zarr

A minimal implementation of chunked, compressed, N-dimensional arrays for Python.

• Source code: https://github.com/alimanfoo/zarr
• Download: https://pypi.python.org/pypi/zarr
• Release notes: https://github.com/alimanfoo/zarr/releases

Installation

Installation requires Numpy and Cython pre-installed. Can only be installed on Linux currently.

Install from PyPI:

$ pip install -U zarr

Install from GitHub:

$ pip install -U git+https://github.com/alimanfoo/zarr.git@master

Status

Experimental, proof-of-concept. This is alpha-quality software. Things may break, change or disappear without warning.

Bug reports and suggestions welcome.

Design goals

• Chunking in multiple dimensions
• Resize any dimension
• Concurrent reads
• Concurrent writes
• Release the GIL during compression and decompression

Usage

Create an array:

>>> import numpy as np
>>> import zarr
>>> z = zarr.empty(shape=(10000, 1000), dtype='i4', chunks=(1000, 100))
>>> z
zarr.ext.SynchronizedArray((10000, 1000), int32, chunks=(1000, 100))
  cname: blosclz; clevel: 5; shuffle: 1 (BYTESHUFFLE)
  nbytes: 38.1M; cbytes: 0; initialized: 0/100

Fill it with some data:

>>> z[:] = np.arange(10000000, dtype='i4').reshape(10000, 1000)
>>> z
zarr.ext.SynchronizedArray((10000, 1000), int32, chunks=(1000, 100))
  cname: blosclz; clevel: 5; shuffle: 1 (BYTESHUFFLE)
  nbytes: 38.1M; cbytes: 2.0M; ratio: 19.3; initialized: 100/100

Obtain a NumPy array by slicing:

>>> z[:]
array([[      0,       1,       2, ...,     997,     998,     999],
       [   1000,    1001,    1002, ...,    1997,    1998,    1999],
       [   2000,    2001,    2002, ...,    2997,    2998,    2999],
       ...,
       [9997000, 9997001, 9997002, ..., 9997997, 9997998, 9997999],
       [9998000, 9998001, 9998002, ..., 9998997, 9998998, 9998999],
       [9999000, 9999001, 9999002, ..., 9999997, 9999998, 9999999]], dtype=int32)
>>> z[:100]
array([[    0,     1,     2, ...,   997,   998,   999],
       [ 1000,  1001,  1002, ...,  1997,  1998,  1999],
       [ 2000,  2001,  2002, ...,  2997,  2998,  2999],
       ...,
       [97000, 97001, 97002, ..., 97997, 97998, 97999],
       [98000, 98001, 98002, ..., 98997, 98998, 98999],
       [99000, 99001, 99002, ..., 99997, 99998, 99999]], dtype=int32)
>>> z[:, :100]
array([[      0,       1,       2, ...,      97,      98,      99],
       [   1000,    1001,    1002, ...,    1097,    1098,    1099],
       [   2000,    2001,    2002, ...,    2097,    2098,    2099],
       ...,
       [9997000, 9997001, 9997002, ..., 9997097, 9997098, 9997099],
       [9998000, 9998001, 9998002, ..., 9998097, 9998098, 9998099],
       [9999000, 9999001, 9999002, ..., 9999097, 9999098, 9999099]], dtype=int32)

Resize the array and add more data:

>>> z.resize(20000, 1000)
>>> z
zarr.ext.SynchronizedArray((20000, 1000), int32, chunks=(1000, 100))
  cname: blosclz; clevel: 5; shuffle: 1 (BYTESHUFFLE)
  nbytes: 76.3M; cbytes: 2.0M; ratio: 38.5; initialized: 100/200
>>> z[10000:, :] = np.arange(10000000, dtype='i4').reshape(10000, 1000)
>>> z
zarr.ext.SynchronizedArray((20000, 1000), int32, chunks=(1000, 100))
  cname: blosclz; clevel: 5; shuffle: 1 (BYTESHUFFLE)
  nbytes: 76.3M; cbytes: 4.0M; ratio: 19.3; initialized: 200/200

For convenience, an append() method is also available, which can be used to append data to any axis:

>>> a = np.arange(10000000, dtype='i4').reshape(10000, 1000)
>>> z = zarr.array(a, chunks=(1000, 100))
>>> z.append(a+a)
>>> z
zarr.ext.SynchronizedArray((20000, 1000), int32, chunks=(1000, 100))
  cname: blosclz; clevel: 5; shuffle: 1 (BYTESHUFFLE)
  nbytes: 76.3M; cbytes: 3.6M; ratio: 21.2; initialized: 200/200
>>> z.append(np.vstack([a, a]), axis=1)
>>> z
zarr.ext.SynchronizedArray((20000, 2000), int32, chunks=(1000, 100))
  cname: blosclz; clevel: 5; shuffle: 1 (BYTESHUFFLE)
  nbytes: 152.6M; cbytes: 7.6M; ratio: 20.2; initialized: 400/400

Persistence

Create a persistent array (data stored on disk):

>>> path = 'example.zarr'
>>> z = zarr.open(path, mode='w', shape=(10000, 1000), dtype='i4', chunks=(1000, 100))
>>> z[:] = np.arange(10000000, dtype='i4').reshape(10000, 1000)
>>> z
zarr.ext.SynchronizedPersistentArray((10000, 1000), int32, chunks=(1000, 100))
  cname: blosclz; clevel: 5; shuffle: 1 (BYTESHUFFLE)
  nbytes: 38.1M; cbytes: 2.0M; ratio: 19.3; initialized: 100/100
  mode: w; path: example.zarr

There is no need to close a persistent array. Data are automatically flushed to disk.

If you're working with really big arrays, try the 'lazy' option:

>>> path = 'big.zarr'
>>> z = zarr.open(path, mode='w', shape=(1e8, 1e7), dtype='i4', chunks=(1000, 1000), lazy=True)
>>> z
zarr.ext.SynchronizedLazyPersistentArray((100000000, 10000000), int32, chunks=(1000, 1000))
  cname: blosclz; clevel: 5; shuffle: 1 (BYTESHUFFLE)
  nbytes: 3.6P; cbytes: 0; initialized: 0/1000000000
  mode: w; path: big.zarr

See the persistence documentation for more details of the file format.

Tuning

zarr is optimised for accessing and storing data in contiguous slices, of the same size or larger than chunks. It is not and probably never will be optimised for single item access.

Chunks sizes >= 1M are generally good. Optimal chunk shape will depend on the correlation structure in your data.

zarr is designed for use in parallel computations working chunk-wise over data. Try it with dask.array. If using in a multi-threaded, set zarr to use blosc in contextual mode:

>>> zarr.set_blosc_options(use_context=True)

Acknowledgments

zarr uses c-blosc internally for compression and decompression and borrows code heavily from bcolz.