snappy/framing_format.txt

Snappy framing format description
Last revised: 2013-10-25

This format decribes a framing format for Snappy, allowing compressing to
files or streams that can then more easily be decompressed without having
to hold the entire stream in memory. It also provides data checksums to
help verify integrity. It does not provide metadata checksums, so it does
not protect against e.g. all forms of truncations.

Implementation of the framing format is optional for Snappy compressors and
decompressor; it is not part of the Snappy core specification.


1. General structure

The file consists solely of chunks, lying back-to-back with no padding
in between. Each chunk consists first a single byte of chunk identifier,
then a three-byte little-endian length of the chunk in bytes (from 0 to
16777215, inclusive), and then the data if any. The four bytes of chunk
header is not counted in the data length.

The different chunk types are listed below. The first chunk must always
be the stream identifier chunk (see section 4.1, below). The stream
ends when the file ends -- there is no explicit end-of-file marker.


2. File type identification

The following identifiers for this format are recommended where appropriate.
However, note that none have been registered officially, so this is only to
be taken as a guideline. We use "Snappy framed" to distinguish between this
format and raw Snappy data.

  File extension:         .sz
  MIME type:              application/x-snappy-framed
  HTTP Content-Encoding:  x-snappy-framed


3. Checksum format

Some chunks have data protected by a checksum (the ones that do will say so
explicitly). The checksums are always masked CRC-32Cs.

A description of CRC-32C can be found in RFC 3720, section 12.1, with
examples in section B.4.

Checksums are not stored directly, but masked, as checksumming data and
then its own checksum can be problematic. The masking is the same as used
in Apache Hadoop: Rotate the checksum by 15 bits, then add the constant
0xa282ead8 (using wraparound as normal for unsigned integers). This is
equivalent to the following C code:

  uint32_t mask_checksum(uint32_t x) {
    return ((x >> 15) | (x << 17)) + 0xa282ead8;
  }

Note that the masking is reversible.

The checksum is always stored as a four bytes long integer, in little-endian.


4. Chunk types

The currently supported chunk types are described below. The list may
be extended in the future.


4.1. Stream identifier (chunk type 0xff)

The stream identifier is always the first element in the stream.
It is exactly six bytes long and contains "sNaPpY" in ASCII. This means that
a valid Snappy framed stream always starts with the bytes

  0xff 0x06 0x00 0x00 0x73 0x4e 0x61 0x50 0x70 0x59

The stream identifier chunk can come multiple times in the stream besides
the first; if such a chunk shows up, it should simply be ignored, assuming
it has the right length and contents. This allows for easy concatenation of
compressed files without the need for re-framing.


4.2. Compressed data (chunk type 0x00)

Compressed data chunks contain a normal Snappy compressed bitstream;
see the compressed format specification. The compressed data is preceded by
the CRC-32C (see section 3) of the _uncompressed_ data.

Note that the data portion of the chunk, i.e., the compressed contents,
can be at most 16777211 bytes (2^24 - 1, minus the checksum).
However, we place an additional restriction that the uncompressed data
in a chunk must be no longer than 65536 bytes. This allows consumers to
easily use small fixed-size buffers.


4.3. Uncompressed data (chunk type 0x01)

Uncompressed data chunks allow a compressor to send uncompressed,
raw data; this is useful if, for instance, uncompressible or
near-incompressible data is detected, and faster decompression is desired.

As in the compressed chunks, the data is preceded by its own masked
CRC-32C (see section 3).

An uncompressed data chunk, like compressed data chunks, should contain
no more than 65536 data bytes, so the maximum legal chunk length with the
checksum is 65540.


4.4. Padding (chunk type 0xfe)

Padding chunks allow a compressor to increase the size of the data stream
so that it complies with external demands, e.g. that the total number of
bytes is a multiple of some value.

All bytes of the padding chunk, except the chunk byte itself and the length,
should be zero, but decompressors must not try to interpret or verify the
padding data in any way.


4.5. Reserved unskippable chunks (chunk types 0x02-0x7f)

These are reserved for future expansion. A decoder that sees such a chunk
should immediately return an error, as it must assume it cannot decode the
stream correctly.

Future versions of this specification may define meanings for these chunks.


4.6. Reserved skippable chunks (chunk types 0x80-0xfd)

These are also reserved for future expansion, but unlike the chunks
described in 4.5, a decoder seeing these must skip them and continue
decoding.

Future versions of this specification may define meanings for these chunks.