Bloscpack

Author:	Valentin Haenel
Contact:	[email protected]
List:	http://groups.google.com/group/blosc
Github:	https://github.com/Blosc/bloscpack
PyPi:	https://pypi.python.org/pypi/bloscpack
Conda-Forge:	https://github.com/conda-forge/bloscpack-feedstock
Anaconda:	https://anaconda.org/pypi/bloscpack
Ohloh:	https://www.ohloh.net/p/bloscpack
Version:
Travis CI:
GitHub Actions:
Coveralls:
Python Versions:
License:
And...:

Description

Command line interface to and serialization format for Blosc, a high performance, multi-threaded, blocking and shuffling compressor. Uses python-blosc bindings to interface with Blosc. Also comes with native support for efficiently serializing and deserializing Numpy arrays.

Code of Conduct

The Blosc community has adopted a Code of Conduct that we expect project participants to adhere to. Please read the full text so that you can understand what actions will and will not be tolerated.

Dependencies

• Python 3.6 3.7 3.8 and 3.9
• python-blosc (provides Blosc) and Numpy (as listed in requirements.txt) for running the code
• The Python packages listed in test_requirements.txt for testing and releasing

Stability of File Format

The tool is considered alpha-stage, experimental, research software. It is not unlikely that the internal storage format for the compressed files will change in future. Please do not depend critically on the files generated (unless you know what you are doing) by Bloscpack. See the warranty disclaimer in the licence at the end of this file.

Installation

Disclaimer: There are a myriad ways of installing Python packages (and their dependencies) these days and it is a futile endeavour to explain the procedures in great detail again and again. Below are three methods that are known to work. Depending on the method you choose and the system your are using you may require any or all of: super user privileges, a C++ compiler and/or a virtual environment. If you do run into problems or are unsure, it's best to send an email to the aforementioned mailing list asking for help.

The package is available on PyPi, so you may use pip to install the dependencies and bloscpack itself:

$ pip install bloscpack

The package is also available on anaconda.org via conda-forge. You may use conda to install it:

$ conda install -c conda-forge bloscpack

If you want to install straight from GitHub, use pip's VCS support:

$ pip install git+https://github.com/Blosc/bloscpack

Or, of course, download the source code or clone the repository and then use the standard setup.py:

$ git clone https://github.com/Blosc/bloscpack
$ cd bloscpack
$ python setup.py install

Usage

Bloscpack is accessible from the command line using the blpk executable this has a number of global options and four subcommands: [c | compress], [d | decompress], [a | append] and [i | info] most of which each have their own options.

Help for global options and subcommands:

$ blpk --help
[...]

Help for each one of the subcommands:

$ blpk compress --help
[...]
$ blpk decompress --help
[...]
$ blpk info --help
[...]
$ blpk append --help
[...]

Examples

Basics

Basic compression:

$ blpk compress data.dat

Or:

$ blpk c data.dat

... will compress the file data.dat to data.dat.blp

Basic decompression:

$ blpk decompress data.dat.blp data.dcmp

Or:

$ blpk d data.dat.blp data.dcmp

... will decompress the file data.dat.blp to the file data.dcmp. If you leave out the [<out_file>] argument, Bloscpack will complain that the file data.dat exists already and refuse to overwrite it:

$ blpk decompress data.dat.blp
blpk: error: output file 'data.dat' exists!

If you know what you are doing, you can use the global option [-f | --force] to override the overwrite checks:

$ blpk --force decompress data.dat.blp

Incidentally this works for compression too:

$ blpk compress data.dat
blpk: error: output file 'data.dat.blp' exists!
$ blpk --force compress data.dat

Lastly, if you want a different filename:

$ blpk compress data.dat custom.filename.blp

... will compress the file data.dat to custom.filename.blp

Settings

By default, the number of threads that Blosc uses during compression and decompression is determined by the number of cores detected on your system. You can change this using the [-n | --nthreads] option:

$ blpk --nthreads 1 compress data.dat

Compression with Blosc is controlled with the following options:

• [-t | --typesize] Typesize used by Blosc (default: 8): $ blpk compress --typesize 8 data.dat
• [-l | --level] Compression level (default: 7): $ blpk compress --level 3 data.dat
• [-s | --no-shuffle] Deactivate shuffle: $ blpk compress --no-shuffle data.dat
• [-c | --codec] Use alternative codec: $ blpk compress --codec lz4 data.dat

In addition, there are the following options that control the Bloscpack file:

• [-z | --chunk-size] Desired approximate size of the chunks, where you can use human readable strings like 8M or 128K or max to use the maximum chunk size of apprx. 2GB (default: 1MB): $ blpk compress --chunk-size 128K data.dat or $ blpk c -z max data.dat
• [-k | --checksum <checksum>] Chose which checksum to use. The following values are permissible: None, adler32, crc32, md5, sha1, sha224, sha256, sha384, sha512, (default: adler32). As described in the header format, each compressed chunk can be stored with a checksum, which aids corruption detection on decompression: $ blpk compress --checksum crc32 data.dat
• [-o | --no-offsets] By default, offsets to the individual chunks are stored. These are included to allow for partial decompression in the future. This option disables that feature. Also, a certain number of offsets (default: 10 * 'nchunks') are preallocated to allow for appending data to the file: $ blpk compress --no-offsets data.dat

Info Subcommand

If you just need some info on how the file was compressed [i | info]:

$ blpk info data.dat.blp
blpk: BloscpackHeader:
blpk:     format_version: 3
blpk:     offsets: True
blpk:     metadata: False
blpk:     checksum: 'adler32'
blpk:     typesize: 8
blpk:     chunk_size: 1.0M (1048576B)
blpk:     last_chunk: 900.0K (921600B)
blpk:     nchunks: 1526
blpk:     max_app_chunks: 15260
blpk: 'offsets':
blpk: [134320,459218,735869,986505,1237646,...]
blpk: First chunk blosc header:
blpk: OrderedDict([('version', 2), ('versionlz', 1), ('flags', 1), ('typesize', 8), ('nbytes', 1048576), ('blocksize', 131072), ('ctbytes', 324894)])
blpk: First chunk blosc flags:
blpk: OrderedDict([('byte_shuffle', True), ('pure_memcpy', False), ('bit_shuffle', False), ('split_blocks', False), ('codec', 'blosclz')])

Importantly, the header and flag information are for the first chunk only. Usually this isn't a problem because bloscpack compressed files do tend to have homogeneous settings like codec used, typesize etc... However, there is nothing that will stop you from appending to an existing bloscpack file using different settings. For example, half the file might be compressed using 'blosclz' whereas the other half of the file might be compressed with 'lz4'. In any case, just be aware that the output is to be seen as an indication that is likely to be correct for all chunks but must not be so necessarily.

Adding Metdata

Using the [-m | --metadata] option you can include JSON from a file:

$ cat meta.json
{"dtype": "float64", "shape": [200000000], "container": "numpy"}
$ blpk compress --chunk-size=512M --metadata meta.json data.dat
$ blpk info data.dat.blp
blpk: BloscpackHeader:
blpk:     format_version: 3
blpk:     offsets: True
blpk:     metadata: True
blpk:     checksum: 'adler32'
blpk:     typesize: 8
blpk:     chunk_size: 512.0M (536870912B)
blpk:     last_chunk: 501.88M (526258176B)
blpk:     nchunks: 3
blpk:     max_app_chunks: 30
blpk: 'offsets':
blpk: [922,78074943,140783242,...]
blpk: 'metadata':
blpk: {   u'container': u'numpy', u'dtype': u'float64', u'shape': [200000000]}
blpk: MetadataHeader:
blpk:     magic_format: 'JSON'
blpk:     meta_options: '00000000'
blpk:     meta_checksum: 'adler32'
blpk:     meta_codec: 'zlib'
blpk:     meta_level: 6
blpk:     meta_size: 59.0B (59B)
blpk:     max_meta_size: 590.0B (590B)
blpk:     meta_comp_size: 58.0B (58B)
blpk:     user_codec: ''

It will be printed when decompressing:

$ blpk decompress data.dat.blp
blpk: Metadata is:
blpk: '{u'dtype': u'float64', u'shape': [200000000], u'container': u'numpy'}'

Appending

You can also append data to an existing bloscpack compressed file:

$ blpk append data.dat.blp data.dat

However there are certain limitations on the amount of data can be appended. For example, if there is an offsets section, there must be enough room to store the offsets for the appended chunks. If no offsets exists, you may append as much data as possible given the limitations governed by the maximum number of chunks and the chunk-size. Additionally, there are limitations on the compression options. For example, one cannot change the checksum used. It is however possible to change the compression level, the typesize and the shuffle option for the appended chunks.

Also note that appending is still considered experimental as of v0.5.0.

Verbose and Debug mode

Lastly there are two mutually exclusive options to control how much output is produced.

The first causes basic info to be printed, [-v | --verbose]:

$ blpk --verbose compress --chunk-size 0.5G data.dat
blpk: using 4 threads
blpk: getting ready for compression
blpk: input file is: 'data.dat'
blpk: output file is: 'data.dat.blp'
blpk: input file size: 1.49G (1600000000B)
blpk: nchunks: 3
blpk: chunk_size: 512.0M (536870912B)
blpk: last_chunk_size: 501.88M (526258176B)
blpk: output file size: 198.39M (208028617B)
blpk: compression ratio: 7.691250
blpk: done

... and [-d | --debug] prints a detailed account of what is going on:

$ blpk --debug compress --chunk-size 0.5G data.dat
blpk: command line argument parsing complete
blpk: command line arguments are:
blpk:     force: False
blpk:     verbose: False
blpk:     offsets: True
blpk:     checksum: adler32
blpk:     subcommand: compress
blpk:     out_file: None
blpk:     metadata: None
blpk:     cname: blosclz
blpk:     in_file: data.dat
blpk:     chunk_size: 536870912
blpk:     debug: True
blpk:     shuffle: True
blpk:     typesize: 8
blpk:     clevel: 7
blpk:     nthreads: 4
blpk: using 4 threads
blpk: getting ready for compression
blpk: input file is: 'data.dat'
blpk: output file is: 'data.dat.blp'
blpk: input file size: 1.49G (1600000000B)
blpk: nchunks: 3
blpk: chunk_size: 512.0M (536870912B)
blpk: last_chunk_size: 501.88M (526258176B)
blpk: BloscArgs:
blpk:     typesize: 8
blpk:     clevel: 7
blpk:     shuffle: True
blpk:     cname: 'blosclz'
blpk: BloscpackArgs:
blpk:     offsets: True
blpk:     checksum: 'adler32'
blpk:     max_app_chunks: <function <lambda> at 0x1182de8>
blpk: metadata_args will be silently ignored
blpk: max_app_chunks is a callable
blpk: max_app_chunks was set to: 30
blpk: BloscpackHeader:
blpk:     format_version: 3
blpk:     offsets: True
blpk:     metadata: False
blpk:     checksum: 'adler32'
blpk:     typesize: 8
blpk:     chunk_size: 512.0M (536870912B)
blpk:     last_chunk: 501.88M (526258176B)
blpk:     nchunks: 3
blpk:     max_app_chunks: 30
blpk: raw_bloscpack_header: 'blpk\x03\x01\x01\x08\x00\x00\x00 \x00\x10^\x1f\x03\x00\x00\x00\x00\x00\x00\x00\x1e\x00\x00\x00\x00\x00\x00\x00'
blpk: Handle chunk '0'
blpk: checksum (adler32): '\x1f\xed\x1e\xf4'
blpk: chunk handled, in: 512.0M (536870912B) out: 74.46M (78074017B)
blpk: Handle chunk '1'
blpk: checksum (adler32): ')\x1e\x08\x88'
blpk: chunk handled, in: 512.0M (536870912B) out: 59.8M (62708295B)
blpk: Handle chunk '2' (last)
blpk: checksum (adler32): '\xe8\x18\xa4\xac'
blpk: chunk handled, in: 501.88M (526258176B) out: 64.13M (67245997B)
blpk: Writing '3' offsets: '[296, 78074317, 140782616]'
blpk: Raw offsets: '(\x01\x00\x00\x00\x00\x00\x00\xcdQ\xa7\x04\x00\x00\x00\x00\x18,d\x08\x00\x00\x00\x00'
blpk: output file size: 198.39M (208028617B)
blpk: compression ratio: 7.691250
blpk: done

Python API

Bloscpack has a versatile yet simple API consisting of a series of 'arguments' objects and high-level functions that can be invoked dependding on your input and output needs.

Nomenclature wise, Python 3 has done a lot for Bloscpack, because we always need to represent compressed data as bytes deliberatey. This makes it easier and more natural to distinguish between text, such a filenames and binary and bytes objects such as compressed data.

Arguments

The three argument types are:

• BloscArgs
• BloscpackArgs
• MetadataArgs

as defined in bloscpack/args.py. Instantiating any of them will create an object with the defaults setup. The defaults are defined in bloscpack/defaults.py. You can use these in the high-level functions listed below.

You can override any and all defaults by passing in the respective keyword-arguments, for example:

>>> b = BloscArgs()               # will create a default args object
>>> b = BloscArgs(clevel=4)       # change compression level to 4
>>> b = BloscArgs(typesize=4,     # change the typesize to 4
>>> ...           clevel=9,       # change the compression level to 9
>>> ...           shuffle=False,  # deactivate the shuffle filter
>>> ...           cname='lz4')    # let lz4 be the internal codec

class BloscArgs(MutableMappingObject):
    """ Object to hold Blosc arguments.

    Parameters
    ----------
    typesize : int
        The typesize used
    clevel : int
        Compression level
    shuffle : boolean
        Whether or not to activate the shuffle filter
    cname: str
        Name of the internal code to use

    """

class BloscpackArgs(MutableMappingObject):
    """ Object to hold BloscPack arguments.

    Parameters
    ----------
    offsets : boolean
        Whether to include space for offsets
    checksum : str
        Name of the checksum to use or None/'None'
    max_app_chunks : int or callable on number of chunks
        How much space to reserve in the offsets for chunks to be appended.

    """

class MetadataArgs(MutableMappingObject):
    """ Object to hold the metadata arguments.

    Parameters
    ----------
    magic_format : 8 bytes
        Format identifier for the metadata
    meta_checksum : str
        Checksum to be used for the metadata
    meta_codec : str
        Codec to be used to compress the metadata
    meta_level : int
        Compression level for metadata
    max_meta_size : int or callable on metadata size
        How much space to reserve for additional metadata

    """

File / Bytes

The following high-level functions exist for compressing and decompressing to and from files and byte objects:

• pack_file_to_file
• unpack_file_from_file
• pack_bytes_to_file
• unpack_bytes_from_file
• pack_bytes_to_bytes
• unpack_bytes_from_bytes

Beyond the target arguments such as the files and the bytes, each pack_* function takes the following arguments:

chunk_size : int
    the desired chunk size in bytes
metadata : dict
    the metadata dict
blosc_args : BloscArgs
    blosc args
bloscpack_args : BloscpackArgs
    bloscpack args
metadata_args : MetadataArgs
    metadata args

Below are their sigantures:

def pack_file_to_file(in_file, out_file,
                      chunk_size=DEFAULT_CHUNK_SIZE,
                      metadata=None,
                      blosc_args=None,
                      bloscpack_args=None,
                      metadata_args=None):

def unpack_file_from_file(in_file, out_file):


def pack_bytes_to_file(bytes_, out_file,
                       chunk_size=DEFAULT_CHUNK_SIZE,
                       metadata=None,
                       blosc_args=None,
                       bloscpack_args=None,
                       metadata_args=None):

def unpack_bytes_from_file(compressed_file):

def pack_bytes_to_bytes(bytes_,
                        chunk_size=DEFAULT_CHUNK_SIZE,
                        metadata=None,
                        blosc_args=None,
                        bloscpack_args=None,
                        metadata_args=None,
                        ):


def unpack_bytes_from_bytes(bytes_):

Numpy

Numpy arrays can be serialized as Bloscpack files, here is a very brief example:

>>> a = np.linspace(0, 1, 3e8)
>>> print a.size, a.dtype
300000000 float64
>>> bp.pack_ndarray_to_file(a, 'a.blp')
>>> b = bp.unpack_ndarray_from_file('a.blp')
>>> (a == b).all()
True

Looking at the generated file, we can see the Numpy metadata being saved:

$ lh a.blp
-rw------- 1 esc esc 266M Aug 13 23:21 a.blp

$ blpk info a.blp
blpk: BloscpackHeader:
blpk:     format_version: 3
blpk:     offsets: True
blpk:     metadata: True
blpk:     checksum: 'adler32'
blpk:     typesize: 8
blpk:     chunk_size: 1.0M (1048576B)
blpk:     last_chunk: 838.0K (858112B)
blpk:     nchunks: 2289
blpk:     max_app_chunks: 22890
blpk: 'offsets':
blpk: [202170,408064,554912,690452,819679,...]
blpk: 'metadata':
blpk: {   u'container': u'numpy',
blpk:     u'dtype': u'<f8',
blpk:     u'order': u'C',
blpk:     u'shape': [300000000]}
blpk: MetadataHeader:
blpk:     magic_format: 'JSON'
blpk:     meta_options: '00000000'
blpk:     meta_checksum: 'adler32'
blpk:     meta_codec: 'zlib'
blpk:     meta_level: 6
blpk:     meta_size: 67.0B (67B)
blpk:     max_meta_size: 670.0B (670B)
blpk:     meta_comp_size: 62.0B (62B)
blpk:     user_codec: ''

Alternatively, we can also use a string as storage:

>>> a = np.linspace(0, 1, 3e8)
>>> c = pack_ndarray_to_bytes(a)
>>> b = unpack_ndarray_from_bytes(c)
>>> (a == b).all()
True

Or use alternate compressors:

>>> a = np.linspace(0, 1, 3e8)
>>> c = pack_ndarray_to_bytes(a, blosc_args=BloscArgs(cname='lz4'))
>>> b = unpack_ndarray_from_bytes(c)
>>> (a == b).all()
True

def pack_ndarray_to_file(ndarray, filename,
                         chunk_size=DEFAULT_CHUNK_SIZE,
                         blosc_args=None,
                         bloscpack_args=None,
                         metadata_args=None):

def pack_ndarray_to_bytes(ndarray,
                          chunk_size=DEFAULT_CHUNK_SIZE,
                          blosc_args=None,
                          bloscpack_args=None,
                          metadata_args=None):

def unpack_ndarray_from_file(filename):

def unpack_ndarray_from_bytes(str_):

If you are interested in the performance of Bloscpack compared to other serialization formats for Numpy arrays, please look at the benchmarks presented in the Bloscpack paper from the EuroScipy 2013 conference proceedings.

Testing

Installing Dependencies

Testing requires some additional libraries, which you can install from PyPi with:

$ pip install -r test_requirements.txt
[...]

Basic Tests

Basic tests, runs quickly:

$ PYTHONPATH=. pytest -m "not slow"
[...]

Heavier Tests

Extended tests using a larger file, may take some time, but will be nice to memory:

$ PYTHONPATH=. pytest test/test_file_io.py::test_pack_unpack_hard
[...]

Extended tests using a huge file. This one take forever and needs loads (5G-6G) of memory and loads of disk-space (10G). Use -s to print progress:

$ PYTHONPATH=. test/test_file_io.py::test_pack_unpack_extreme
[...]

Note that, some compression/decompression tests create temporary files (on UNIXoid systems this is under /tmp/blpk*) which are deleted upon completion of the respective test, both successful and unsuccessful, or when the test is aborted with e.g. ctrl-c (using atexit magic).

Under rare circumstances, for example when aborting the deletion which is triggered on abort you may be left with large files polluting your temporary space. Depending on your partitioning scheme etc.. doing this repeatedly, may lead to you running out of space on the file-system.

Command Line Interface Tests

The command line interface is tested with cram:

$ cram --verbose test_cmdline/*.cram
[...]

Coverage

To determine coverage you can pool together the coverage from the cram tests and the unit tests:

$ COVERAGE=1 cram --verbose test_cmdline/*.cram
[...]
$ PYTHONPATH=. pytest --cov=bloscpack --cov-append -m "not slow" test
[...]

Test Runner

To run the command line interface tests and the unit tests and analyse coverage, use the convenience test.sh runner:

$ ./test.sh
[...]

Benchmark

Using the provided bench/blpk_vs_gzip.py script on a Intel(R) Core(TM) i7-3667U CPU @ 2.00GHz CPU with 2 cores and 4 threads (active hyperthreading), cpu frequency scaling activated but set to the performance governor (all cores scaled to 2.0 GHz), 8GB of DDR3 memory and a Luks encrypted SSD, we get:

$ PYTHONPATH=. ./bench/blpk_vs_gzip.py
create the test data..........done

Input file size: 1.49G
Will now run bloscpack...
Time: 2.06 seconds
Output file size: 198.55M
Ratio: 7.69
Will now run gzip...
Time: 134.20 seconds
Output file size: 924.05M
Ratio: 1.65

As was expected from previous benchmarks of Blosc using the python-blosc bindings, Blosc is both much faster and has a better compression ratio for this kind of structured data. One thing to note here, is that we are not dropping the system file cache after every step, so the file to read will be cached in memory. To get a more accurate picture we can use the --drop-caches switch of the benchmark which requires you however, to run the benchmark as root, since dropping the caches requires root privileges:

$ PYTHONPATH=. ./bench/blpk_vs_gzip.py --drop-caches
will drop caches
create the test data..........done

Input file size: 1.49G
Will now run bloscpack...
Time: 13.49 seconds
Output file size: 198.55M
Ratio: 7.69
Will now run gzip...
Time: 137.49 seconds
Output file size: 924.05M
Ratio: 1.65

Optimizing Chunk Size

You can use the provided bench/compression_time_vs_chunk_size.py file to optimize the chunk-size for a given machine. For example:

$ sudo env PATH=$PATH PYTHONPATH=.  bench/compression_time_vs_chunk_size.py
create the test data..........done
chunk_size    comp-time       decomp-time      ratio
512.0K        8.106235        10.243908        7.679094
724.08K       4.424007        12.284307        7.092846
1.0M          6.243544        11.978932        7.685173
1.41M         4.715511        10.780901        7.596981
2.0M          4.548568        10.676304        7.688216
2.83M         4.851359        11.668394        7.572480
4.0M          4.557665        10.127647        7.689736
5.66M         4.589349        9.579627         7.667467
8.0M          5.290080        10.525652        7.690499

Running the script requires super user privileges, since you need to synchronize disk writes and drop the file system caches for less noisy results. Also, you should probably run this script a couple of times and inspect the variability of the results.

Bloscpack Format

The input is split into chunks since a) we wish to put less stress on main memory and b) because Blosc has a buffer limit of 2GB (Version 1.0.0 and above). By default the chunk-size is a moderate 1MB which should be fine, even for less powerful machines.

In addition to the chunks some additional information must be added to the file for housekeeping:

The layout of the file is then:

|-header-|-meta-|-offsets-|-chunk-|-checksum-|-chunk-|-checksum-|...|

Description of the header

The following 32 byte header is used for Bloscpack as of version 0.3.0. The design goals of the header format are to contain as much information as possible to achieve interesting things in the future and to be as general as possible such that the persistence layer of Blaze/BLZ can be implemented without modification of the header format.

The following ASCII representation shows the layout of the header:

|-0-|-1-|-2-|-3-|-4-|-5-|-6-|-7-|-8-|-9-|-A-|-B-|-C-|-D-|-E-|-F-|
| b   l   p   k | ^ | ^ | ^ | ^ |   chunk-size  |  last-chunk   |
                  |   |   |   |
      version ----+   |   |   |
      options --------+   |   |
     checksum ------------+   |
     typesize ----------------+

|-0-|-1-|-2-|-3-|-4-|-5-|-6-|-7-|-8-|-9-|-A-|-B-|-C-|-D-|-E-|-F-|
|            nchunks            |        max-app-chunks         |

The first 4 bytes are the magic string blpk. Then there are 4 bytes which hold information about the activated features in this file. This is followed by 4 bytes for the chunk-size, another 4 bytes for the last-chunk-size, 8 bytes for the number of chunks, nchunks and lastly 8 bytes for the total number of chunks that can be appended to this file, max-app-chunks.

Effectively, storing the number of chunks as a signed 8 byte integer, limits the number of chunks to 2**63-1 = 9223372036854775807, but this should not be relevant in practice, since, even with the moderate default value of 1MB for chunk-size, we can still store files as large as 8ZB (!) Given that in 2012 the maximum size of a single file in the Zettabye File System (zfs) is 16EB, Bloscpack should be safe for a few more years.

Description of the header entries

All entries are little-endian.

The overall file-size can be computed as chunk-size * (nchunks - 1) + last-chunk-size. In a streaming scenario -1 can be used as a placeholder. For example if the total number of chunks, or the size of the last chunk is not known at the time the header is created.

The following constraints exist on the header entries:

• last-chunk must be less than or equal to chunk-size.
• nchunks + max_app_chunks must be less than or equal to the maximum value of an int64.

Description of the metadata section

This section goes after the header. It consists of a metadata-section header followed by a serialized and potentially compressed data section, followed by preallocated space to resize the data section, possibly followed by a checksum.

The layout of the section is thus:

|-metadata-header-|-data-|-prealloc-|-checksum-|

The header has the following layout:

|-0-|-1-|-2-|-3-|-4-|-5-|-6-|-7-|-8-|-9-|-A-|-B-|-C-|-D-|-E-|-F-|
|         magic-format          | ^ | ^ | ^ | ^ |   meta-size   |
                                  |   |   |   |
              meta-options -------+   |   |   |
              meta-checksum ----------+   |   |
              meta-codec -----------------+   |
              meta-level ---------------------+

|-0-|-1-|-2-|-3-|-4-|-5-|-6-|-7-|-8-|-9-|-A-|-B-|-C-|-D-|-E-|-F-|
| max-meta-size |meta-comp-size |            user-codec         |

The total space left for enlarging the metadata section is simply: max-meta-size - meta-comp-size.

JSON Example of serialized metadata:

'{"dtype": "float64", "shape": [1024], "others": []}'

If compression is requested, but not beneficial, because the compressed size would be larger than the uncompressed size, compression of the metadata is automatically deactivated.

As of Bloscpack version 0.3.0 only the JSON serializer is supported and used the string JSON followed by four whitespace bytes as identifier. Since JSON and any other of the suggested serializers has limitations, only a subset of Python structures can be stored, so probably some additional object handling must be done prior to serialize certain kinds of metadata.

Description of the offsets entries

Following the metadata section, comes a variable length section of chunk offsets. Offsets of the chunks into the file are to be used for accelerated seeking. The offsets (if activated) follow the header. Each offset is a 64 bit signed little-endian integer (int64). A value of -1 denotes an unknown offset. Initially, all offsets should be initialized to -1 and filled in after writing all chunks. Thus, If the compression of the file fails prematurely or is aborted, all offsets should have the value -1. Also, any unused offset entries preallocated to allow the file to grow should be set to -1. Each offset denotes the exact position of the chunk in the file such that seeking to the offset, will position the file pointer such that, reading the next 16 bytes gives the Blosc header, which is at the start of the desired chunk.

Description of the chunk format

As mentioned previously, each chunk is just a Blosc compressed string including header. The Blosc header (as of v1.0.0) is 16 bytes as follows:

|-0-|-1-|-2-|-3-|-4-|-5-|-6-|-7-|-8-|-9-|-A-|-B-|-C-|-D-|-E-|-F-|
  ^   ^   ^   ^ |     nbytes    |   blocksize   |    ctbytes    |
  |   |   |   |
  |   |   |   +--typesize
  |   |   +------flags
  |   +----------versionlz
  +--------------version

The first four are simply bytes, the last three are are each unsigned ints (uint32) each occupying 4 bytes. The header is always little-endian. ctbytes is the length of the buffer including header and nbytes is the length of the data when uncompressed. A more detailed description of the Blosc header can be found in the README_HEADER.rst of the Blosc repository

Overhead

Depending on which configuration for the file is used a constant, or linear overhead may be added to the file. The Bloscpack header adds 32 bytes in any case. If the data is non-compressible, Blosc will add 16 bytes of header to each chunk. The metadata section obviously adds a constant overhead, and if used, both the checksum and the offsets will add overhead to the file. The offsets add 8 bytes per chunk and the checksum adds a fixed constant value which depends on the checksum to each chunk. For example, 32 bytes for the adler32 checksum.

Coding Conventions

• Numpy rst style docstrings
• README cli examples should use long options
• testing: expected before received nt.assert_equal(expected, received)
• Debug messages: as close to where the data was generated
• Single quotes around ambiguities in messages overwriting existing file: 'testfile'
• Exceptions instead of exit
• Use the Wikipedia definition of compression ratio: http://en.wikipedia.org/wiki/Data_compression_ratio

How to Optimize Logging

Some care must be taken when logging in the inner loop. For example consider the following two commits:

• https://github.com/Blosc/bloscpack/commit/0854930514eebaf7dbc6c4dcf3589dbcb9f2fdc9
• https://github.com/Blosc/bloscpack/commit/355bf90a8c13a2a1f792d43228c2a68c61476621

If there are a larger number of chunks, calls to double_pretty_size will be executed (and may be costly) even if no logging is needed.

Consider the following script, loop-bench.py:

import numpy as np
import bloscpack as bp
import blosc

shuffle = True
clevel = 9
cname = 'lz4'

a = np.arange(2.5e8)

bargs = bp.args.BloscArgs(clevel=clevel, shuffle=shuffle, cname=cname)
bpargs = bp.BloscpackArgs(offsets=False, checksum='None', max_app_chunks=0)

Timing with v0.7.0:

In [1]: %run loop-bench.py

In [2]: %timeit bpc = bp.pack_ndarray_str(a, blosc_args=bargs, bloscpack_args=bpargs)
1 loops, best of 3: 423 ms per loop

In [3]: %timeit bpc = bp.pack_ndarray_str(a, blosc_args=bargs, bloscpack_args=bpargs)
1 loops, best of 3: 421 ms per loop

In [4]: bpc = bp.pack_ndarray_str(a, blosc_args=bargs, bloscpack_args=bpargs)

In [5]: %timeit a3 = bp.unpack_ndarray_str(bpc)
1 loops, best of 3: 727 ms per loop

In [6]: %timeit a3 = bp.unpack_ndarray_str(bpc)
1 loops, best of 3: 725 ms per loop

And then using a development version that contains the two optimization commits:

In [1]: %run loop-bench.py

In [2]: %timeit bpc = bp.pack_ndarray_str(a, blosc_args=bargs, bloscpack_args=bpargs)
1 loops, best of 3: 357 ms per loop

In [3]: %timeit bpc = bp.pack_ndarray_str(a, blosc_args=bargs, bloscpack_args=bpargs)
1 loops, best of 3: 357 ms per loop

In [4]: bpc = bp.pack_ndarray_str(a, blosc_args=bargs, bloscpack_args=bpargs)

In [5]: %timeit a3 = bp.unpack_ndarray_str(bpc)
1 loops, best of 3: 658 ms per loop

In [6]: %timeit a3 = bp.unpack_ndarray_str(bpc)
1 loops, best of 3: 655 ms per loop

Comparison to HDF5/PyTables

Since Blosc has already been supported for use in HDF5 files from within PyTables, one might be tempted to question why yet another file format has to be invented. This section aims to differentiate between HDF5/PyTables and effectively argues that they are not competitors.

• Lightweight vs. Heavyweight. Bloscpack is a lightweight format. The format specification can easily be digested within a day and the dependencies are minimal. PyTables is a complex piece of software and the HDF5 file format specification is a large document.
• Persistence vs. Database. Bloscpack is designed to allow for fast serialization and deserialization of in-memory data. PyTables is more of a database which for example allows complex queries to be computed on the data.

Additionally there are two network uses cases which Bloscpack is suited for (but does not have support for as of yet):

1. Streaming: Since bloscpack without offsets can be written in a single pass it is ideally suited for streaming over a network, where you can compress send and decompress individual chunks in a streaming fashion.
2. Expose a file over HTTP and do partial reads from it, for example when storing a compressed file in S3. You can easily just store a file on a web server and then use the header information to read and decompress individual chunks.

Prior Art

The following is a list of important resources that were read during the conception and initial stages of Bloscpack.

• The 6pack utility included with FastLZ (the codec that BloscLZ was derived from) was the initial inspiration for writing a command line interface to Blosc.
• The Wikipedia article on the PNG format contains some interesting details about the PNG header and file headers in general.
• The XZ File Format Specification gave rise to some ideas and techniques about writing file format specifications and using checksums for data integrity. Although the format and the document itself was a bit to heavyweight for my tastes.
• The Snappy framing format and the file container format for LZ4 were also consulted, but I can't remember if and what inspiration they gave rise to.
• The homepages of zlib and gzip were also consulted at some point. The command line interface of gzip/gunzip was deemed to be from a different era and as a result git-style subcommands are used in Bloscpack.

Resources and Related Publications

• Main Blosc website
• Francesc Alted. *The Data Access Problem* EuroScipy 2009 Keynote Presentation
• Francesc Alted. *Why modern CPUs are starving and what can be done about it*, Computing in Science & Engineering, Vol. 12, No. 2. (March 2010), pp. 68-71
• Francesc Alted: Sending Data from Memory to CPU (and back) faster than memcpy(). PyData London 2014 slides0 video0
• The Blosc Github organization
• Valentin Haenel. *Introducing Bloscpack* EuroScipy 2013 Presentation
• Valentin Haenel. *Bloscpack: a compressed lightweight serialization format for numerical data*. Proceedings of the 6th European Conference on Python in Science (EuroSciPy 2013).
• Valentin Haenel. Fast Serialization of Numpy Arrays with Bloscpack. PyData Berlin 2014 slides1, video1

Maintainers Notes on Cutting a Release

1. Set the version as environment variable VERSION=vX.X.X
2. Update the changelog and ANNOUNCE.rst
3. Commit using git commit -m "$VERSION changelog and ANNOUNCE.rst"
4. Set the version number in bloscpack/version.py
5. Commit with git commit -m "$VERSION"
6. Make the tag using git tag -s -m "Bloscpack $VERSION" $VERSION
7. Push commits to Blosc github git push blosc master
8. Push commits to own github git push esc master
9. Push the tag to Blosc github git push blosc $VERSION
10. Push the tag to own github git push esc $VERSION
11. Make a source distribution using python setup.py sdist bdist_wheel
12. Upload to PyPi using twine upload dist/bloscpack-$VERSION*
13. Bump version number to next dev version and reset ANNOUNCE.rst
14. Announce release on the Blosc list
15. Announce release via Twitter

TODO

Documentation

• Refactor monolithic readme into Sphinx and publish
• Cleanup and double check the docstrings for the public API classes

Command Line

• quiet verbosity level
• Expose the ability to set 'max_app_chunks' from the command line
• Allow to save metadata to a file during decompression
• subcommand e or estimate to estimate the size of the uncompressed data.
• subcommand v or verify to verify the integrity of the data
• add --raw-input and --raw-output switches to allow stuff like: cat file | blpk --raw-input --raw-output compress > file.blp
• Establish and document proper exit codes
• Document the metadata saved during Numpy serialization

Profiling and Optimization

• Use the faster version of struct where you have a single string
• Memory profiler, might be able to reduce memory used by reusing the buffer during compression and decompression
• Benchmark different codecs
• Use line profiler to check code
• Select different defaults for Numpy arrays, no offsets? no pre-alloc?

Library Features

• possibly provide a BloscPackFile abstraction, like GzipFile
• Allow to not-prealloc additional space for metadata
• Refactor certain collections of functions that operate on data into objects
  • Offsets (maybe)
• partial decompression?
• since we now have potentially small chunks, the progressbar becomes relevant again
• configuration file to store commonly used options on a given machine
• print the compression time, either as verbose or debug
• Investigate if we can use a StringIO object that returns memoryviews on read.
• Implement a memoryview Compressed/PlainSource
• Use a bytearray to read chunks from a file. Then re-use that bytearray during every read to avoid allocating deallocating strings the whole time.
• The keyword arguments to many functions are global dicts, this is a bad idea, Make the immutable with a forzendict.
• Check that the checksum is really being checked for all PlainSinks
• Bunch of NetworkSource/Sinks
• HTTPSource/Sink

Miscellaneous

• Announce on scipy/numpy lists, comp.compression, freshmeat, ohloh ...

Packaging and Infrastructure

• Debian packages (for python-blosc and bloscpack)
• Conda recipes (for python-blosc and bloscpack)
• Use tox for testing multiple python versions
• Build on travis and drone.io using pre-compiled

Changelog

• v0.16.0 - Thu 27 Dec 2018
  • Update of Python API and docs
  • various minor fixes
• v0.15.0 - Wed 31 Oct 2018
  • Halloween Release!
  • Adding the Blosc code of conduct (#79 by @esc)
  • Two new high-level functions: 'pack_bytes_to_bytes' and 'unpack_bytes_from_bytes' (#83 by @esc)
  • Fix incorrect check for typesize-chunksize mismatch (#81 by @esc)
  • Fix test to append without shuffle (#82 by @esc)
  • Fix tests to respect snappy not being available by default (#85 by @esc)
  • Fix tests to account for new default blocksize (#86 by @esc)
  • Enable testing on Python 3.7 via Travis (#84 by @esc)
• v0.14.0 - Thu Oct 18 2018
  • Remove official support for Python 2.6 (#77 by @esc)
• v0.13.0 - Thu May 24 2018
  • Add license file and include in sdist packages (#75 by @toddrme2178)
  • Print codec on info (#73 by @esc)
  • Decode Blosc flags (#72 by @esc)
  • Fix an embarrassing typo (#71 by @esc)
  • Test zstd (#70 by @esc)
  • Document args object (#69 by @esc)
  • Various pep8 fixes by @esc
  • Support for uploading wheels and using twine by @esc
  • Fix use of coverage by @esc
  • Better support for Python 2.6 by @esc
• v0.12.0 - Fri Mar 09 2018
  • Allow Pythonic None as checksum (#60 by @esc)
  • Fix failing tests to comply with latest Blosc (#63 and #64 by FrancescElies)
  • Support testing with Python 3.6 via Travis (#65 by @esc)
  • Unpinn Blosc in conda recipe (who uses this?) (#61 by @esc)
  • Cleanup README (#66 by @esc)
  • Fix Trove classifiers (#67 by @esc)
  • Random pep8 fixes by @esc
• v0.11.0 - Mon Aug 22 2016
  • Unpinn python-blosc and fix unit-tests (#51 and #57 fixed by @oogali)
  • Improve the computation of the chunksize when it is not divisible by typesize (#52 by FrancescAlted)
• v0.10.0 - Thu Dec 10 2015
  • Fix for compressing sliced arrays (#43 reported by @mistycheney)
  • Fix un/pack_bytes_file to be available from toplevel
  • Fix the badges to come (mostly) from https://img.shields.io
  • Fixes for travis-ci, test Python 3.5 too
  • Pin Blosc version to 1.2.7 via requirements.txt and setup.py due to breakage with Blosc 1.2.8.
• v0.9.0 - Tue Aug 18 2015
  • Use ast.literal_eval instead of np.safe_eval which is much faster (#39 @cpcloud)
  • Support for packing/unpacking bytes to/from file (#41)
• v0.8.0 - Sun Jul 12 2015
  • Python 3.x compatibility (#14)
• v0.7.3 - Sat Jul 11 2015
  • Fix deserialization of numpy arrays with nested dtypes that were created with versions v0.7.1 and before. (#37)
• v0.7.2 - Wed Mar 25 2015
  • Fix support for zero length arrays (and input in general) (#17 reported by @dmbelov)
  • Catch when typesize doesn't divide chunk_size (#18 reported by @dmbelov)
  • Fix serialization of object arrays (#16 reported by @dmbelov)
  • Reject Object dtype arrays since they cannot be compressed with Bloscpack
  • Provide backwards compatibility for older Numpy serializations
  • Fix win32 compatibility of tests (#27 fixed by @mindw)
  • Fix using setuptools for scripts and dependencies (#28 fixed by @mindw)
  • Various misc fixes
• v0.7.1 - Sun Jun 29 2014
  • Fix a bug related to setting the correct typesize when compressing Numpy arrays
  • Optimization of debug statements in the inner loops
• v0.7.0 - Wed May 28 2014
  • Modularize cram tests, even has something akin to a harness
  • Refactored, tweaked and simplified Source/Sink code and semantics
  • Various documentation improvements: listing prior art, comparison to HDF5
  • Improve benchmarking scripts
  • Introduce a BloscArgs object for saner handling of the BloscArgs
  • Introduce a BloscpackArgs object for saner handling of the BloscpackArgs
  • Introduce MetadataHeader and MetdataArgs objects too
  • Fix all (hopefully) incorrect uses of the term 'compression ratio'
  • Various miscellaneous fixes and improvements
• v0.6.0 - Fri Mar 28 2014
  • Complete refactor of Bloscpack codebase to support modularization
  • Support for drone.io CI service
  • Improved dependency specification for Python 2.6
  • Improved installation instructions
• v0.5.2 - Fri Mar 07 2014
  • Fix project url in setup.py
• v0.5.1 - Sat Feb 22 2014
  • Documentation fixes and improvements
• v0.5.0 - Sun Feb 02 2014
  • Moved project to the Blosc organization on Github
• v0.5.0-rc1 - Thu Jan 30 2014
  • Support for Blosc 1.3.x (alternative codecs)
• v0.4.1 - Fri Sep 27 2013
  • Fixed the pack_unpack_hard test suite
  • Fixed handling Numpy record and nested record arrays
• v0.4.0 - Sun Sep 15 2013
  • Fix a bug when serializing numpy arrays to strings
• v0.4.0-rc2 - Tue Sep 03 2013
  • Package available via PyPi (since 0.4.0-rc1)
  • Support for packing/unpacking numpy arrays to/from string
  • Check that string and record arrays work
  • Fix installation problems with PyPi package (Thanks to Olivier Grisel)
• v0.4.0-rc1 - Sun Aug 18 2013
  • BloscpackHeader class introduced
  • The info subcommand shows human readable sizes when printing the header
  • Now using Travis-CI for testing and Coveralls for coverage
  • Further work on the Plain/Compressed-Source/Sink abstractions
  • Start using memoryview in places
  • Learned to serialize Numpy arrays
• v0.3.0 - Sun Aug 04 2013
  • Minor readme fixes
  • Increase number of cram tests
• v0.3.0-rc1 - Thu Aug 01 2013
  • Bloscpack format changes (format version 3)
    • Variable length metadata section with it's own header
    • Ability to preallocate offsets for appending data (max_app_chunks)
  • Refactor compression and decompression to use file pointers instead of file name strings, allows using StringIO/cStringIO.
  • Sanitize calculation of nchunks and chunk-size
  • Special keyword max for use with chunk-size in the CLI
  • Support appending to a file and append subcommand (including the ability to preallocate offsets)
  • Support rudimentary info subcommand
  • Add tests of the command line interface using cram
  • Minor bugfixes and corrections as usual
• v0.2.1 - Mon Nov 26 2012
  • Backport to Python 2.6
  • Typo fixes in documentation
• v0.2.0 - Fri Sep 21 2012
  • Use atexit magic to remove test data on abort
  • Change prefix of temp directory to /tmp/blpk*
  • Merge header RFC into monolithic readme
• v0.2.0-rc2 - Tue Sep 18 2012
  • Don't bail out if the file is smaller than default chunk
  • Set the default typesize to 8 bytes
  • Upgrade dependencies to python-blosc v1.0.5 and fix tests
  • Make extreme test less resource intensive
  • Minor bugfixes and corrections
• v0.2.0-rc1 - Thu Sep 13 2012
  • Implement new header format as described in RFC
  • Implement checksumming compressed chunks with various checksums
  • Implement offsets of the chunks into the file
  • Efforts to make the library re-entrant, better control of side-effects
  • README is now rst not md (flirting with sphinx)
  • Tons of trivial fixes, typos, wording, refactoring, renaming, pep8 etc..
• v0.1.1 - Sun Jul 15 2012
  • Fix the memory issue with the tests
  • Two new suites: hard and extreme
  • Minor typo fixes and corrections
• v0.1.0 - Thu Jun 14 2012
  • Freeze the first 8 bytes of the header (hopefully for ever)
  • Fail to decompress on non-matching format version
  • Minor typo fixes and corrections
• v0.1.0-rc3 - Tue Jun 12 2012
  • Limit the chunk-size benchmark to a narrower range
  • After more careful experiments, a default chunk-size of 1MB was deemed most appropriate
  • Fixed a terrible bug, where during testing and benchmarking, temporary files were not removed, oups...
  • Adapted the header to have space for more chunks, include special marker for unknown chunk number (-1) and format version of the compressed file
  • Added a note in the README about instability of the file format
  • Various minor fixes and enhancements
• v0.1.0-rc2 - Sat Jun 09 2012
  • Default chunk-size now 4MB
  • Human readable chunk-size argument
  • Last chunk now contains remainder
  • Pure python benchmark to compare against gzip
  • Benchmark to measure the effect of chunk-size
  • Various minor fixes and enhancements
• v0.1.0-rc1 - Sun May 27 2012
  • Initial version
  • Compression/decompression
  • Command line argument parser
  • README, setup.py, tests and benchmark

Thanks

• Francesc Alted for writing Blosc in the first place, for providing continual code-review and feedback on Bloscpack and for co-authoring the Bloscpack file-format specification.

Author, Copyright and License

© 2012-2020 Valentin Haenel <[email protected]>

Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:

1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.

2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.

3. Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.