fflate

High performance (de)compression in an 8kB package

Why fflate?

fflate (short for fast flate) is the fastest, smallest, and most versatile pure JavaScript compression and decompression library in existence, handily beating pako, tiny-inflate, and UZIP.js in performance benchmarks while being multiple times more lightweight. Its compression ratios are often better than even the original Zlib C library. It includes support for DEFLATE, GZIP, and Zlib data. Data compressed by fflate can be decompressed by other tools, and vice versa.

In addition to the base decompression and compression APIs, fflate supports high-speed ZIP file archiving for an extra 3 kB. In fact, the compressor, in synchronous mode, compresses both more quickly and with a higher compression ratio than most compression software (even Info-ZIP, a C program), and in asynchronous mode it can utilize multiple threads to achieve over 3x the performance of virtually any other utility.

	pako	tiny-inflate	UZIP.js	fflate
Decompression performance	1x	Up to 40% slower	Up to 40% faster	Up to 40% faster
Compression performance	1x	N/A	Up to 25% faster	Up to 50% faster
Base bundle size (minified)	45.6kB	3kB (inflate only)	14.2kB	8kB (3kB for inflate only)
Decompression support	✅	✅	✅	✅
Compression support	✅	❌	✅	✅
ZIP support	❌	❌	✅	✅
Streaming support	✅	❌	❌	✅
GZIP support	✅	❌	❌	✅
Supports files up to 4GB	✅	❌	❌	✅
Doesn't hang on error	✅	❌	❌	✅
Dictionary support	✅	❌	❌	✅
Multi-thread/Asynchronous	❌	❌	❌	✅
Streaming ZIP support	❌	❌	❌	✅
Uses ES Modules	❌	❌	❌	✅

Demo

If you'd like to try fflate for yourself without installing it, you can take a look at the browser demo. Since fflate is a pure JavaScript library, it works in both the browser and Node.js (see Browser support for more info).

Usage

Install fflate:

1npm i fflate # or yarn add fflate, or pnpm add fflate

Import:

1// I will assume that you use the following for the rest of this guide
2import * as fflate from 'fflate';
3
4// However, you should import ONLY what you need to minimize bloat.
5// So, if you just need GZIP compression support:
6import { gzipSync } from 'fflate';
7// Woo! You just saved 20 kB off your bundle with one line.

If your environment doesn't support ES Modules (e.g. Node.js):

1// Try to avoid this when using fflate in the browser, as it will import
2// all of fflate's components, even those that you aren't using.
3const fflate = require('fflate');

If you want to load from a CDN in the browser:

1<!--
2You should use either UNPKG or jsDelivr (i.e. only one of the following)
3
4Note that tree shaking is completely unsupported from the CDN. If you want
5a small build without build tools, please ask me and I will make one manually
6with only the features you need. This build is about 31kB, or 11.5kB gzipped.
7-->
8<script src="https://unpkg.com/fflate@0.8.2"></script>
9<script src="https://cdn.jsdelivr.net/npm/fflate@0.8.2/umd/index.js"></script>
10<!-- Now, the global variable fflate contains the library -->
11
12<!-- If you're going buildless but want ESM, import from Skypack -->
13<script type="module">
14  import * as fflate from 'https://cdn.skypack.dev/fflate@0.8.2?min';
15</script>

If you are using Deno:

1// Don't use the ?dts Skypack flag; it isn't necessary for Deno support
2// The @deno-types comment adds TypeScript typings
3
4// @deno-types="https://cdn.skypack.dev/fflate@0.8.2/lib/index.d.ts"
5import * as fflate from 'https://cdn.skypack.dev/fflate@0.8.2?min';

If your environment doesn't support bundling:

1// Again, try to import just what you need
2
3// For the browser:
4import * as fflate from 'fflate/esm/browser.js';
5// If the standard ESM import fails on Node (i.e. older version):
6import * as fflate from 'fflate/esm';

And use:

1// This is an ArrayBuffer of data
2const massiveFileBuf = await fetch('/aMassiveFile').then(
3  res => res.arrayBuffer()
4);
5// To use fflate, you need a Uint8Array
6const massiveFile = new Uint8Array(massiveFileBuf);
7// Note that Node.js Buffers work just fine as well:
8// const massiveFile = require('fs').readFileSync('aMassiveFile.txt');
9
10// Higher level means lower performance but better compression
11// The level ranges from 0 (no compression) to 9 (max compression)
12// The default level is 6
13const notSoMassive = fflate.zlibSync(massiveFile, { level: 9 });
14const massiveAgain = fflate.unzlibSync(notSoMassive);
15const gzipped = fflate.gzipSync(massiveFile, {
16  // GZIP-specific: the filename to use when decompressed
17  filename: 'aMassiveFile.txt',
18  // GZIP-specific: the modification time. Can be a Date, date string,
19  // or Unix timestamp
20  mtime: '9/1/16 2:00 PM'
21});

fflate can autodetect a compressed file's format as well:

1const compressed = new Uint8Array(
2  await fetch('/GZIPorZLIBorDEFLATE').then(res => res.arrayBuffer())
3);
4// Above example with Node.js Buffers:
5// Buffer.from('H4sIAAAAAAAAE8tIzcnJBwCGphA2BQAAAA==', 'base64');
6
7const decompressed = fflate.decompressSync(compressed);

Using strings is easy with fflate's string conversion API:

1const buf = fflate.strToU8('Hello world!');
2
3// The default compression method is gzip
4// Increasing mem may increase performance at the cost of memory
5// The mem ranges from 0 to 12, where 4 is the default
6const compressed = fflate.compressSync(buf, { level: 6, mem: 8 });
7
8// When you need to decompress:
9const decompressed = fflate.decompressSync(compressed);
10const origText = fflate.strFromU8(decompressed);
11console.log(origText); // Hello world!

If you need to use an (albeit inefficient) binary string, you can set the second argument to true.

1const buf = fflate.strToU8('Hello world!');
2
3// The second argument, latin1, is a boolean that indicates that the data
4// is not Unicode but rather should be encoded and decoded as Latin-1.
5// This is useful for creating a string from binary data that isn't
6// necessarily valid UTF-8. However, binary strings are incredibly
7// inefficient and tend to double file size, so they're not recommended.
8const compressedString = fflate.strFromU8(
9  fflate.compressSync(buf),
10  true
11);
12const decompressed = fflate.decompressSync(
13  fflate.strToU8(compressedString, true)
14);
15const origText = fflate.strFromU8(decompressed);
16console.log(origText); // Hello world!

You can use streams as well to incrementally add data to be compressed or decompressed:

1// This example uses synchronous streams, but for the best experience
2// you'll definitely want to use asynchronous streams.
3
4let outStr = '';
5const gzipStream = new fflate.Gzip({ level: 9 }, (chunk, isLast) => {
6  // accumulate in an inefficient binary string (just an example)
7  outStr += fflate.strFromU8(chunk, true);
8});
9
10// You can also attach the data handler separately if you don't want to
11// do so in the constructor.
12gzipStream.ondata = (chunk, final) => { ... }
13
14// Since this is synchronous, all errors will be thrown by stream.push()
15gzipStream.push(chunk1);
16gzipStream.push(chunk2);
17
18...
19
20// You should mark the last chunk by using true in the second argument
21// In addition to being necessary for the stream to work properly, this
22// will also set the isLast parameter in the handler to true.
23gzipStream.push(lastChunk, true);
24
25console.log(outStr); // The compressed binary string is now available
26
27// The options parameter for compression streams is optional; you can
28// provide one parameter (the handler) or none at all if you set
29// deflateStream.ondata later.
30const deflateStream = new fflate.Deflate((chunk, final) => {
31  console.log(chunk, final);
32});
33
34// If you want to create a stream from strings, use EncodeUTF8
35const utfEncode = new fflate.EncodeUTF8((data, final) => {
36  // Chaining streams together is done by pushing to the
37  // next stream in the handler for the previous stream
38  deflateStream.push(data, final);
39});
40
41utfEncode.push('Hello'.repeat(1000));
42utfEncode.push(' '.repeat(100));
43utfEncode.push('world!'.repeat(10), true);
44
45// The deflateStream has logged the compressed data
46
47const inflateStream = new fflate.Inflate();
48inflateStream.ondata = (decompressedChunk, final) => { ... };
49
50let stringData = '';
51
52// Streaming UTF-8 decode is available too
53const utfDecode = new fflate.DecodeUTF8((data, final) => {
54  stringData += data;
55});
56
57// Decompress streams auto-detect the compression method, as the
58// non-streaming decompress() method does.
59const dcmpStrm = new fflate.Decompress((chunk, final) => {
60  console.log(chunk, 'was encoded with GZIP, Zlib, or DEFLATE');
61  utfDecode.push(chunk, final);
62});
63
64dcmpStrm.push(zlibJSONData1);
65dcmpStrm.push(zlibJSONData2, true);
66
67// This succeeds; the UTF-8 decoder chained with the unknown compression format
68// stream to reach a string as a sink.
69console.log(JSON.parse(stringData));

You can create multi-file ZIP archives easily as well. Note that by default, compression is enabled for all files, which is not useful when ZIPping many PNGs, JPEGs, PDFs, etc. because those formats are already compressed. You should either override the level on a per-file basis or globally to avoid wasting resources.

1// Note that the asynchronous version (see below) runs in parallel and
2// is *much* (up to 3x) faster for larger archives.
3const zipped = fflate.zipSync({
4  // Directories can be nested structures, as in an actual filesystem
5  'dir1': {
6    'nested': {
7      // You can use Unicode in filenames
8      '你好.txt': fflate.strToU8('Hey there!')
9    },
10    // You can also manually write out a directory path
11    'other/tmp.txt': new Uint8Array([97, 98, 99, 100])
12  },
13
14  // You can also provide compression options
15  'massiveImage.bmp': [aMassiveFile, {
16    level: 9,
17    mem: 12
18  }],
19  // PNG is pre-compressed; no need to waste time
20  'superTinyFile.png': [aPNGFile, { level: 0 }],
21
22  // Directories take options too
23  'exec': [{
24    'hello.sh': [fflate.strToU8('echo hello world'), {
25      // ZIP only: Set the operating system to Unix
26      os: 3,
27      // ZIP only: Make this file executable on Unix
28      attrs: 0o755 << 16
29    }]
30  }, {
31    // ZIP and GZIP support mtime (defaults to current time)
32    mtime: new Date('10/20/2020')
33  }]
34}, {
35  // These options are the defaults for all files, but file-specific
36  // options take precedence.
37  level: 1,
38  // Obfuscate last modified time by default 
39  mtime: new Date('1/1/1980')
40});
41
42// If you write the zipped data to myzip.zip and unzip, the folder
43// structure will be outputted as:
44
45// myzip.zip (original file)
46// dir1
47// |-> nested
48// |   |-> 你好.txt
49// |-> other
50// |   |-> tmp.txt
51// massiveImage.bmp
52// superTinyFile.png
53
54// When decompressing, folders are not nested; all filepaths are fully
55// written out in the keys. For example, the return value may be:
56// { 'nested/directory/structure.txt': Uint8Array(2) [97, 97] }
57const decompressed = fflate.unzipSync(zipped, {
58  // You may optionally supply a filter for files. By default, all files in a
59  // ZIP archive are extracted, but a filter can save resources by telling
60  // the library not to decompress certain files
61  filter(file) {
62    // Don't decompress the massive image or any files larger than 10 MiB
63    return file.name != 'massiveImage.bmp' && file.originalSize <= 10_000_000;
64  }
65});

If you need extremely high performance or custom ZIP compression formats, you can use the highly-extensible ZIP streams. They take streams as both input and output. You can even use custom compression/decompression algorithms from other libraries, as long as they are defined in the ZIP spec (see section 4.4.5). If you'd like more info on using custom compressors, feel free to ask.

1// ZIP object
2// Can also specify zip.ondata outside of the constructor
3const zip = new fflate.Zip((err, dat, final) => {
4  if (!err) {
5    // output of the streams
6    console.log(dat, final);
7  }
8});
9
10const helloTxt = new fflate.ZipDeflate('hello.txt', {
11  level: 9
12});
13
14// Always add streams to ZIP archives before pushing to those streams
15zip.add(helloTxt);
16
17helloTxt.push(chunk1);
18// Last chunk
19helloTxt.push(chunk2, true);
20
21// ZipPassThrough is like ZipDeflate with level 0, but allows for tree shaking
22const nonStreamingFile = new fflate.ZipPassThrough('test.png');
23zip.add(nonStreamingFile);
24// If you have data already loaded, just .push(data, true)
25nonStreamingFile.push(pngData, true);
26
27// You need to call .end() after finishing
28// This ensures the ZIP is valid
29zip.end();
30
31// Unzip object
32const unzipper = new fflate.Unzip();
33
34// This function will almost always have to be called. It is used to support
35// compression algorithms such as BZIP2 or LZMA in ZIP files if just DEFLATE
36// is not enough (though it almost always is).
37// If your ZIP files are not compressed, this line is not needed.
38unzipper.register(fflate.UnzipInflate);
39
40const neededFiles = ['file1.txt', 'example.json'];
41
42// Can specify handler in constructor too
43unzipper.onfile = file => {
44  // file.name is a string, file is a stream
45  if (neededFiles.includes(file.name)) {
46    file.ondata = (err, dat, final) => {
47      // Stream output here
48      console.log(dat, final);
49    };
50    
51    console.log('Reading:', file.name);
52
53    // File sizes are sometimes not set if the ZIP file did not encode
54    // them, so you may want to check that file.size != undefined
55    console.log('Compressed size', file.size);
56    console.log('Decompressed size', file.originalSize);
57
58    // You should only start the stream if you plan to use it to improve
59    // performance. Only after starting the stream will ondata be called.
60    // This method will throw if the compression method hasn't been registered
61    file.start();
62  }
63};
64
65// Try to keep under 5,000 files per chunk to avoid stack limit errors
66// For example, if all files are a few kB, multi-megabyte chunks are OK
67// If files are mostly under 100 bytes, 64kB chunks are the limit
68unzipper.push(zipChunk1);
69unzipper.push(zipChunk2);
70unzipper.push(zipChunk3, true);

As you may have guessed, there is an asynchronous version of every method as well. Unlike most libraries, this will cause the compression or decompression run in a separate thread entirely and automatically by using Web (or Node) Workers. This means that the processing will not block the main thread at all.

Note that there is a significant initial overhead to using workers of about 50ms for each asynchronous function. For instance, if you call unzip ten times, the overhead only applies for the first call, but if you call unzip and zlib, they will each cause the 50ms delay. For small (under about 50kB) payloads, the asynchronous APIs will be much slower. However, if you're compressing larger files/multiple files at once, or if the synchronous API causes the main thread to hang for too long, the callback APIs are an order of magnitude better.

1import {
2  gzip, zlib, AsyncGzip, zip, unzip, strFromU8,
3  Zip, AsyncZipDeflate, Unzip, AsyncUnzipInflate
4} from 'fflate';
5
6// Workers will work in almost any browser (even IE11!)
7// All of the async APIs use a node-style callback as so:
8const terminate = gzip(aMassiveFile, (err, data) => {
9  if (err) {
10    // The compressed data was likely corrupt, so we have to handle
11    // the error.
12    return;
13  }
14  // Use data however you like
15  console.log(data.length);
16});
17
18if (needToCancel) {
19  // The return value of any of the asynchronous APIs is a function that,
20  // when called, will immediately cancel the operation. The callback
21  // will not be called.
22  terminate();
23}
24
25// If you wish to provide options, use the second argument.
26
27// The consume option will render the data inside aMassiveFile unusable,
28// but can improve performance and dramatically reduce memory usage.
29zlib(aMassiveFile, { consume: true, level: 9 }, (err, data) => {
30  // Use the data
31});
32
33// Asynchronous streams are similar to synchronous streams, but the
34// handler has the error that occurred (if any) as the first parameter,
35// and they don't block the main thread.
36
37// Additionally, any buffers that are pushed in will be consumed and
38// rendered unusable; if you need to use a buffer you push in, you
39// should clone it first.
40const gzs = new AsyncGzip({ level: 9, mem: 12, filename: 'hello.txt' });
41let wasCallbackCalled = false;
42gzs.ondata = (err, chunk, final) => {
43  // Note the new err parameter
44  if (err) {
45    // Note that after this occurs, the stream becomes corrupt and must
46    // be discarded. You can't continue pushing chunks and expect it to
47    // work.
48    console.error(err);
49    return;
50  }
51  wasCallbackCalled = true;
52}
53gzs.push(chunk);
54
55// Since the stream is asynchronous, the callback will not be called
56// immediately. If such behavior is absolutely necessary (it shouldn't
57// be), use synchronous streams.
58console.log(wasCallbackCalled) // false
59
60// To terminate an asynchronous stream's internal worker, call
61// stream.terminate().
62gzs.terminate();
63
64// This is way faster than zipSync because the compression of multiple
65// files runs in parallel. In fact, the fact that it's parallelized
66// makes it faster than most standalone ZIP CLIs. The effect is most
67// significant for multiple large files; less so for many small ones.
68zip({ f1: aMassiveFile, 'f2.txt': anotherMassiveFile }, {
69  // The options object is still optional, you can still do just
70  // zip(archive, callback)
71  level: 6
72}, (err, data) => {
73  // Save the ZIP file
74});
75
76// unzip is the only async function without support for consume option
77// It is parallelized, so unzip is also often much faster than unzipSync
78unzip(aMassiveZIPFile, (err, unzipped) => {
79  // If the archive has data.xml, log it here
80  console.log(unzipped['data.xml']);
81  // Conversion to string
82  console.log(strFromU8(unzipped['data.xml']))
83});
84
85// Streaming ZIP archives can accept asynchronous streams. This automatically
86// uses multicore compression.
87const zip = new Zip();
88zip.ondata = (err, chunk, final) => { ... };
89// The JSON and BMP are compressed in parallel
90const exampleFile = new AsyncZipDeflate('example.json');
91zip.add(exampleFile);
92exampleFile.push(JSON.stringify({ large: 'object' }), true);
93const exampleFile2 = new AsyncZipDeflate('example2.bmp', { level: 9 });
94zip.add(exampleFile2);
95exampleFile2.push(ec2a);
96exampleFile2.push(ec2b);
97exampleFile2.push(ec2c);
98...
99exampleFile2.push(ec2Final, true);
100zip.end();
101
102// Streaming Unzip should register the asynchronous inflation algorithm
103// for parallel processing.
104const unzip = new Unzip(stream => {
105  if (stream.name.endsWith('.json')) {
106    stream.ondata = (err, chunk, final) => { ... };
107    stream.start();
108
109    if (needToCancel) {
110      // To cancel these streams, call .terminate()
111      stream.terminate();
112    }
113  }
114});
115unzip.register(AsyncUnzipInflate);
116unzip.push(data, true);

See the documentation for more detailed information about the API.

Bundle size estimates

The bundle size measurements for fflate on sites like Bundlephobia include every feature of the library and should be seen as an upper bound. As long as you are using tree shaking or dead code elimination, this table should give you a general idea of fflate's bundle size for the features you need.

The maximum bundle size that is possible with fflate is about 31kB (11.5kB gzipped) if you use every single feature, but feature parity with pako is only around 10kB (as opposed to 45kB from pako). If your bundle size increases dramatically after adding fflate, please create an issue.

Feature	Bundle size (minified)	Nearest competitor
Decompression	3kB	tiny-inflate
Compression	5kB	UZIP.js, 2.84x larger
Async decompression	4kB (1kB + raw decompression)	N/A
Async compression	6kB (1kB + raw compression)	N/A
ZIP decompression	5kB (2kB + raw decompression)	UZIP.js, 2.84x larger
ZIP compression	7kB (2kB + raw compression)	UZIP.js, 2.03x larger
GZIP/Zlib decompression	4kB (1kB + raw decompression)	pako, 11.4x larger
GZIP/Zlib compression	5kB (1kB + raw compression)	pako, 9.12x larger
Streaming decompression	4kB (1kB + raw decompression)	pako, 11.4x larger
Streaming compression	5kB (1kB + raw compression)	pako, 9.12x larger

What makes fflate so fast?

Many JavaScript compression/decompression libraries exist. However, the most popular one, pako, is merely a clone of Zlib rewritten nearly line-for-line in JavaScript. Although it is by no means poorly made, pako doesn't recognize the many differences between JavaScript and C, and therefore is suboptimal for performance. Moreover, even when minified, the library is 45 kB; it may not seem like much, but for anyone concerned with optimizing bundle size (especially library authors), it's more weight than necessary.

Note that there exist some small libraries like tiny-inflate for solely decompression, and with a minified size of 3 kB, it can be appealing; however, its performance is lackluster, typically 40% worse than pako in my tests.

UZIP.js is both faster (by up to 40%) and smaller (14 kB minified) than pako, and it contains a variety of innovations that make it excellent for both performance and compression ratio. However, the developer made a variety of tiny mistakes and inefficient design choices that make it imperfect. Moreover, it does not support GZIP or Zlib data directly; one must remove the headers manually to use UZIP.js.

So what makes fflate different? It takes the brilliant innovations of UZIP.js and optimizes them while adding direct support for GZIP and Zlib data. And unlike all of the above libraries, it uses ES Modules to allow for partial builds through tree shaking, meaning that it can rival even tiny-inflate in size while maintaining excellent performance. The end result is a library that, in total, weighs 8kB minified for the core build (3kB for decompression only and 5kB for compression only), is about 15% faster than UZIP.js or up to 60% faster than pako, and achieves the same or better compression ratio than the rest.

Before you decide that fflate is the end-all compression library, you should note that JavaScript simply cannot rival the performance of a native program. If you're only using Node.js, it's probably better to use the native Zlib bindings, which tend to offer the best performance. Though note that even against Zlib, fflate is only around 30% slower in decompression and 10% slower in compression, and can still achieve better compression ratios!

What about CompressionStream?

Like fflate, the Compression Streams API provides DEFLATE, GZIP, and Zlib compression and decompression support. It's a good option if you'd like to compress or decompress data without installing any third-party libraries, and it wraps native Zlib bindings to achieve better performance than what most JavaScript programs can achieve.

However, browsers do not offer any native non-streaming compression API, and CompressionStream has surprisingly poor performance on data already loaded into memory; fflate tends to be faster even for files that are dozens of megabytes large. Similarly, fflate is much faster for files under a megabyte because it avoids marshalling overheads. Even when streaming hundreds of megabytes of data, the native API usually performs between 30% faster and 10% slower than fflate. And Compression Streams have many other disadvantages - no ability to control compression level, poor support for older browsers, no ZIP support, etc.

If you'd still prefer to depend upon a native browser API but want to support older browsers, you can use an fflate-based Compression Streams ponyfill.

Browser support

fflate makes heavy use of typed arrays (Uint8Array, Uint16Array, etc.). Typed arrays can be polyfilled at the cost of performance, but the most recent browser that doesn't support them is from 2011, so I wouldn't bother.

The asynchronous APIs also use Worker, which is not supported in a few browsers (however, the vast majority of browsers that support typed arrays support Worker).

Other than that, fflate is completely ES3, meaning you probably won't even need a bundler to use it.

Testing

You can validate the performance of fflate with npm test. It validates that the module is working as expected, ensures the outputs are no more than 5% larger than competitors at max compression, and outputs performance metrics to test/results.

Note that the time it takes for the CLI to show the completion of each test is not representative of the time each package took, so please check the JSON output if you want accurate measurements.

License

This software is MIT Licensed, with special exemptions for projects and organizations as noted below:

• SheetJS is exempt from MIT licensing and may license any source code from this software under the BSD Zero Clause License