21 KiB
zstd
Zstandard is a real-time compression algorithm, providing high compression ratios. It offers a very wide range of compression / speed trade-off, while being backed by a very fast decoder. A high performance compression algorithm is implemented. For now focused on speed.
This package provides compression to and decompression of Zstandard content.
This package is pure Go and without use of "unsafe".
The zstd package is provided as open source software using a Go standard license.
Currently the package is heavily optimized for 64 bit processors and will be significantly slower on 32 bit processors.
Installation
Install using go get -u github.com/klauspost/compress. The package is located in github.com/klauspost/compress/zstd.
Godoc Documentation: https://godoc.org/github.com/klauspost/compress/zstd
Compressor
Status:
STABLE - there may always be subtle bugs, a wide variety of content has been tested and the library is actively used by several projects. This library is being fuzz-tested for all updates.
There may still be specific combinations of data types/size/settings that could lead to edge cases, so as always, testing is recommended.
For now, a high speed (fastest) and medium-fast (default) compressor has been implemented.
- The "Fastest" compression ratio is roughly equivalent to zstd level 1.
- The "Default" compression ratio is roughly equivalent to zstd level 3 (default).
- The "Better" compression ratio is roughly equivalent to zstd level 7.
- The "Best" compression ratio is roughly equivalent to zstd level 11.
In terms of speed, it is typically 2x as fast as the stdlib deflate/gzip in its fastest mode. The compression ratio compared to stdlib is around level 3, but usually 3x as fast.
Usage
An Encoder can be used for either compressing a stream via the
io.WriteCloser interface supported by the Encoder or as multiple independent
tasks via the EncodeAll function.
Smaller encodes are encouraged to use the EncodeAll function.
Use NewWriter to create a new instance that can be used for both.
To create a writer with default options, do like this:
// Compress input to output.
func Compress(in io.Reader, out io.Writer) error {
enc, err := zstd.NewWriter(out)
if err != nil {
return err
}
_, err = io.Copy(enc, in)
if err != nil {
enc.Close()
return err
}
return enc.Close()
}
Now you can encode by writing data to enc. The output will be finished writing when Close() is called.
Even if your encode fails, you should still call Close() to release any resources that may be held up.
The above is fine for big encodes. However, whenever possible try to reuse the writer.
To reuse the encoder, you can use the Reset(io.Writer) function to change to another output.
This will allow the encoder to reuse all resources and avoid wasteful allocations.
Currently stream encoding has 'light' concurrency, meaning up to 2 goroutines can be working on part
of a stream. This is independent of the WithEncoderConcurrency(n), but that is likely to change
in the future. So if you want to limit concurrency for future updates, specify the concurrency
you would like.
You can specify your desired compression level using WithEncoderLevel() option. Currently only pre-defined
compression settings can be specified.
Future Compatibility Guarantees
This will be an evolving project. When using this package it is important to note that both the compression efficiency and speed may change.
The goal will be to keep the default efficiency at the default zstd (level 3). However the encoding should never be assumed to remain the same, and you should not use hashes of compressed output for similarity checks.
The Encoder can be assumed to produce the same output from the exact same code version. However, the may be modes in the future that break this, although they will not be enabled without an explicit option.
This encoder is not designed to (and will probably never) output the exact same bitstream as the reference encoder.
Also note, that the cgo decompressor currently does not report all errors on invalid input, omits error checks, ignores checksums and seems to ignore concatenated streams, even though it is part of the spec.
Blocks
For compressing small blocks, the returned encoder has a function called EncodeAll(src, dst []byte) []byte.
EncodeAll will encode all input in src and append it to dst.
This function can be called concurrently, but each call will only run on a single goroutine.
Encoded blocks can be concatenated and the result will be the combined input stream.
Data compressed with EncodeAll can be decoded with the Decoder, using either a stream or DecodeAll.
Especially when encoding blocks you should take special care to reuse the encoder. This will effectively make it run without allocations after a warmup period. To make it run completely without allocations, supply a destination buffer with space for all content.
import "github.com/klauspost/compress/zstd"
// Create a writer that caches compressors.
// For this operation type we supply a nil Reader.
var encoder, _ = zstd.NewWriter(nil)
// Compress a buffer.
// If you have a destination buffer, the allocation in the call can also be eliminated.
func Compress(src []byte) []byte {
return encoder.EncodeAll(src, make([]byte, 0, len(src)))
}
You can control the maximum number of concurrent encodes using the WithEncoderConcurrency(n)
option when creating the writer.
Using the Encoder for both a stream and individual blocks concurrently is safe.
Performance
I have collected some speed examples to compare speed and compression against other compressors.
fileis the input file.outis the compressor used.zskpis this package.zstdis the Datadog cgo library.gzstd/gzkpis gzip standard and this library.levelis the compression level used. Forzskplevel 1 is "fastest", level 2 is "default"; 3 is "better", 4 is "best".insize/outsizeis the input/output size.millisis the number of milliseconds used for compression.mb/sis megabytes (2^20 bytes) per second.
Silesia Corpus:
http://sun.aei.polsl.pl/~sdeor/corpus/silesia.zip
This package:
file out level insize outsize millis mb/s
silesia.tar zskp 1 211947520 73101992 643 313.87
silesia.tar zskp 2 211947520 67504318 969 208.38
silesia.tar zskp 3 211947520 65177448 1899 106.44
silesia.tar zskp 4 211947520 61381950 8115 24.91
cgo zstd:
silesia.tar zstd 1 211947520 73605392 543 371.56
silesia.tar zstd 3 211947520 66793289 864 233.68
silesia.tar zstd 6 211947520 62916450 1913 105.66
silesia.tar zstd 9 211947520 60212393 5063 39.92
gzip, stdlib/this package:
silesia.tar gzstd 1 211947520 80007735 1654 122.21
silesia.tar gzkp 1 211947520 80369488 1168 173.06
GOB stream of binary data. Highly compressible.
https://files.klauspost.com/compress/gob-stream.7z
file out level insize outsize millis mb/s
gob-stream zskp 1 1911399616 235022249 3088 590.30
gob-stream zskp 2 1911399616 205669791 3786 481.34
gob-stream zskp 3 1911399616 185792019 9324 195.48
gob-stream zskp 4 1911399616 171537212 32113 56.76
gob-stream zstd 1 1911399616 249810424 2637 691.26
gob-stream zstd 3 1911399616 208192146 3490 522.31
gob-stream zstd 6 1911399616 193632038 6687 272.56
gob-stream zstd 9 1911399616 177620386 16175 112.70
gob-stream gzstd 1 1911399616 357382641 10251 177.82
gob-stream gzkp 1 1911399616 362156523 5695 320.08
The test data for the Large Text Compression Benchmark is the first
10^9 bytes of the English Wikipedia dump on Mar. 3, 2006.
http://mattmahoney.net/dc/textdata.html
file out level insize outsize millis mb/s
enwik9 zskp 1 1000000000 343848582 3609 264.18
enwik9 zskp 2 1000000000 317276632 5746 165.97
enwik9 zskp 3 1000000000 294540704 11725 81.34
enwik9 zskp 4 1000000000 276609671 44029 21.66
enwik9 zstd 1 1000000000 358072021 3110 306.65
enwik9 zstd 3 1000000000 313734672 4784 199.35
enwik9 zstd 6 1000000000 295138875 10290 92.68
enwik9 zstd 9 1000000000 278348700 28549 33.40
enwik9 gzstd 1 1000000000 382578136 9604 99.30
enwik9 gzkp 1 1000000000 383825945 6544 145.73
Highly compressible JSON file.
https://files.klauspost.com/compress/github-june-2days-2019.json.zst
file out level insize outsize millis mb/s
github-june-2days-2019.json zskp 1 6273951764 699045015 10620 563.40
github-june-2days-2019.json zskp 2 6273951764 617881763 11687 511.96
github-june-2days-2019.json zskp 3 6273951764 537511906 29252 204.54
github-june-2days-2019.json zskp 4 6273951764 512796117 97791 61.18
github-june-2days-2019.json zstd 1 6273951764 766284037 8450 708.00
github-june-2days-2019.json zstd 3 6273951764 661889476 10927 547.57
github-june-2days-2019.json zstd 6 6273951764 642756859 22996 260.18
github-june-2days-2019.json zstd 9 6273951764 601974523 52413 114.16
github-june-2days-2019.json gzstd 1 6273951764 1164400847 29948 199.79
github-june-2days-2019.json gzkp 1 6273951764 1128755542 19236 311.03
VM Image, Linux mint with a few installed applications:
https://files.klauspost.com/compress/rawstudio-mint14.7z
file out level insize outsize millis mb/s
rawstudio-mint14.tar zskp 1 8558382592 3667489370 20210 403.84
rawstudio-mint14.tar zskp 2 8558382592 3364592300 31873 256.07
rawstudio-mint14.tar zskp 3 8558382592 3224594213 71751 113.75
rawstudio-mint14.tar zskp 4 8558382592 3027332295 486243 16.79
rawstudio-mint14.tar zstd 1 8558382592 3609250104 17136 476.27
rawstudio-mint14.tar zstd 3 8558382592 3341679997 29262 278.92
rawstudio-mint14.tar zstd 6 8558382592 3235846406 77904 104.77
rawstudio-mint14.tar zstd 9 8558382592 3160778861 140946 57.91
rawstudio-mint14.tar gzstd 1 8558382592 3926257486 57722 141.40
rawstudio-mint14.tar gzkp 1 8558382592 3970463184 41749 195.49
CSV data:
https://files.klauspost.com/compress/nyc-taxi-data-10M.csv.zst
file out level insize outsize millis mb/s
nyc-taxi-data-10M.csv zskp 1 3325605752 641339945 8925 355.35
nyc-taxi-data-10M.csv zskp 2 3325605752 591748091 11268 281.44
nyc-taxi-data-10M.csv zskp 3 3325605752 538490114 19880 159.53
nyc-taxi-data-10M.csv zskp 4 3325605752 495986829 89368 35.49
nyc-taxi-data-10M.csv zstd 1 3325605752 687399637 8233 385.18
nyc-taxi-data-10M.csv zstd 3 3325605752 598514411 10065 315.07
nyc-taxi-data-10M.csv zstd 6 3325605752 570522953 20038 158.27
nyc-taxi-data-10M.csv zstd 9 3325605752 517554797 64565 49.12
nyc-taxi-data-10M.csv gzstd 1 3325605752 928656485 23876 132.83
nyc-taxi-data-10M.csv gzkp 1 3325605752 924718719 16388 193.53
Decompressor
Staus: STABLE - there may still be subtle bugs, but a wide variety of content has been tested.
This library is being continuously fuzz-tested, kindly supplied by fuzzit.dev. The main purpose of the fuzz testing is to ensure that it is not possible to crash the decoder, or run it past its limits with ANY input provided.
Usage
The package has been designed for two main usages, big streams of data and smaller in-memory buffers.
There are two main usages of the package for these. Both of them are accessed by creating a Decoder.
For streaming use a simple setup could look like this:
import "github.com/klauspost/compress/zstd"
func Decompress(in io.Reader, out io.Writer) error {
d, err := zstd.NewReader(in)
if err != nil {
return err
}
defer d.Close()
// Copy content...
_, err = io.Copy(out, d)
return err
}
It is important to use the "Close" function when you no longer need the Reader to stop running goroutines. See "Allocation-less operation" below.
For decoding buffers, it could look something like this:
import "github.com/klauspost/compress/zstd"
// Create a reader that caches decompressors.
// For this operation type we supply a nil Reader.
var decoder, _ = zstd.NewReader(nil)
// Decompress a buffer. We don't supply a destination buffer,
// so it will be allocated by the decoder.
func Decompress(src []byte) ([]byte, error) {
return decoder.DecodeAll(src, nil)
}
Both of these cases should provide the functionality needed.
The decoder can be used for concurrent decompression of multiple buffers.
It will only allow a certain number of concurrent operations to run.
To tweak that yourself use the WithDecoderConcurrency(n) option when creating the decoder.
Dictionaries
Data compressed with dictionaries can be decompressed.
Dictionaries are added individually to Decoders.
Dictionaries are generated by the zstd --train command and contains an initial state for the decoder.
To add a dictionary use the WithDecoderDicts(dicts ...[]byte) option with the dictionary data.
Several dictionaries can be added at once.
The dictionary will be used automatically for the data that specifies them. A re-used Decoder will still contain the dictionaries registered.
When registering multiple dictionaries with the same ID, the last one will be used.
It is possible to use dictionaries when compressing data.
To enable a dictionary use WithEncoderDict(dict []byte). Here only one dictionary will be used
and it will likely be used even if it doesn't improve compression.
The used dictionary must be used to decompress the content.
For any real gains, the dictionary should be built with similar data. If an unsuitable dictionary is used the output may be slightly larger than using no dictionary. Use the zstd commandline tool to build a dictionary from sample data. For information see zstd dictionary information.
For now there is a fixed startup performance penalty for compressing content with dictionaries. This will likely be improved over time. Just be aware to test performance when implementing.
Allocation-less operation
The decoder has been designed to operate without allocations after a warmup.
This means that you should store the decoder for best performance.
To re-use a stream decoder, use the Reset(r io.Reader) error to switch to another stream.
A decoder can safely be re-used even if the previous stream failed.
To release the resources, you must call the Close() function on a decoder.
After this it can no longer be reused, but all running goroutines will be stopped.
So you must use this if you will no longer need the Reader.
For decompressing smaller buffers a single decoder can be used. When decoding buffers, you can supply a destination slice with length 0 and your expected capacity. In this case no unneeded allocations should be made.
Concurrency
The buffer decoder does everything on the same goroutine and does nothing concurrently.
It can however decode several buffers concurrently. Use WithDecoderConcurrency(n) to limit that.
The stream decoder operates on
- One goroutine reads input and splits the input to several block decoders.
- A number of decoders will decode blocks.
- A goroutine coordinates these blocks and sends history from one to the next.
So effectively this also means the decoder will "read ahead" and prepare data to always be available for output.
Since "blocks" are quite dependent on the output of the previous block stream decoding will only have limited concurrency.
In practice this means that concurrency is often limited to utilizing about 2 cores effectively.
Benchmarks
These are some examples of performance compared to datadog cgo library.
The first two are streaming decodes and the last are smaller inputs.
BenchmarkDecoderSilesia-8 3 385000067 ns/op 550.51 MB/s 5498 B/op 8 allocs/op
BenchmarkDecoderSilesiaCgo-8 6 197666567 ns/op 1072.25 MB/s 270672 B/op 8 allocs/op
BenchmarkDecoderEnwik9-8 1 2027001600 ns/op 493.34 MB/s 10496 B/op 18 allocs/op
BenchmarkDecoderEnwik9Cgo-8 2 979499200 ns/op 1020.93 MB/s 270672 B/op 8 allocs/op
Concurrent performance:
BenchmarkDecoder_DecodeAllParallel/kppkn.gtb.zst-16 28915 42469 ns/op 4340.07 MB/s 114 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/geo.protodata.zst-16 116505 9965 ns/op 11900.16 MB/s 16 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/plrabn12.txt.zst-16 8952 134272 ns/op 3588.70 MB/s 915 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/lcet10.txt.zst-16 11820 102538 ns/op 4161.90 MB/s 594 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/asyoulik.txt.zst-16 34782 34184 ns/op 3661.88 MB/s 60 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/alice29.txt.zst-16 27712 43447 ns/op 3500.58 MB/s 99 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/html_x_4.zst-16 62826 18750 ns/op 21845.10 MB/s 104 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/paper-100k.pdf.zst-16 631545 1794 ns/op 57078.74 MB/s 2 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/fireworks.jpeg.zst-16 1690140 712 ns/op 172938.13 MB/s 1 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/urls.10K.zst-16 10432 113593 ns/op 6180.73 MB/s 1143 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/html.zst-16 113206 10671 ns/op 9596.27 MB/s 15 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallel/comp-data.bin.zst-16 1530615 779 ns/op 5229.49 MB/s 0 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/kppkn.gtb.zst-16 65217 16192 ns/op 11383.34 MB/s 46 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/geo.protodata.zst-16 292671 4039 ns/op 29363.19 MB/s 6 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/plrabn12.txt.zst-16 26314 46021 ns/op 10470.43 MB/s 293 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/lcet10.txt.zst-16 33897 34900 ns/op 12227.96 MB/s 205 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/asyoulik.txt.zst-16 104348 11433 ns/op 10949.01 MB/s 20 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/alice29.txt.zst-16 75949 15510 ns/op 9805.60 MB/s 32 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/html_x_4.zst-16 173910 6756 ns/op 60624.29 MB/s 37 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/paper-100k.pdf.zst-16 923076 1339 ns/op 76474.87 MB/s 1 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/fireworks.jpeg.zst-16 922920 1351 ns/op 91102.57 MB/s 2 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/urls.10K.zst-16 27649 43618 ns/op 16096.19 MB/s 407 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/html.zst-16 279073 4160 ns/op 24614.18 MB/s 6 B/op 0 allocs/op
BenchmarkDecoder_DecodeAllParallelCgo/comp-data.bin.zst-16 749938 1579 ns/op 2581.71 MB/s 0 B/op 0 allocs/op
This reflects the performance around May 2020, but this may be out of date.
Contributions
Contributions are always welcome. For new features/fixes, remember to add tests and for performance enhancements include benchmarks.
For sending files for reproducing errors use a service like goobox or similar to share your files.
For general feedback and experience reports, feel free to open an issue or write me on Twitter.
This package includes the excellent github.com/cespare/xxhash package Copyright (c) 2016 Caleb Spare.