Python 2.X CFFI bindings for libRaptorQ v0.1.x - C++11 implementation of RaptorQ Forward Error Correction codes, as described in RFC6330.
Warning: Using libRaptorQ RFC6330 API (which this module wraps around) properly requires knowledge of some concepts and parameters described in that RFC, and not using correct ones may result in undecodable data! See "Usage" section below for more details.
Warning: As far as I know (not a lawyer), there are lots of patents around the use of this technology, which might be important for any high-profile and commercial projects, especially in US and Canada.
Quoting wikipedia on Raptor code:
Raptor codes, as with fountain codes in general, encode a given message consisting of a number of symbols, k, into a potentially limitless sequence of encoding symbols such that knowledge of any k or more encoding symbols allows the message to be recovered with some non-zero probability.
Raptor ("RAPid TORnado") codes are the first known class of fountain codes with linear time encoding and decoding.
And RFC6330:
RaptorQ codes are a new family of codes that provide superior flexibility, support for larger source block sizes, and better coding efficiency than Raptor codes in RFC 5053.
... in most cases, a set of cardinality equal to the number of source symbols is sufficient; in rare cases, a set of cardinality slightly more than the number of source symbols is required.
In practice this means that source data block of size 1 MiB (for example) can (with very high probability) be recovered from any 1.002 MiB of the received symbols for it (from "Application Layer Forward Error Correction for Mobile Multimedia Broadcasting Case Study" paper).
Note that being a probablilistic algorithm, RaptorQ can have highly-improbable pathological cases and be exploited through these e.g. by dropping specific data blocks (see "Stopping a Rapid Tornado with a Puff" paper for more details).
Encoded data will be roughly same size as original plus the "repair symbols", i.e. almost no size overhead, except for what is intentionally generated.
Module includes command-line script ("rq", when installed or as symlink in the repo), which has example code for both encoding and decoding, and can be used as a standalone tool, or for basic algorithm testing/showcase.
Can also be used from command-line via python2 -m libraptorq ... invocation
(when installed as module), e.g. python2 -m libraptorq --help.
Important: With current 0.1.x libRaptorQ API, specifying unsuitable parameters for encoding, such as having symbol_size=16 and max_memory=200 for encoding 200K+ of data WILL result in silently producing encoded data that cannot be decoded.
Note: it's just an example/testing script to run and check if module works with specific parameters or see how to use it, don't rely on it as a production tool or anything like that.
To encode file, with 50% extra symbols (resulting indivisible data chunks to be stored/transmitted intact or lost entirely) and 30% of total from these (K required symbols + X repair symbols) dropped (for testing purposes) before saving them to "setup.py.enc":
% ./rq --debug encode -s16 -m200 --repair-symbols-rate 0.5 --drop-rate 0.3 setup.py setup.py.enc Initialized RQEncoder (0.063s)... Precomputed blocks (0.002s)... Finished encoding symbols (9 blocks, 0.008s)... Closed RQEncoder (0.002s)... Encoded 1,721 B into 167 symbols (needed: >108, repair rate: 50%), 45 dropped (30%), 122 left in output (1,952 B without ids)
Decode original file back from these:
% ./rq --debug decode setup.py.enc setup.py.dec
Initialized RQDecoder (0.064s)...
Decoded enough symbols to recover data (0.010s)...
Closed RQDecoder (0.002s)...
Decoded 1,721 B of data from 108 processed symbols (1,728 B without ids, symbols total: 122)
% sha256sum -b setup.py{,.dec}
36c50348459b51821a2715b0f5c4ef08647d66f77a29913121af4f0f4dfef454 *setup.py
36c50348459b51821a2715b0f5c4ef08647d66f77a29913121af4f0f4dfef454 *setup.py.dec
No matter which chunks are dropped (get picked by random.choice), file
should be recoverable from output as long as number of chunks left (in each
"block") is slightly (by ~0.02%) above K.
Output data ("setup.py.enc" in the example) for the script is JSON-encoded list
of base64-encoded symbols, as well as some parameters for lib init
(oti_scheme, oti_common).
Input data length and sha256 hash of source data are only there to make sure that decoded data is same as original (or exit with error otherwise).
See output with --help option for all the other script parameters.
To use as a python2 module:
from libraptorq import RQEncoder
data = 'some input string' * 500
# Data size must be divisible by RQEncoder.data_size_div
data_len, n = len(data), RQEncoder.data_size_div
if data_len % n: data += '\0' * (n - data_len % n)
with RQEncoder(data, min_subsymbol_size=4, symbol_size=16, max_memory=200) as enc:
symbols = dict()
oti_scheme, oti_common = enc.oti_scheme, enc.oti_common
for block in enc:
symbols.update(block.encode_iter(repair_rate=0))
data_encoded = data_len, oti_scheme, oti_common, symbols
oti_scheme and oti_common are two integers specifying encoder options,
needed to initialize decoder, which can be hard-coded (if constant) on both ends.
block.encode_iter() can be used without options to produce max possible
amount of symbols, up to block.symbols + block.max_repair.
Above example only produces K symbols - min amount required.
For decoding (reverse operation):
from libraptorq import RQDecoder data_len, oti_scheme, oti_common, symbols = data_encoded with RQDecoder(oti_common, oti_scheme) as dec: for sym_id, sym in symbols.viewitems(): dec.add_symbol(sym, sym_id) data = dec.decode()[:data_len]
Note that in practice, e.g. when transmitting each symbol in a udp packet, one'd
want to send something like sym_id || sym_data || checksum, and keep sending
these from block.encode_iter() until other side acknowledges that it can
decode a block (i.e. enough symbols received, see RQDecoder.decode_block()),
then start streaming the next block in similar fashion.
See __main__.py file (cli script) for an extended example, and libRaptorQ docs for info on its C API, which this module wraps around.
It's a regular package for Python 2.7 (not 3.X).
It uses and needs CFFI (can/should be installed by pip) and libRaptorQ v0.1.x installed (as libRaptorQ.so) on the system.
libRaptorQ v1.x (as opposed to current stable version 0.1.9) has different API and will not work with this module.
Using pip is the best way:
% pip install libraptorq
If you don't have it, use:
% easy_install pip % pip install libraptorq
Alternatively (see also pip2014.com and pip install guide):
% curl https://raw.github.com/pypa/pip/master/contrib/get-pip.py | python2 % pip install libraptorq
Or, if you absolutely must:
% easy_install libraptorq
But, you really shouldn't do that.
Current-git version can be installed like this:
% pip install 'git+https://github.com/mk-fg/python-libraptorq.git#egg=libraptorq'
Note that to install stuff in system-wide PATH and site-packages, elevated privileges are often required. Use "install --user", ~/.pydistutils.cfg or virtualenv to do unprivileged installs into custom paths.
Alternatively, ./rq tool can be run right from the checkout tree without any
installation, if that's the only thing you need there.
- See github-issue-1 for more info on what happens when encoding parameters (such as symbol_size and max_memory) are specified carelessly, and why command-line interface of this module does not have defaults for these.
- libRaptorQ is currently used via CFFI in "ABI Mode" to avoid any extra hassle with compilation and the need for compiler, see CFFI docs on the subject for more info on what it means.
- When decoding, libRaptorQ can raise errors for
add_symbol()calls, when source block is already decoded and that extra symbol is not needed. - libRaptorQ allows to specify "rq_type" parameter for internal data alignment size (C++ iterator element), which is hard-coded to RQ_ENC_32/RQ_DEC_32 in the module, for simplicity.
- Lack of Python 3.X compatibility is due to me not using it at all (yet?), so don't need it, have nothing against it in principle.