Snappy compressed format description
Last revised: 2011-10-05
This is not a formal specification, but should suffice to explain most
relevant parts of how the Snappy format works. It is originally based on
text by Zeev Tarantov.
Snappy is a LZ77-type compressor with a fixed, byte-oriented encoding.
There is no entropy encoder backend nor framing layer -- the latter is
assumed to be handled by other parts of the system.
This document only describes the format, not how the Snappy compressor nor
decompressor actually works. The correctness of the decompressor should not
depend on implementation details of the compressor, and vice versa.
The stream starts with the uncompressed length (up to a maximum of 2^32 - 1),
stored as a little-endian varint. Varints consist of a series of bytes,
where the lower 7 bits are data and the upper bit is set iff there are
more bytes to be read. In other words, an uncompressed length of 64 would
be stored as 0x40, and an uncompressed length of 2097150 (0x1FFFFE)
would be stored as 0xFE 0xFF 0x7F.
2. The compressed stream itself
There are two types of elements in a Snappy stream: Literals and
copies (backreferences). There is no restriction on the order of elements,
except that the stream naturally cannot start with a copy. (Having
two literals in a row is never optimal from a compression point of
view, but nevertheless fully permitted.) Each element starts with a tag byte,
and the lower two bits of this tag byte signal what type of element will
01: Copy with 1-byte offset
10: Copy with 2-byte offset
11: Copy with 4-byte offset
The interpretation of the upper six bits are element-dependent.
2.1. Literals (00)
Literals are uncompressed data stored directly in the byte stream.
The literal length is stored differently depending on the length
of the literal:
- For literals up to and including 60 bytes in length, the upper
six bits of the tag byte contain (len-1). The literal follows
immediately thereafter in the bytestream.
- For longer literals, the (len-1) value is stored after the tag byte,
little-endian. The upper six bits of the tag byte describe how
many bytes are used for the length; 60, 61, 62 or 63 for
1-4 bytes, respectively. The literal itself follows after the
Copies are references back into previous decompressed data, telling
the decompressor to reuse data it has previously decoded.
They encode two values: The _offset_, saying how many bytes back
from the current position to read, and the _length_, how many bytes
to copy. Offsets of zero can be encoded, but are not legal;
similarly, it is possible to encode backreferences that would
go past the end of the block (offset > current decompressed position),
which is also nonsensical and thus not allowed.
As in most LZ77-based compressors, the length can be larger than the offset,
yielding a form of run-length encoding (RLE). For instance,
"xababab" could be encoded as
<literal: "xab"> <copy: offset=2 length=4>
Note that since the current Snappy compressor works in 32 kB
blocks and does not do matching across blocks, it will never produce
a bitstream with offsets larger than about 32768. However, the
decompressor should not rely on this, as it may change in the future.
There are several different kinds of copy elements, depending on
the amount of bytes to be copied (length), and how far back the
data to be copied is (offset).
2.2.1. Copy with 1-byte offset (01)
These elements can encode lengths between [4..11] bytes and offsets
between [0..2047] bytes. (len-4) occupies three bits and is stored
in bits [2..4] of the tag byte. The offset occupies 11 bits, of which the
upper three are stored in the upper three bits ([5..7]) of the tag byte,
and the lower eight are stored in a byte following the tag byte.
2.2.2. Copy with 2-byte offset (10)
These elements can encode lengths between [1..64] and offsets from
[0..65535]. (len-1) occupies six bits and is stored in the upper
six bits ([2..7]) of the tag byte. The offset is stored as a
little-endian 16-bit integer in the two bytes following the tag byte.
2.2.3. Copy with 4-byte offset (11)
These are like the copies with 2-byte offsets (see previous subsection),
except that the offset is stored as a 32-bit integer instead of a
16-bit integer (and thus will occupy four bytes).