Skip to main content

Fast random access of gzip files in Python

Project description


PyPi version Anaconda versionTest status

Fast random access of gzip files in Python


The indexed_gzip project is a Python extension which aims to provide a drop-in replacement for the built-in Python gzip.GzipFile class, the IndexedGzipFile.

indexed_gzip was written to allow fast random access of compressed NIFTI image files (for which GZIP is the de-facto compression standard), but will work with any GZIP file. indexed_gzip is easy to use with nibabel (

The standard gzip.GzipFile class exposes a random access-like interface (via its seek and read methods), but every time you seek to a new point in the uncompressed data stream, the GzipFile instance has to start decompressing from the beginning of the file, until it reaches the requested location.

An IndexedGzipFile instance gets around this performance limitation by building an index, which contains seek points, mappings between corresponding locations in the compressed and uncompressed data streams. Each seek point is accompanied by a chunk (32KB) of uncompressed data which is used to initialise the decompression algorithm, allowing us to start reading from any seek point. If the index is built with a seek point spacing of 1MB, we only have to decompress (on average) 512KB of data to read from any location in the file.

Intended use

You may find indexed_gzip useful if you need to read from large GZIP files. A major advantage of indexed_gzip is that it will work with any GZIP file. However, if you have control over the creation of your GZIP files, you may wish to consider some alternatives:

  • mgzip provides an accelerated GZIP compression and decompression library.
  • Compression formats other than GZIP, such as bzip2 and xz, have better support for random access.


indexed_gzip is available on PyPi - to install, simply type:

pip install indexed_gzip

You can also install indexed_gzip from conda-forge:

conda install -c conda-forge indexed_gzip

To compile indexed_gzip, make sure you have cython installed (and numpy if you want to compile the tests), and then run:

python develop

To run the tests, type the following; you will need numpy, nibabel, pytest, pytest-cov, and coverage installed:

python -m indexed_gzip.tests


You can use the indexed_gzip module directly:

import indexed_gzip as igzip

# You can create an IndexedGzipFile instance
# by specifying a file name, or an open file
# handle. For the latter use, the file handle
# must be opened in read-only binary mode.
# Write support is currently non-existent.
myfile = igzip.IndexedGzipFile('big_file.gz')

some_offset_into_uncompressed_data = 234195

# The index will be automatically
# built on-demand when seeking.
data =

Using with in-memory data

You can use indexed_gzip with any Python file-like object. For example:

import io
import indexed_gzip as igzip

# Load some gzip data from somewhere
with open('my_file.gz') as f:
    data =

# Create an IndexedGzipFile based on the
# in-memory data buffer
gzf = igzip.IndexedGzipFile(fileobj=io.BytesIO(data))
uncompressed =

Using with nibabel

You can use indexed_gzip with nibabel. nibabel >= 2.3.0 will automatically use indexed_gzip if it is present:

import nibabel as nib

image = nib.load('big_image.nii.gz')

If you are using nibabel 2.2.x, you need to explicitly set the keep_file_open flag:

import nibabel as nib

image = nib.load('big_image.nii.gz', keep_file_open='auto')

To use indexed_gzip with nibabel 2.1.0 or older, you need to do a little more work:

import nibabel      as nib
import indexed_gzip as igzip

# Here we are using 4MB spacing between
# seek points, and using a larger read
# buffer (than the default size of 16KB).
fobj = igzip.IndexedGzipFile(

# Create a nibabel image using
# the existing file handle.
fmap = nib.Nifti1Image.make_file_map()
fmap['image'].fileobj = fobj
image = nib.Nifti1Image.from_file_map(fmap)

# Use the image ArrayProxy to access the
# data - the index will automatically be
# built as data is accessed.
vol3 = image.dataobj[:, :, :, 3]

Index import/export

If you have a large file, you may wish to pre-generate the index once, and save it out to an index file:

import indexed_gzip as igzip

# Load the file, pre-generate the
# index, and save it out to disk.
fobj = igzip.IndexedGzipFile('big_file.gz')

The next time you open the same file, you can load in the index:

import indexed_gip as igzip
fobj = igzip.IndexedGzipFile('big_file.gz', index_file='big_file.gzidx')

Write support

indexed_gzip does not currently have any support for writing. Currently if you wish to write to a file, you will need to save the file by alternate means (e.g. via gzip or nibabel), and then re-create a new IndexedGzipFile instance. For example:

import nibabel as nib

# Load the entire image into memory
image = nib.load('big_image.nii.gz')
data = image.get_data()

# Make changes to the data
data[:, :, :, 5] *= 100

# Save the image using nibabel, 'big_image.nii.gz')

# Re-load the image
image = nib.load('big_image.nii.gz')


A small test script is included with indexed_gzip; this script compares the performance of the IndexedGzipFile class with the gzip.GzipFile class. This script does the following:

  1. Generates a test file.

  2. Generates a specified number of seek locations, uniformly spaced throughout the test file.

  3. Randomly shuffles these locations

  4. Seeks to each location, and reads a chunk of data from the file.

This plot shows the results of this test for a few compresed files of varying sizes, with 500 seeks:

Indexed gzip performance


The indexed_gzip project is based upon the zran.c example (written by Mark Alder) which ships with the zlib source code.

indexed_gzip was originally inspired by Zalan Rajna's (@zrajna) zindex project:

Z. Rajna, A. Keskinarkaus, V. Kiviniemi and T. Seppanen
"Speeding up the file access of large compressed NIfTI neuroimaging data"
Engineering in Medicine and Biology Society (EMBC), 2015 37th Annual
International Conference of the IEEE, Milan, 2015, pp. 654-657.

Initial work on indexed_gzip took place at Brainhack Paris, at the Institut Pasteur, 24th-26th February 2016, with the support of the FMRIB Centre, at the University of Oxford, UK.

Many thanks to the following contributors (listed chronologically):

  • Zalan Rajna (@zrajna): Bug fixes (#2)
  • Martin Craig (@mcraig-ibme): Porting indexed_gzip to Windows (#3)
  • Chris Markiewicz (@effigies): Option to drop file handles (#6)
  • Omer Ozarslan (@ozars): Index import/export (#8)
  • @DarioDaF: Windows overflow bug (#30)
  • Sławomir Zborowski (@szborows): seek_points method (#35), README fixes (#34)
  • Ashwin Ramaswami (@epicfaace): Support for in-memory file objects (#55), bug fixes (#63, #64, #65).
  • Michał Górny (@mgorny): Remove hard dependency on nibabel from test suite (#78).
  • Alexander Gorban (@alexgorban) Fix memory leak (#82, #83).
  • Maximilian Knespel (@mxmlnkn) Change default read buffer size to improve performance (#90).
  • Ben Beasley (@musicinmybrain) Python 3.12 compatibility (#126).


indexed_gzip inherits the zlib license, available for perusal in the LICENSE file.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

indexed_gzip-1.8.5.tar.gz (74.7 kB view hashes)

Uploaded source

Built Distributions

indexed_gzip-1.8.5-cp311-cp311-win_amd64.whl (843.5 kB view hashes)

Uploaded cp311

indexed_gzip-1.8.5-cp311-cp311-win32.whl (794.7 kB view hashes)

Uploaded cp311

indexed_gzip-1.8.5-cp310-cp310-win_amd64.whl (840.1 kB view hashes)

Uploaded cp310

indexed_gzip-1.8.5-cp310-cp310-win32.whl (794.7 kB view hashes)

Uploaded cp310

indexed_gzip-1.8.5-cp39-cp39-win_amd64.whl (840.8 kB view hashes)

Uploaded cp39

indexed_gzip-1.8.5-cp39-cp39-win32.whl (795.3 kB view hashes)

Uploaded cp39

indexed_gzip-1.8.5-cp38-cp38-win_amd64.whl (843.2 kB view hashes)

Uploaded cp38

indexed_gzip-1.8.5-cp38-cp38-win32.whl (797.4 kB view hashes)

Uploaded cp38

indexed_gzip-1.8.5-cp37-cp37m-win_amd64.whl (820.0 kB view hashes)

Uploaded cp37

indexed_gzip-1.8.5-cp37-cp37m-win32.whl (784.0 kB view hashes)

Uploaded cp37

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page