S2 Compression

S2 is an extension of Snappy.

S2 is aimed for high throughput, which is why it features concurrent compression for bigger payloads.

Decoding is compatible with Snappy compressed content, but content compressed with S2 cannot be decompressed by Snappy. This means that S2 can seamlessly replace Snappy without converting compressed content.

S2 can produce Snappy compatible output, faster and better than Snappy. If you want full benefit of the changes you should use s2 without Snappy compatibility.

S2 is designed to have high throughput on content that cannot be compressed. This is important, so you don't have to worry about spending CPU cycles on already compressed data.

Benefits over Snappy

• Better compression
• Adjustable compression (3 levels)
• Concurrent stream compression
• Faster decompression, even for Snappy compatible content
• Concurrent Snappy/S2 stream decompression
• Skip forward in compressed stream
• Random seeking with indexes
• Compatible with reading Snappy compressed content
• Smaller block size overhead on incompressible blocks
• Block concatenation
• Block Dictionary support
• Uncompressed stream mode
• Automatic stream size padding
• Snappy compatible block compression

Drawbacks over Snappy

• Not optimized for 32 bit systems
• Streams use slightly more memory due to larger blocks and concurrency (configurable)

Usage

Installation: go get -u github.com/klauspost/compress/s2

Full package documentation:

Compression

func EncodeStream(src io.Reader, dst io.Writer) error {
    enc := s2.NewWriter(dst)
    _, err := io.Copy(enc, src)
    if err != nil {
        enc.Close()
        return err
    }
    // Blocks until compression is done.
    return enc.Close() 
}

You should always call enc.Close(), otherwise you will leak resources and your encode will be incomplete.

For the best throughput, you should attempt to reuse the Writer using the Reset() method.

The Writer in S2 is always buffered, therefore NewBufferedWriter in Snappy can be replaced with NewWriter in S2. It is possible to flush any buffered data using the Flush() method. This will block until all data sent to the encoder has been written to the output.

S2 also supports the io.ReaderFrom interface, which will consume all input from a reader.

As a final method to compress data, if you have a single block of data you would like to have encoded as a stream, a slightly more efficient method is to use the EncodeBuffer method. This will take ownership of the buffer until the stream is closed.

func EncodeStream(src []byte, dst io.Writer) error {
    enc := s2.NewWriter(dst)
    // The encoder owns the buffer until Flush or Close is called.
    err := enc.EncodeBuffer(buf)
    if err != nil {
        enc.Close()
        return err
    }
    // Blocks until compression is done.
    return enc.Close()
}

Each call to EncodeBuffer will result in discrete blocks being created without buffering, so it should only be used a single time per stream. If you need to write several blocks, you should use the regular io.Writer interface.

Decompression

func DecodeStream(src io.Reader, dst io.Writer) error {
    dec := s2.NewReader(src)
    _, err := io.Copy(dst, dec)
    return err
}

Similar to the Writer, a Reader can be reused using the Reset method.

For the best possible throughput, there is a EncodeBuffer(buf []byte) function available. However, it requires that the provided buffer isn't used after it is handed over to S2 and until the stream is flushed or closed.

For smaller data blocks, there is also a non-streaming interface: Encode(), EncodeBetter() and Decode(). Do however note that these functions (similar to Snappy) does not provide validation of data, so data corruption may be undetected. Stream encoding provides CRC checks of data.

It is possible to efficiently skip forward in a compressed stream using the Skip() method. For big skips the decompressor is able to skip blocks without decompressing them.

Single Blocks

Similar to Snappy S2 offers single block compression. Blocks do not offer the same flexibility and safety as streams, but may be preferable for very small payloads, less than 100K.

Using a simple dst := s2.Encode(nil, src) will compress src and return the compressed result. It is possible to provide a destination buffer. If the buffer has a capacity of s2.MaxEncodedLen(len(src)) it will be used. If not a new will be allocated.

Alternatively EncodeBetter/EncodeBest can also be used for better, but slightly slower compression.

Similarly to decompress a block you can use dst, err := s2.Decode(nil, src). Again an optional destination buffer can be supplied. The s2.DecodedLen(src) can be used to get the minimum capacity needed. If that is not satisfied a new buffer will be allocated.

Block function always operate on a single goroutine since it should only be used for small payloads.

Commandline tools

Some very simply commandline tools are provided; s2c for compression and s2d for decompression.

Binaries can be downloaded on the Releases Page.

Installing then requires Go to be installed. To install them, use:

go install github.com/klauspost/compress/s2/cmd/s2c@latest && go install github.com/klauspost/compress/s2/cmd/s2d@latest

To build binaries to the current folder use:

go build github.com/klauspost/compress/s2/cmd/s2c && go build github.com/klauspost/compress/s2/cmd/s2d

s2c

Usage: s2c [options] file1 file2

Compresses all files supplied as input separately.
Output files are written as 'filename.ext.s2' or 'filename.ext.snappy'.
By default output files will be overwritten.
Use - as the only file name to read from stdin and write to stdout.

Wildcards are accepted: testdir/*.txt will compress all files in testdir ending with .txt
Directories can be wildcards as well. testdir/*/*.txt will match testdir/subdir/b.txt

File names beginning with 'http://' and 'https://' will be downloaded and compressed.
Only http response code 200 is accepted.

Options:
  -bench int
    	Run benchmark n times. No output will be written
  -blocksize string
    	Max  block size. Examples: 64K, 256K, 1M, 4M. Must be power of two and <= 4MB (default "4M")
  -c	Write all output to stdout. Multiple input files will be concatenated
  -cpu int
    	Compress using this amount of threads (default 32)
  -faster
    	Compress faster, but with a minor compression loss
  -help
    	Display help
  -index
        Add seek index (default true)    	
  -o string
        Write output to another file. Single input file only
  -pad string
    	Pad size to a multiple of this value, Examples: 500, 64K, 256K, 1M, 4M, etc (default "1")
  -q	Don't write any output to terminal, except errors
  -rm
    	Delete source file(s) after successful compression
  -safe
    	Do not overwrite output files
  -slower
    	Compress more, but a lot slower
  -snappy
        Generate Snappy compatible output stream
  -verify
    	Verify written files  

s2d

Usage: s2d [options] file1 file2

Decompresses all files supplied as input. Input files must end with '.s2' or '.snappy'.
Output file names have the extension removed. By default output files will be overwritten.
Use - as the only file name to read from stdin and write to stdout.

Wildcards are accepted: testdir/*.txt will compress all files in testdir ending with .txt
Directories can be wildcards as well. testdir/*/*.txt will match testdir/subdir/b.txt

File names beginning with 'http://' and 'https://' will be downloaded and decompressed.
Extensions on downloaded files are ignored. Only http response code 200 is accepted.

Options:
  -bench int
    	Run benchmark n times. No output will be written
  -c	Write all output to stdout. Multiple input files will be concatenated
  -help
    	Display help
  -o string
        Write output to another file. Single input file only
  -offset string
        Start at offset. Examples: 92, 64K, 256K, 1M, 4M. Requires Index
  -q    Don't write any output to terminal, except errors
  -rm
        Delete source file(s) after successful decompression
  -safe
        Do not overwrite output files
  -tail string
        Return last of compressed file. Examples: 92, 64K, 256K, 1M, 4M. Requires Index
  -verify
    	Verify files, but do not write output                                      

s2sx: self-extracting archives

s2sx allows creating self-extracting archives with no dependencies.

By default, executables are created for the same platforms as the host os, but this can be overridden with -os and -arch parameters.

Extracted files have 0666 permissions, except when untar option used.

Usage: s2sx [options] file1 file2

Compresses all files supplied as input separately.
If files have '.s2' extension they are assumed to be compressed already.
Output files are written as 'filename.s2sx' and with '.exe' for windows targets.
If output is big, an additional file with ".more" is written. This must be included as well.
By default output files will be overwritten.

Wildcards are accepted: testdir/*.txt will compress all files in testdir ending with .txt
Directories can be wildcards as well. testdir/*/*.txt will match testdir/subdir/b.txt

Options:
  -arch string
        Destination architecture (default "amd64")
  -c    Write all output to stdout. Multiple input files will be concatenated
  -cpu int
        Compress using this amount of threads (default 32)
  -help
        Display help
  -max string
        Maximum executable size. Rest will be written to another file. (default "1G")
  -os string
        Destination operating system (default "windows")
  -q    Don't write any output to terminal, except errors
  -rm
        Delete source file(s) after successful compression
  -safe
        Do not overwrite output files
  -untar
        Untar on destination

Available platforms are:

• darwin-amd64
• darwin-arm64
• linux-amd64
• linux-arm
• linux-arm64
• linux-mips64
• linux-ppc64le
• windows-386
• windows-amd64

By default, there is a size limit of 1GB for the output executable.

When this is exceeded the remaining file content is written to a file called output+.more. This file must be included for a successful extraction and placed alongside the executable for a successful extraction.

This file must have the same name as the executable, so if the executable is renamed, so must the .more file.

This functionality is disabled with stdin/stdout.

Self-extracting TAR files

If you wrap a TAR file you can specify -untar to make it untar on the destination host.

Files are extracted to the current folder with the path specified in the tar file.

Note that tar files are not validated before they are wrapped.

For security reasons files that move below the root folder are not allowed.

Performance

This section will focus on comparisons to Snappy. This package is solely aimed at replacing Snappy as a high speed compression package. If you are mainly looking for better compression zstandard gives better compression, but typically at speeds slightly below “better” mode in this package.

Compression is increased compared to Snappy, mostly around 5-20% and the throughput is typically 25-40% increased (single threaded) compared to the Snappy Go implementation.

Streams are concurrently compressed. The stream will be distributed among all available CPU cores for the best possible throughput.

A “better” compression mode is also available. This allows to trade a bit of speed for a minor compression gain. The content compressed in this mode is fully compatible with the standard decoder.

Snappy vs S2 compression speed on 16 core (32 thread) computer, using all threads and a single thread (1 CPU):

Legend

• S2 Speed: Speed of S2 compared to Snappy, using 16 cores and 1 core.
• S2 Throughput: Throughput of S2 in MB/s.
• S2 % smaller: How many percent of the Snappy output size is S2 better.
• S2 "better": Speed when enabling “better” compression mode in S2 compared to Snappy.
• "better" throughput: Speed when enabling “better” compression mode in S2 compared to Snappy.
• "better" % smaller: How many percent of the Snappy output size is S2 better when using “better” compression.

There is a good speedup across the board when using a single thread and a significant speedup when using multiple threads.

Machine generated data gets by far the biggest compression boost, with size being reduced by up to 35% of Snappy size.

The “better” compression mode sees a good improvement in all cases, but usually at a performance cost.

Incompressible content (sharnd.out.2gb, 2GB random data) sees the smallest speedup. This is likely dominated by synchronization overhead, which is confirmed by the fact that single threaded performance is higher (see above).

Decompression

S2 attempts to create content that is also fast to decompress, except in “better” mode where the smallest representation is used.

S2 vs Snappy decompression speed. Both operating on single core:

Legend

• S2 Throughput: Decompression speed of S2 encoded content.
• Better Throughput: Decompression speed of S2 “better” encoded content.
• vs Snappy: Decompression speed of S2 “better” mode compared to Snappy and absolute speed.

While the decompression code hasn't changed, there is a significant speedup in decompression speed. S2 prefers longer matches and will typically only find matches that are 6 bytes or longer. While this reduces compression a bit, it improves decompression speed.

The “better” compression mode will actively look for shorter matches, which is why it has a decompression speed quite similar to Snappy.

Without assembly decompression is also very fast; single goroutine decompression speed. No assembly:

Even though S2 typically compresses better than Snappy, decompression speed is always better.

Concurrent Stream Decompression

For full stream decompression S2 offers a DecodeConcurrent that will decode a full stream using multiple goroutines.

Example scaling, AMD Ryzen 3950X, 16 cores, decompression using s2d -bench=3 <input>, best of 3:

Scaling can be expected to be pretty linear until memory bandwidth is saturated.

For now the DecodeConcurrent can only be used for full streams without seeking or combining with regular reads.

Block compression

When compressing blocks no concurrent compression is performed just as Snappy. This is because blocks are for smaller payloads and generally will not benefit from concurrent compression.

An important change is that incompressible blocks will not be more than at most 10 bytes bigger than the input. In rare, worst case scenario Snappy blocks could be significantly bigger than the input.

Mixed content blocks

The most reliable is a wide dataset. For this we use webdevdata.org-2015-01-07-subset, 53927 files, total input size: 4,014,735,833 bytes. Single goroutine used.

S2 delivers both the best single threaded throughput with regular mode and the best compression rate with “best”. “Better” mode provides the same compression speed as LZ4 with better compression ratio.

When outputting Snappy compatible output it still delivers better throughput (150MB/s more) and better compression.

As can be seen from the other benchmarks decompression should also be easier on the S2 generated output.

Though they cannot be compared due to different decompression speeds here are the speed/size comparisons for other Go compressors:

Standard block compression

Benchmarking single block performance is subject to a lot more variation since it only tests a limited number of file patterns. So individual benchmarks should only be seen as a guideline and the overall picture is more important.

These micro-benchmarks are with data in cache and trained branch predictors. For a more realistic benchmark see the mixed content above.

Block compression. Parallel benchmark running on 16 cores, 16 goroutines.

AMD64 assembly is use for both S2 and Snappy.

Speed is generally at or above Snappy. Small blocks gets a significant speedup, although at the expense of size.

Decompression speed is better than Snappy, except in one case.

Since payloads are very small the variance in terms of size is rather big, so they should only be seen as a general guideline.

Size is on average around Snappy, but varies on content type. In cases where compression is worse, it usually is compensated by a speed boost.

Better compression

Benchmarking single block performance is subject to a lot more variation since it only tests a limited number of file patterns. So individual benchmarks should only be seen as a guideline and the overall picture is more important.

Except for the mostly incompressible JPEG image compression is better and usually in the double digits in terms of percentage reduction over Snappy.

The PDF sample shows a significant slowdown compared to Snappy, as this mode tries harder to compress the data. Very small blocks are also not favorable for better compression, so throughput is way down.

This mode aims to provide better compression at the expense of performance and achieves that without a huge performance penalty, except on very small blocks.

Decompression speed suffers a little compared to the regular S2 mode, but still manages to be close to Snappy in spite of increased compression.

Best compression mode

S2 offers a “best” compression mode.

This will compress as much as possible with little regard to CPU usage.

Mainly for offline compression, but where decompression speed should still be high and compatible with other S2 compressed data.

Some examples compared on 16 core CPU, amd64 assembly used:

* enwik10
Default... 10000000000 -> 4759950115 [47.60%]; 1.03s, 9263.0MB/s
Better...  10000000000 -> 4084706676 [40.85%]; 2.16s, 4415.4MB/s
Best...    10000000000 -> 3615520079 [36.16%]; 42.259s, 225.7MB/s

* github-june-2days-2019.json
Default... 6273951764 -> 1041700255 [16.60%]; 431ms, 13882.3MB/s
Better...  6273951764 -> 945841238 [15.08%]; 547ms, 10938.4MB/s
Best...    6273951764 -> 826392576 [13.17%]; 9.455s, 632.8MB/s

* nyc-taxi-data-10M.csv
Default... 3325605752 -> 1093516949 [32.88%]; 324ms, 9788.7MB/s
Better...  3325605752 -> 885394158 [26.62%]; 491ms, 6459.4MB/s
Best...    3325605752 -> 773681257 [23.26%]; 8.29s, 412.0MB/s

* 10gb.tar
Default... 10065157632 -> 5915541066 [58.77%]; 1.028s, 9337.4MB/s
Better...  10065157632 -> 5453844650 [54.19%]; 1.597s, 4862.7MB/s
Best...    10065157632 -> 5192495021 [51.59%]; 32.78s, 308.2MB/

* consensus.db.10gb
Default... 10737418240 -> 4549762344 [42.37%]; 882ms, 12118.4MB/s
Better...  10737418240 -> 4438535064 [41.34%]; 1.533s, 3500.9MB/s
Best...    10737418240 -> 4210602774 [39.21%]; 42.96s, 254.4MB/s

Decompression speed should be around the same as using the ‘better’ compression mode.

Dictionaries

Note: S2 dictionary compression is currently at an early implementation stage, with no assembly for neither encoding nor decoding. Performance improvements can be expected in the future.

Adding dictionaries allow providing a custom dictionary that will serve as lookup in the beginning of blocks.

The same dictionary must be used for both encoding and decoding. S2 does not keep track of whether the same dictionary is used, and using the wrong dictionary will most often not result in an error when decompressing.

Blocks encoded without dictionaries can be decompressed seamlessly with a dictionary. This means it is possible to switch from an encoding without dictionaries to an encoding with dictionaries and treat the blocks similarly.

Similar to zStandard dictionaries, the same usage scenario applies to S2 dictionaries.

Training works if there is some correlation in a family of small data samples. The more data-specific a dictionary is, the more efficient it is (there is no universal dictionary). Hence, deploying one dictionary per type of data will provide the greatest benefits. Dictionary gains are mostly effective in the first few KB. Then, the compression algorithm will gradually use previously decoded content to better compress the rest of the file.

S2 further limits the dictionary to only be enabled on the first 64KB of a block. This will remove any negative (speed) impacts of the dictionaries on bigger blocks.

Compression

Using the github_users_sample_set and a 64KB dictionary trained with zStandard the following sizes can be achieved.

So for highly repetitive content, this case provides an almost 3x reduction in size.

For less uniform data we will use the Go source code tree. Compressing First 64KB of all .go files in go/src, Go 1.19.5, 8912 files, 51253563 bytes input:

Creating Dictionaries

There are no tools to create dictionaries in S2. However, there are multiple ways to create a useful dictionary:

Using a Sample File

If your input is very uniform, you can just use a sample file as the dictionary.

For example in the github_users_sample_set above, the average compression only goes up from 10.49% to 11.48% by using the first file as dictionary compared to using a dedicated dictionary.

    // Read a sample
    sample, err := os.ReadFile("sample.json")

    // Create a dictionary.
    dict := s2.MakeDict(sample, nil)
	
    // b := dict.Bytes() will provide a dictionary that can be saved
    // and reloaded with s2.NewDict(b).
	
    // To encode:
    encoded := dict.Encode(nil, file)

    // To decode:
    decoded, err := dict.Decode(nil, file)

Using Zstandard

Zstandard dictionaries can easily be converted to S2 dictionaries.

This can be helpful to generate dictionaries for files that don't have a fixed structure.

Example, with training set files placed in ./training-set:

λ zstd -r --train-fastcover training-set/* --maxdict=65536 -o name.dict

This will create a dictionary of 64KB, that can be converted to a dictionary like this:

    // Decode the Zstandard dictionary.
    insp, err := zstd.InspectDictionary(zdict)
    if err != nil {
        panic(err)
    }
	
    // We are only interested in the contents.
    // Assume that files start with "// Copyright (c) 2023".
    // Search for the longest match for that.
    // This may save a few bytes.
    dict := s2.MakeDict(insp.Content(), []byte("// Copyright (c) 2023"))

    // b := dict.Bytes() will provide a dictionary that can be saved
    // and reloaded with s2.NewDict(b).

    // We can now encode using this dictionary
    encodedWithDict := dict.Encode(nil, payload)

    // To decode content:
    decoded, err := dict.Decode(nil, encodedWithDict)

It is recommended to save the dictionary returned by b:= dict.Bytes(), since that will contain only the S2 dictionary.

This dictionary can later be loaded using s2.NewDict(b). The dictionary then no longer requires zstd to be initialized.

Also note how s2.MakeDict allows you to search for a common starting sequence of your files. This can be omitted, at the expense of a few bytes.

Snappy Compatibility

S2 now offers full compatibility with Snappy.

This means that the efficient encoders of S2 can be used to generate fully Snappy compatible output.

There is a snappy package that can be used by simply changing imports from github.com/golang/snappy to github.com/klauspost/compress/snappy. This uses “better” mode for all operations. If you would like more control, you can use the s2 package as described below:

Blocks

Snappy compatible blocks can be generated with the S2 encoder. Compression and speed is typically a bit better MaxEncodedLen is also smaller for smaller memory usage. Replace

s2.EncodeSnappy can be replaced with s2.EncodeSnappyBetter or s2.EncodeSnappyBest to get more efficiently compressed snappy compatible output.

s2.ConcatBlocks is compatible with snappy blocks.

Comparison of webdevdata.org-2015-01-07-subset, 53927 files, total input size: 4,014,735,833 bytes. amd64, single goroutine used:

Streams

For streams, replace enc = snappy.NewBufferedWriter(w) with enc = s2.NewWriter(w, s2.WriterSnappyCompat()). All other options are available, but note that block size limit is different for snappy.

Comparison of different streams, AMD Ryzen 3950x, 16 cores. Size and throughput:

Decompression

All decompression functions map directly to equivalent s2 functions.

Features like quick forward skipping without decompression are also available for Snappy streams.

If you know you are only decompressing snappy streams, setting ReaderMaxBlockSize(64<<10) on your Reader will reduce memory consumption.

Concatenating blocks and streams.

Concatenating streams will concatenate the output of both without recompressing them. While this is inefficient in terms of compression it might be usable in certain scenarios. The 10 byte ‘stream identifier’ of the second stream can optionally be stripped, but it is not a requirement.

Blocks can be concatenated using the ConcatBlocks function.

Snappy blocks/streams can safely be concatenated with S2 blocks and streams. Streams with indexes (see below) will currently not work on concatenated streams.

Stream Seek Index

S2 and Snappy streams can have indexes. These indexes will allow random seeking within the compressed data.

The index can either be appended to the stream as a skippable block or returned for separate storage.

When the index is appended to a stream it will be skipped by regular decoders, so the output remains compatible with other decoders.

Creating an Index

To automatically add an index to a stream, add WriterAddIndex() option to your writer. Then the index will be added to the stream when Close() is called.

	// Add Index to stream...
	enc := s2.NewWriter(w, s2.WriterAddIndex())
	io.Copy(enc, r)
	enc.Close()

If you want to store the index separately, you can use CloseIndex() instead of the regular Close(). This will return the index. Note that CloseIndex() should only be called once, and you shouldn't call Close().

	// Get index for separate storage... 
	enc := s2.NewWriter(w)
	io.Copy(enc, r)
	index, err := enc.CloseIndex()

The index can then be used needing to read from the stream. This means the index can be used without needing to seek to the end of the stream or for manually forwarding streams. See below.

Finally, an existing S2/Snappy stream can be indexed using the s2.IndexStream(r io.Reader) function.

Using Indexes

To use indexes there is a ReadSeeker(random bool, index []byte) (*ReadSeeker, error) function available.

Calling ReadSeeker will return an io.ReadSeeker compatible version of the reader.

If ‘random’ is specified the returned io.Seeker can be used for random seeking, otherwise only forward seeking is supported. Enabling random seeking requires the original input to support the io.Seeker interface.

	dec := s2.NewReader(r)
	rs, err := dec.ReadSeeker(false, nil)
	rs.Seek(wantOffset, io.SeekStart)	

Get a seeker to seek forward. Since no index is provided, the index is read from the stream. This requires that an index was added and that r supports the io.Seeker interface.

A custom index can be specified which will be used if supplied. When using a custom index, it will not be read from the input stream.

	dec := s2.NewReader(r)
	rs, err := dec.ReadSeeker(false, index)
	rs.Seek(wantOffset, io.SeekStart)	

This will read the index from index. Since we specify non-random (forward only) seeking r does not have to be an io.Seeker

	dec := s2.NewReader(r)
	rs, err := dec.ReadSeeker(true, index)
	rs.Seek(wantOffset, io.SeekStart)	

Finally, since we specify that we want to do random seeking r must be an io.Seeker.

The returned ReadSeeker contains a shallow reference to the existing Reader, meaning changes performed to one is reflected in the other.

To check if a stream contains an index at the end, the (*Index).LoadStream(rs io.ReadSeeker) error can be used.

Manually Forwarding Streams

Indexes can also be read outside the decoder using the Index type. This can be used for parsing indexes, either separate or in streams.

In some cases it may not be possible to serve a seekable stream. This can for instance be an HTTP stream, where the Range request is sent at the start of the stream.

With a little bit of extra code it is still possible to use indexes to forward to specific offset with a single forward skip.

It is possible to load the index manually like this:

	var index s2.Index
	_, err = index.Load(idxBytes)

This can be used to figure out how much to offset the compressed stream:

	compressedOffset, uncompressedOffset, err := index.Find(wantOffset)

The compressedOffset is the number of bytes that should be skipped from the beginning of the compressed file.

The uncompressedOffset will then be offset of the uncompressed bytes returned when decoding from that position. This will always be <= wantOffset.

When creating a decoder it must be specified that it should not expect a stream identifier at the beginning of the stream. Assuming the io.Reader r has been forwarded to compressedOffset we create the decoder like this:

	dec := s2.NewReader(r, s2.ReaderIgnoreStreamIdentifier())

We are not completely done. We still need to forward the stream the uncompressed bytes we didn't want. This is done using the regular “Skip” function:

	err = dec.Skip(wantOffset - uncompressedOffset)

This will ensure that we are at exactly the offset we want, and reading from dec will start at the requested offset.

Compact storage

For compact storage RemoveIndexHeaders can be used to remove any redundant info from a serialized index. If you remove the header it must be restored before Loading.

This is expected to save 20 bytes. These can be restored using RestoreIndexHeaders. This removes a layer of security, but is the most compact representation. Returns nil if headers contains errors.

Index Format:

Each block is structured as a snappy skippable block, with the chunk ID 0x99.

The block can be read from the front, but contains information so it can be read from the back as well.

Numbers are stored as fixed size little endian values or zigzag encoded base 128 varints, with un-encoded value length of 64 bits, unless other limits are specified.

For regular streams the uncompressed offsets are fully predictable, so HasUncompressedOffsets allows to specify that compressed blocks all have exactly EstBlockSize bytes of uncompressed content.

Entries must be in order, starting with the lowest offset, and there must be no uncompressed offset duplicates.
Entries may point to the start of a skippable block, but it is then not allowed to also have an entry for the next block since that would give an uncompressed offset duplicate.

There is no requirement for all blocks to be represented in the index. In fact there is a maximum of 65536 block entries in an index.

The writer can use any method to reduce the number of entries. An implicit block start at 0,0 can be assumed.

Decoding entries:

// Read Uncompressed entries.
// Each assumes EstBlockSize delta from previous.
for each entry {
    uOff = 0
    if HasUncompressedOffsets == 1 {
        uOff = ReadVarInt // Read value from stream
    }
   
    // Except for the first entry, use previous values.
    if entryNum == 0 {
        entry[entryNum].UncompressedOffset = uOff
        continue
    }
    
    // Uncompressed uses previous offset and adds EstBlockSize
    entry[entryNum].UncompressedOffset = entry[entryNum-1].UncompressedOffset + EstBlockSize + uOff
}


// Guess that the first block will be 50% of uncompressed size.
// Integer truncating division must be used.
CompressGuess := EstBlockSize / 2

// Read Compressed entries.
// Each assumes CompressGuess delta from previous.
// CompressGuess is adjusted for each value.
for each entry {
    cOff = ReadVarInt // Read value from stream
    
    // Except for the first entry, use previous values.
    if entryNum == 0 {
        entry[entryNum].CompressedOffset = cOff
        continue
    }
    
    // Compressed uses previous and our estimate.
    entry[entryNum].CompressedOffset = entry[entryNum-1].CompressedOffset + CompressGuess + cOff
        
     // Adjust compressed offset for next loop, integer truncating division must be used. 
     CompressGuess += cOff/2               
}

To decode from any given uncompressed offset (wantOffset):

• Iterate entries until entry[n].UncompressedOffset > wantOffset.
• Start decoding from entry[n-1].CompressedOffset.
• Discard entry[n-1].UncompressedOffset - wantOffset bytes from the decoded stream.

See using indexes for functions that perform the operations with a simpler interface.

Format Extensions

• Frame Stream identifier changed from sNaPpY to S2sTwO.
• Framed compressed blocks can be up to 4MB (up from 64KB).
• Compressed blocks can have an offset of 0, which indicates to repeat the last seen offset.

Repeat offsets must be encoded as a 2.2.1. Copy with 1-byte offset (01), where the offset is 0.

The length is specified by reading the 3-bit length specified in the tag and decode using this table:

This allows any repeat offset + length to be represented by 2 to 5 bytes. It also allows to emit matches longer than 64 bytes with one copy + one repeat instead of several 64 byte copies.

Lengths are stored as little endian values.

The first copy of a block cannot be a repeat offset and the offset is reset on every block in streams.

Default streaming block size is 1MB.

Dictionary Encoding

Adding dictionaries allow providing a custom dictionary that will serve as lookup in the beginning of blocks.

A dictionary provides an initial repeat value that can be used to point to a common header.

Other than that the dictionary contains values that can be used as back-references.

Often used data should be placed at the end of the dictionary since offsets < 2048 bytes will be smaller.

Format

Dictionary content must at least 16 bytes and less or equal to 64KiB (65536 bytes).

Encoding: [repeat value (uvarint)][dictionary content...]

Before the dictionary content, an unsigned base-128 (uvarint) encoded value specifying the initial repeat offset. This value is an offset into the dictionary content and not a back-reference offset, so setting this to 0 will make the repeat value point to the first value of the dictionary.

The value must be less than the dictionary length-8

Encoding

From the decoder point of view the dictionary content is seen as preceding the encoded content.

[dictionary content][decoded output]

Backreferences to the dictionary are encoded as ordinary backreferences that have an offset before the start of the decoded block.

Matches copying from the dictionary are not allowed to cross from the dictionary into the decoded data. However, if a copy ends at the end of the dictionary the next repeat will point to the start of the decoded buffer, which is allowed.

The first match can be a repeat value, which will use the repeat offset stored in the dictionary.

When 64KB (65536 bytes) has been en/decoded it is no longer allowed to reference the dictionary, neither by a copy nor repeat operations. If the boundary is crossed while copying from the dictionary, the operation should complete, but the next instruction is not allowed to reference the dictionary.

Valid blocks encoded without a dictionary can be decoded with any dictionary. There are no checks whether the supplied dictionary is the correct for a block. Because of this there is no overhead by using a dictionary.

Example

This is the dictionary content. Elements are separated by [].

Dictionary: [0x0a][Yesterday 25 bananas were added to Benjamins brown bag].

Initial repeat offset is set at 10, which is the letter 2.

Encoded [LIT "10"][REPEAT len=10][LIT "hich"][MATCH off=50 len=6][MATCH off=31 len=6][MATCH off=61 len=10]

Decoded: [10][ bananas w][hich][ were ][brown ][were added]

Output: 10 bananas which were brown were added

Streams

For streams each block can use the dictionary.

The dictionary cannot not currently be provided on the stream.

LICENSE

This code is based on the Snappy-Go implementation.

Use of this source code is governed by a BSD-style license that can be found in the LICENSE file.