-memoize/issues/17. It is not known when this bug made its way into the code.
2020-02-30 v1.1.10 Moved growl to dev dependency.
2020-01-30 v1.1.9 Code style improvements.
2019-11-29 v1.1.8 Corrected typos in documentation.
2019-09-25 v1.1.7 Manually created an IE 11 compatible version by removing arrow functions and replacing Object.assign
with a custom function. Minor size increase from 660 to 780 Brotli bytes. Also eliminated complex bind approach in
favor of closures. The latest v8 engine handles closures as fast as bound arguments for our case and we are not as
concerned with older browser performace (particularly given how fast the code is anyway). Converted for loops to run in
reverse, which for our case is faster (but is not always guranteed to be faster).
2019-09-17 v1.1.6 Added a manually transpiled es5_ie11.html file with an Object.assign polyfill to the test directory to verify
compatibility with IE11. Modified unit tests so they are ES5 compatible. All tests pass. Added sideEffects=false to package.json.
2019-06-28 v1.1.5 Improved documentation. Updated version of micro-memoize used for benchmark testing. No code changes.
2019-05-31 v1.1.4 Fixed Issue 7.
2019-04-09 v1.1.3 Fixed Issue 6. Minor speed and size improvements.
2019-04-02 v1.1.2 Speed improvements for multiple arguments. Now consistently faster than fast-memoize and nano-memoize across multiple test runs. Benchmarks run in a new test environment. The benchmarks for v1.1.1 although correct from a relative perspective, grossly understated actual performance due to a corrupt testing environment.
2019-03-25 v1.1.1 Pushed incorrect version with v1.1.0. This corrects the version push.
2019-03-25 v1.1.0 Added use of WeakMap for high-speed caching of single argument functions when passed objects. The serializer option no longer defaults to (value) => JSON.stringify(value) so if you want to treat objects that have the same string representation as the same, you will have to provide a serializer.
2019-03-24 v1.0.8 Updated/corrected documentation.
2019-03-24 v1.0.7 Made smaller and faster. Renamed sngl to sng and mltpl to mlt. Converted all functions to arrow functions except sng and mlt. Simplified and optimized mlt. Removed () around args to single argument arrow function definitions, e.g. (a) => ... became a => .... Replaced the arg signature () in arrow functions with _, which is shorter. Eliminated multiple character variables except for those used in options to request memoization. Collapsed setTimeout into a single location. Defined const I = Infinity. Eliminated () around ? : condition expressions. Changed {} to Object.create(null). Documented all variables. Moved some variables around for clarity. Moved options into a destructing argument.
2019-03-20 v1.0.6 Updated documentation.
2019-03-11 v1.0.5 Now supports setting maxAge to Infinity and no timers will be created to expire caches.
2019-02-26 v1.0.4 Further optimized cache expiration. See Issue 4
2019-02-16 v1.0.3 Fixed README formatting
2019-02-16 v1.0.2 Further optimizations to deal with Issue 4. expireInterval introduced in v1.0.1 removed since it is no longer needed. Also, 25% reduction in size. Code no longer thrashes when memoizing a large number of functions.
2019-02-16 v1.0.1 Memo expiration optimization. Special appreciation to @titoBouzout and @popbee who spent a good bit of time reviewing code for optimization and making recommendations. See Issue 4 for the conversation.
2018-04-13 v1.0.0 Code style improvements.
2018-02-07 v0.1.2 Documentation updates
2018-02-07 v0.1.1 Documentationand benchmark updates
2018-02-01 v0.1.0 Documentation updates. 50 byte decrease.
2018-01-27 v0.0.7b BETA Documentation updates.
2018-01-27 v0.0.6b BETA Minor size and speed improvements.
2018-01-27 v0.0.5b BETA Fixed edge case where multi-arg key may be shorter than current args.
2018-01-27 v0.0.4b BETA Fixed benchmarks. Removed maxSize. More unit tests. Fixed maxAge.
2018-01-27 v0.0.3b BETA More unit tests. Documentation. Benchmark code in repository not yet running.
2018-01-24 v0.0.2a ALPHA Minor speed enhancements. Benchmark code in repository not yet running.
2018=01-24 v0.0.1a ALPHA First public release. Benchmark code in repository not yet running.