Tutorial#

Zarr provides classes and functions for working with N-dimensional arrays that behave like NumPy arrays but whose data is divided into chunks and each chunk is compressed. If you are already familiar with HDF5 then Zarr arrays provide similar functionality, but with some additional flexibility.

Creating an array#

Zarr has several functions for creating arrays. For example:

>>> import zarr
>>> z = zarr.zeros((10000, 10000), chunks=(1000, 1000), dtype='i4')
>>> z
<zarr.core.Array (10000, 10000) int32>

The code above creates a 2-dimensional array of 32-bit integers with 10000 rows and 10000 columns, divided into chunks where each chunk has 1000 rows and 1000 columns (and so there will be 100 chunks in total).

For a complete list of array creation routines see the zarr.creation module documentation.

Reading and writing data#

Zarr arrays support a similar interface to NumPy arrays for reading and writing data. For example, the entire array can be filled with a scalar value:

>>> z[:] = 42

Regions of the array can also be written to, e.g.:

>>> import numpy as np
>>> z[0, :] = np.arange(10000)
>>> z[:, 0] = np.arange(10000)

The contents of the array can be retrieved by slicing, which will load the requested region into memory as a NumPy array, e.g.:

>>> z[0, 0]
0
>>> z[-1, -1]
42
>>> z[0, :]
array([   0,    1,    2, ..., 9997, 9998, 9999], dtype=int32)
>>> z[:, 0]
array([   0,    1,    2, ..., 9997, 9998, 9999], dtype=int32)
>>> z[:]
array([[   0,    1,    2, ..., 9997, 9998, 9999],
       [   1,   42,   42, ...,   42,   42,   42],
       [   2,   42,   42, ...,   42,   42,   42],
       ...,
       [9997,   42,   42, ...,   42,   42,   42],
       [9998,   42,   42, ...,   42,   42,   42],
       [9999,   42,   42, ...,   42,   42,   42]], dtype=int32)

Persistent arrays#

In the examples above, compressed data for each chunk of the array was stored in main memory. Zarr arrays can also be stored on a file system, enabling persistence of data between sessions. For example:

>>> z1 = zarr.open('data/example.zarr', mode='w', shape=(10000, 10000),
...                chunks=(1000, 1000), dtype='i4')

The array above will store its configuration metadata and all compressed chunk data in a directory called ‘data/example.zarr’ relative to the current working directory. The zarr.convenience.open() function provides a convenient way to create a new persistent array or continue working with an existing array. Note that although the function is called “open”, there is no need to close an array: data are automatically flushed to disk, and files are automatically closed whenever an array is modified.

Persistent arrays support the same interface for reading and writing data, e.g.:

>>> z1[:] = 42
>>> z1[0, :] = np.arange(10000)
>>> z1[:, 0] = np.arange(10000)

Check that the data have been written and can be read again:

>>> z2 = zarr.open('data/example.zarr', mode='r')
>>> np.all(z1[:] == z2[:])
True

If you are just looking for a fast and convenient way to save NumPy arrays to disk then load back into memory later, the functions zarr.convenience.save() and zarr.convenience.load() may be useful. E.g.:

>>> a = np.arange(10)
>>> zarr.save('data/example.zarr', a)
>>> zarr.load('data/example.zarr')
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])

Please note that there are a number of other options for persistent array storage, see the section on Storage alternatives below.

Resizing and appending#

A Zarr array can be resized, which means that any of its dimensions can be increased or decreased in length. For example:

>>> z = zarr.zeros(shape=(10000, 10000), chunks=(1000, 1000))
>>> z[:] = 42
>>> z.resize(20000, 10000)
>>> z.shape
(20000, 10000)

Note that when an array is resized, the underlying data are not rearranged in any way. If one or more dimensions are shrunk, any chunks falling outside the new array shape will be deleted from the underlying store.

For convenience, Zarr arrays also provide an append() method, which can be used to append data to any axis. E.g.:

>>> a = np.arange(10000000, dtype='i4').reshape(10000, 1000)
>>> z = zarr.array(a, chunks=(1000, 100))
>>> z.shape
(10000, 1000)
>>> z.append(a)
(20000, 1000)
>>> z.append(np.vstack([a, a]), axis=1)
(20000, 2000)
>>> z.shape
(20000, 2000)

Compressors#

A number of different compressors can be used with Zarr. A separate package called NumCodecs is available which provides a common interface to various compressor libraries including Blosc, Zstandard, LZ4, Zlib, BZ2 and LZMA. Different compressors can be provided via the compressor keyword argument accepted by all array creation functions. For example:

>>> from numcodecs import Blosc
>>> compressor = Blosc(cname='zstd', clevel=3, shuffle=Blosc.BITSHUFFLE)
>>> data = np.arange(100000000, dtype='i4').reshape(10000, 10000)
>>> z = zarr.array(data, chunks=(1000, 1000), compressor=compressor)
>>> z.compressor
Blosc(cname='zstd', clevel=3, shuffle=BITSHUFFLE, blocksize=0)

This array above will use Blosc as the primary compressor, using the Zstandard algorithm (compression level 3) internally within Blosc, and with the bit-shuffle filter applied.

When using a compressor, it can be useful to get some diagnostics on the compression ratio. Zarr arrays provide a info property which can be used to print some diagnostics, e.g.:

>>> z.info
Type               : zarr.core.Array
Data type          : int32
Shape              : (10000, 10000)
Chunk shape        : (1000, 1000)
Order              : C
Read-only          : False
Compressor         : Blosc(cname='zstd', clevel=3, shuffle=BITSHUFFLE,
                   : blocksize=0)
Store type         : zarr.storage.KVStore
No. bytes          : 400000000 (381.5M)
No. bytes stored   : 3379344 (3.2M)
Storage ratio      : 118.4
Chunks initialized : 100/100

If you don’t specify a compressor, by default Zarr uses the Blosc compressor. Blosc is generally very fast and can be configured in a variety of ways to improve the compression ratio for different types of data. Blosc is in fact a “meta-compressor”, which means that it can use a number of different compression algorithms internally to compress the data. Blosc also provides highly optimized implementations of byte- and bit-shuffle filters, which can improve compression ratios for some data. A list of the internal compression libraries available within Blosc can be obtained via:

>>> from numcodecs import blosc
>>> blosc.list_compressors()
['blosclz', 'lz4', 'lz4hc', 'snappy', 'zlib', 'zstd']

In addition to Blosc, other compression libraries can also be used. For example, here is an array using Zstandard compression, level 1:

>>> from numcodecs import Zstd
>>> z = zarr.array(np.arange(100000000, dtype='i4').reshape(10000, 10000),
...                chunks=(1000, 1000), compressor=Zstd(level=1))
>>> z.compressor
Zstd(level=1)

Here is an example using LZMA with a custom filter pipeline including LZMA’s built-in delta filter:

>>> import lzma
>>> lzma_filters = [dict(id=lzma.FILTER_DELTA, dist=4),
...                 dict(id=lzma.FILTER_LZMA2, preset=1)]
>>> from numcodecs import LZMA
>>> compressor = LZMA(filters=lzma_filters)
>>> z = zarr.array(np.arange(100000000, dtype='i4').reshape(10000, 10000),
...                chunks=(1000, 1000), compressor=compressor)
>>> z.compressor
LZMA(format=1, check=-1, preset=None, filters=[{'dist': 4, 'id': 3}, {'id': 33, 'preset': 1}])

The default compressor can be changed by setting the value of the zarr.storage.default_compressor variable, e.g.:

>>> import zarr.storage
>>> from numcodecs import Zstd, Blosc
>>> # switch to using Zstandard
... zarr.storage.default_compressor = Zstd(level=1)
>>> z = zarr.zeros(100000000, chunks=1000000)
>>> z.compressor
Zstd(level=1)
>>> # switch back to Blosc defaults
... zarr.storage.default_compressor = Blosc()

To disable compression, set compressor=None when creating an array, e.g.:

>>> z = zarr.zeros(100000000, chunks=1000000, compressor=None)
>>> z.compressor is None
True

Filters#

In some cases, compression can be improved by transforming the data in some way. For example, if nearby values tend to be correlated, then shuffling the bytes within each numerical value or storing the difference between adjacent values may increase compression ratio. Some compressors provide built-in filters that apply transformations to the data prior to compression. For example, the Blosc compressor has built-in implementations of byte- and bit-shuffle filters, and the LZMA compressor has a built-in implementation of a delta filter. However, to provide additional flexibility for implementing and using filters in combination with different compressors, Zarr also provides a mechanism for configuring filters outside of the primary compressor.

Here is an example using a delta filter with the Blosc compressor:

>>> from numcodecs import Blosc, Delta
>>> filters = [Delta(dtype='i4')]
>>> compressor = Blosc(cname='zstd', clevel=1, shuffle=Blosc.SHUFFLE)
>>> data = np.arange(100000000, dtype='i4').reshape(10000, 10000)
>>> z = zarr.array(data, chunks=(1000, 1000), filters=filters, compressor=compressor)
>>> z.info
Type               : zarr.core.Array
Data type          : int32
Shape              : (10000, 10000)
Chunk shape        : (1000, 1000)
Order              : C
Read-only          : False
Filter [0]         : Delta(dtype='<i4')
Compressor         : Blosc(cname='zstd', clevel=1, shuffle=SHUFFLE, blocksize=0)
Store type         : zarr.storage.KVStore
No. bytes          : 400000000 (381.5M)
No. bytes stored   : 1290562 (1.2M)
Storage ratio      : 309.9
Chunks initialized : 100/100

For more information about available filter codecs, see the Numcodecs documentation.

Groups#

Zarr supports hierarchical organization of arrays via groups. As with arrays, groups can be stored in memory, on disk, or via other storage systems that support a similar interface.

To create a group, use the zarr.group() function:

>>> root = zarr.group()
>>> root
<zarr.hierarchy.Group '/'>

Groups have a similar API to the Group class from h5py. For example, groups can contain other groups:

>>> foo = root.create_group('foo')
>>> bar = foo.create_group('bar')

Groups can also contain arrays, e.g.:

>>> z1 = bar.zeros('baz', shape=(10000, 10000), chunks=(1000, 1000), dtype='i4')
>>> z1
<zarr.core.Array '/foo/bar/baz' (10000, 10000) int32>

Arrays are known as “datasets” in HDF5 terminology. For compatibility with h5py, Zarr groups also implement the create_dataset() and require_dataset() methods, e.g.:

>>> z = bar.create_dataset('quux', shape=(10000, 10000), chunks=(1000, 1000), dtype='i4')
>>> z
<zarr.core.Array '/foo/bar/quux' (10000, 10000) int32>

Members of a group can be accessed via the suffix notation, e.g.:

>>> root['foo']
<zarr.hierarchy.Group '/foo'>

The ‘/’ character can be used to access multiple levels of the hierarchy in one call, e.g.:

>>> root['foo/bar']
<zarr.hierarchy.Group '/foo/bar'>
>>> root['foo/bar/baz']
<zarr.core.Array '/foo/bar/baz' (10000, 10000) int32>

The zarr.hierarchy.Group.tree() method can be used to print a tree representation of the hierarchy, e.g.:

>>> root.tree()
/
 └── foo
     └── bar
         ├── baz (10000, 10000) int32
         └── quux (10000, 10000) int32

The zarr.convenience.open() function provides a convenient way to create or re-open a group stored in a directory on the file-system, with sub-groups stored in sub-directories, e.g.:

>>> root = zarr.open('data/group.zarr', mode='w')
>>> root
<zarr.hierarchy.Group '/'>
>>> z = root.zeros('foo/bar/baz', shape=(10000, 10000), chunks=(1000, 1000), dtype='i4')
>>> z
<zarr.core.Array '/foo/bar/baz' (10000, 10000) int32>

Groups can be used as context managers (in a with statement). If the underlying store has a close method, it will be called on exit.

For more information on groups see the zarr.hierarchy and zarr.convenience API docs.

Array and group diagnostics#

Diagnostic information about arrays and groups is available via the info property. E.g.:

>>> root = zarr.group()
>>> foo = root.create_group('foo')
>>> bar = foo.zeros('bar', shape=1000000, chunks=100000, dtype='i8')
>>> bar[:] = 42
>>> baz = foo.zeros('baz', shape=(1000, 1000), chunks=(100, 100), dtype='f4')
>>> baz[:] = 4.2
>>> root.info
Name        : /
Type        : zarr.hierarchy.Group
Read-only   : False
Store type  : zarr.storage.MemoryStore
No. members : 1
No. arrays  : 0
No. groups  : 1
Groups      : foo

>>> foo.info
Name        : /foo
Type        : zarr.hierarchy.Group
Read-only   : False
Store type  : zarr.storage.MemoryStore
No. members : 2
No. arrays  : 2
No. groups  : 0
Arrays      : bar, baz

>>> bar.info
Name               : /foo/bar
Type               : zarr.core.Array
Data type          : int64
Shape              : (1000000,)
Chunk shape        : (100000,)
Order              : C
Read-only          : False
Compressor         : Blosc(cname='lz4', clevel=5, shuffle=SHUFFLE, blocksize=0)
Store type         : zarr.storage.MemoryStore
No. bytes          : 8000000 (7.6M)
No. bytes stored   : 33240 (32.5K)
Storage ratio      : 240.7
Chunks initialized : 10/10

>>> baz.info
Name               : /foo/baz
Type               : zarr.core.Array
Data type          : float32
Shape              : (1000, 1000)
Chunk shape        : (100, 100)
Order              : C
Read-only          : False
Compressor         : Blosc(cname='lz4', clevel=5, shuffle=SHUFFLE, blocksize=0)
Store type         : zarr.storage.MemoryStore
No. bytes          : 4000000 (3.8M)
No. bytes stored   : 23943 (23.4K)
Storage ratio      : 167.1
Chunks initialized : 100/100

Groups also have the zarr.hierarchy.Group.tree() method, e.g.:

>>> root.tree()
/
 └── foo
     ├── bar (1000000,) int64
     └── baz (1000, 1000) float32

If you’re using Zarr within a Jupyter notebook (requires ipytree), calling tree() will generate an interactive tree representation, see the repr_tree.ipynb notebook for more examples.

User attributes#

Zarr arrays and groups support custom key/value attributes, which can be useful for storing application-specific metadata. For example:

>>> root = zarr.group()
>>> root.attrs['foo'] = 'bar'
>>> z = root.zeros('zzz', shape=(10000, 10000))
>>> z.attrs['baz'] = 42
>>> z.attrs['qux'] = [1, 4, 7, 12]
>>> sorted(root.attrs)
['foo']
>>> 'foo' in root.attrs
True
>>> root.attrs['foo']
'bar'
>>> sorted(z.attrs)
['baz', 'qux']
>>> z.attrs['baz']
42
>>> z.attrs['qux']
[1, 4, 7, 12]

Internally Zarr uses JSON to store array attributes, so attribute values must be JSON serializable.

Advanced indexing#

As of version 2.2, Zarr arrays support several methods for advanced or “fancy” indexing, which enable a subset of data items to be extracted or updated in an array without loading the entire array into memory.

Note that although this functionality is similar to some of the advanced indexing capabilities available on NumPy arrays and on h5py datasets, the Zarr API for advanced indexing is different from both NumPy and h5py, so please read this section carefully. For a complete description of the indexing API, see the documentation for the zarr.core.Array class.

Indexing with coordinate arrays#

Items from a Zarr array can be extracted by providing an integer array of coordinates. E.g.:

>>> z = zarr.array(np.arange(10) ** 2)
>>> z[:]
array([ 0,  1,  4,  9, 16, 25, 36, 49, 64, 81])
>>> z.get_coordinate_selection([2, 5])
array([ 4, 25])

Coordinate arrays can also be used to update data, e.g.:

>>> z.set_coordinate_selection([2, 5], [-1, -2])
>>> z[:]
array([ 0,  1, -1,  9, 16, -2, 36, 49, 64, 81])

For multidimensional arrays, coordinates must be provided for each dimension, e.g.:

>>> z = zarr.array(np.arange(15).reshape(3, 5))
>>> z[:]
array([[ 0,  1,  2,  3,  4],
       [ 5,  6,  7,  8,  9],
       [10, 11, 12, 13, 14]])
>>> z.get_coordinate_selection(([0, 2], [1, 3]))
array([ 1, 13])
>>> z.set_coordinate_selection(([0, 2], [1, 3]), [-1, -2])
>>> z[:]
array([[ 0, -1,  2,  3,  4],
       [ 5,  6,  7,  8,  9],
       [10, 11, 12, -2, 14]])

For convenience, coordinate indexing is also available via the vindex property, as well as the square bracket operator, e.g.:

>>> z.vindex[[0, 2], [1, 3]]
array([-1, -2])
>>> z.vindex[[0, 2], [1, 3]] = [-3, -4]
>>> z[:]
array([[ 0, -3,  2,  3,  4],
       [ 5,  6,  7,  8,  9],
       [10, 11, 12, -4, 14]])
>>> z[[0, 2], [1, 3]]
array([-3, -4])

When the indexing arrays have different shapes, they are broadcast together. That is, the following two calls are equivalent:

>>> z[1, [1, 3]]
array([6, 8])
>>> z[[1, 1], [1, 3]]
array([6, 8])

Indexing with a mask array#

Items can also be extracted by providing a Boolean mask. E.g.:

>>> z = zarr.array(np.arange(10) ** 2)
>>> z[:]
array([ 0,  1,  4,  9, 16, 25, 36, 49, 64, 81])
>>> sel = np.zeros_like(z, dtype=bool)
>>> sel[2] = True
>>> sel[5] = True
>>> z.get_mask_selection(sel)
array([ 4, 25])
>>> z.set_mask_selection(sel, [-1, -2])
>>> z[:]
array([ 0,  1, -1,  9, 16, -2, 36, 49, 64, 81])

Here’s a multidimensional example:

>>> z = zarr.array(np.arange(15).reshape(3, 5))
>>> z[:]
array([[ 0,  1,  2,  3,  4],
       [ 5,  6,  7,  8,  9],
       [10, 11, 12, 13, 14]])
>>> sel = np.zeros_like(z, dtype=bool)
>>> sel[0, 1] = True
>>> sel[2, 3] = True
>>> z.get_mask_selection(sel)
array([ 1, 13])
>>> z.set_mask_selection(sel, [-1, -2])
>>> z[:]
array([[ 0, -1,  2,  3,  4],
       [ 5,  6,  7,  8,  9],
       [10, 11, 12, -2, 14]])

For convenience, mask indexing is also available via the vindex property, e.g.:

>>> z.vindex[sel]
array([-1, -2])
>>> z.vindex[sel] = [-3, -4]
>>> z[:]
array([[ 0, -3,  2,  3,  4],
       [ 5,  6,  7,  8,  9],
       [10, 11, 12, -4, 14]])

Mask indexing is conceptually the same as coordinate indexing, and is implemented internally via the same machinery. Both styles of indexing allow selecting arbitrary items from an array, also known as point selection.

Orthogonal indexing#

Zarr arrays also support methods for orthogonal indexing, which allows selections to be made along each dimension of an array independently. For example, this allows selecting a subset of rows and/or columns from a 2-dimensional array. E.g.:

>>> z = zarr.array(np.arange(15).reshape(3, 5))
>>> z[:]
array([[ 0,  1,  2,  3,  4],
       [ 5,  6,  7,  8,  9],
       [10, 11, 12, 13, 14]])
>>> z.get_orthogonal_selection(([0, 2], slice(None)))  # select first and third rows
array([[ 0,  1,  2,  3,  4],
       [10, 11, 12, 13, 14]])
>>> z.get_orthogonal_selection((slice(None), [1, 3]))  # select second and fourth columns
array([[ 1,  3],
       [ 6,  8],
       [11, 13]])
>>> z.get_orthogonal_selection(([0, 2], [1, 3]))       # select rows [0, 2] and columns [1, 4]
array([[ 1,  3],
       [11, 13]])

Data can also be modified, e.g.:

>>> z.set_orthogonal_selection(([0, 2], [1, 3]), [[-1, -2], [-3, -4]])
>>> z[:]
array([[ 0, -1,  2, -2,  4],
       [ 5,  6,  7,  8,  9],
       [10, -3, 12, -4, 14]])

For convenience, the orthogonal indexing functionality is also available via the oindex property, e.g.:

>>> z = zarr.array(np.arange(15).reshape(3, 5))
>>> z.oindex[[0, 2], :]  # select first and third rows
array([[ 0,  1,  2,  3,  4],
       [10, 11, 12, 13, 14]])
>>> z.oindex[:, [1, 3]]  # select second and fourth columns
array([[ 1,  3],
       [ 6,  8],
       [11, 13]])
>>> z.oindex[[0, 2], [1, 3]]  # select rows [0, 2] and columns [1, 4]
array([[ 1,  3],
       [11, 13]])
>>> z.oindex[[0, 2], [1, 3]] = [[-1, -2], [-3, -4]]
>>> z[:]
array([[ 0, -1,  2, -2,  4],
       [ 5,  6,  7,  8,  9],
       [10, -3, 12, -4, 14]])

Any combination of integer, slice, 1D integer array and/or 1D Boolean array can be used for orthogonal indexing.

If the index contains at most one iterable, and otherwise contains only slices and integers, orthogonal indexing is also available directly on the array:

>>> z = zarr.array(np.arange(15).reshape(3, 5))
>>> all(z.oindex[[0, 2], :] == z[[0, 2], :])
True

Block Indexing#

As of version 2.16.0, Zarr also support block indexing, which allows selections of whole chunks based on their logical indices along each dimension of an array. For example, this allows selecting a subset of chunk aligned rows and/or columns from a 2-dimensional array. E.g.:

>>> import zarr
>>> import numpy as np
>>> z = zarr.array(np.arange(100).reshape(10, 10), chunks=(3, 3))

Retrieve items by specifying their block coordinates:

>>> z.get_block_selection(1)
array([[30, 31, 32, 33, 34, 35, 36, 37, 38, 39],
       [40, 41, 42, 43, 44, 45, 46, 47, 48, 49],
       [50, 51, 52, 53, 54, 55, 56, 57, 58, 59]])

Equivalent slicing:

>>> z[3:6]
array([[30, 31, 32, 33, 34, 35, 36, 37, 38, 39],
       [40, 41, 42, 43, 44, 45, 46, 47, 48, 49],
       [50, 51, 52, 53, 54, 55, 56, 57, 58, 59]])

For convenience, the block selection functionality is also available via the blocks property, e.g.:

>>> z.blocks[1]
array([[30, 31, 32, 33, 34, 35, 36, 37, 38, 39],
       [40, 41, 42, 43, 44, 45, 46, 47, 48, 49],
       [50, 51, 52, 53, 54, 55, 56, 57, 58, 59]])

Block index arrays may be multidimensional to index multidimensional arrays. For example:

>>> z.blocks[0, 1:3]
array([[ 3,  4,  5,  6,  7,  8],
       [13, 14, 15, 16, 17, 18],
       [23, 24, 25, 26, 27, 28]])

Data can also be modified. Let’s start by a simple 2D array:

>>> import zarr
>>> import numpy as np
>>> z = zarr.zeros((6, 6), dtype=int, chunks=2)

Set data for a selection of items:

>>> z.set_block_selection((1, 0), 1)
>>> z[...]
array([[0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [1, 1, 0, 0, 0, 0],
       [1, 1, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0],
       [0, 0, 0, 0, 0, 0]])

For convenience, this functionality is also available via the blocks property. E.g.:

>>> z.blocks[:, 2] = 7
>>> z[...]
array([[0, 0, 0, 0, 7, 7],
       [0, 0, 0, 0, 7, 7],
       [1, 1, 0, 0, 7, 7],
       [1, 1, 0, 0, 7, 7],
       [0, 0, 0, 0, 7, 7],
       [0, 0, 0, 0, 7, 7]])

Any combination of integer and slice can be used for block indexing:

>>> z.blocks[2, 1:3]
array([[0, 0, 7, 7],
       [0, 0, 7, 7]])

Indexing fields in structured arrays#

All selection methods support a fields parameter which allows retrieving or replacing data for a specific field in an array with a structured dtype. E.g.:

>>> a = np.array([(b'aaa', 1, 4.2),
...               (b'bbb', 2, 8.4),
...               (b'ccc', 3, 12.6)],
...              dtype=[('foo', 'S3'), ('bar', 'i4'), ('baz', 'f8')])
>>> z = zarr.array(a)
>>> z['foo']
array([b'aaa', b'bbb', b'ccc'],
      dtype='|S3')
>>> z['baz']
array([  4.2,   8.4,  12.6])
>>> z.get_basic_selection(slice(0, 2), fields='bar')
array([1, 2], dtype=int32)
>>> z.get_coordinate_selection([0, 2], fields=['foo', 'baz'])
array([(b'aaa',   4.2), (b'ccc',  12.6)],
      dtype=[('foo', 'S3'), ('baz', '<f8')])

Storage alternatives#

Zarr can use any object that implements the MutableMapping interface from the collections module in the Python standard library as the store for a group or an array.

Some pre-defined storage classes are provided in the zarr.storage module. For example, the zarr.storage.DirectoryStore class provides a MutableMapping interface to a directory on the local file system. This is used under the hood by the zarr.convenience.open() function. In other words, the following code:

>>> z = zarr.open('data/example.zarr', mode='w', shape=1000000, dtype='i4')

…is short-hand for:

>>> store = zarr.DirectoryStore('data/example.zarr')
>>> z = zarr.create(store=store, overwrite=True, shape=1000000, dtype='i4')

…and the following code:

>>> root = zarr.open('data/example.zarr', mode='w')

…is short-hand for:

>>> store = zarr.DirectoryStore('data/example.zarr')
>>> root = zarr.group(store=store, overwrite=True)

Any other compatible storage class could be used in place of zarr.storage.DirectoryStore in the code examples above. For example, here is an array stored directly into a ZIP archive, via the zarr.storage.ZipStore class:

>>> store = zarr.ZipStore('data/example.zip', mode='w')
>>> root = zarr.group(store=store)
>>> z = root.zeros('foo/bar', shape=(1000, 1000), chunks=(100, 100), dtype='i4')
>>> z[:] = 42
>>> store.close()

Re-open and check that data have been written:

>>> store = zarr.ZipStore('data/example.zip', mode='r')
>>> root = zarr.group(store=store)
>>> z = root['foo/bar']
>>> z[:]
array([[42, 42, 42, ..., 42, 42, 42],
       [42, 42, 42, ..., 42, 42, 42],
       [42, 42, 42, ..., 42, 42, 42],
       ...,
       [42, 42, 42, ..., 42, 42, 42],
       [42, 42, 42, ..., 42, 42, 42],
       [42, 42, 42, ..., 42, 42, 42]], dtype=int32)
>>> store.close()

Note that there are some limitations on how ZIP archives can be used, because items within a ZIP archive cannot be updated in place. This means that data in the array should only be written once and write operations should be aligned with chunk boundaries. Note also that the close() method must be called after writing any data to the store, otherwise essential records will not be written to the underlying ZIP archive.

Another storage alternative is the zarr.storage.DBMStore class, added in Zarr version 2.2. This class allows any DBM-style database to be used for storing an array or group. Here is an example using a Berkeley DB B-tree database for storage (requires bsddb3 to be installed):

>>> import bsddb3
>>> store = zarr.DBMStore('data/example.bdb', open=bsddb3.btopen)
>>> root = zarr.group(store=store, overwrite=True)
>>> z = root.zeros('foo/bar', shape=(1000, 1000), chunks=(100, 100), dtype='i4')
>>> z[:] = 42
>>> store.close()

Also added in Zarr version 2.2 is the zarr.storage.LMDBStore class which enables the lightning memory-mapped database (LMDB) to be used for storing an array or group (requires lmdb to be installed):

>>> store = zarr.LMDBStore('data/example.lmdb')
>>> root = zarr.group(store=store, overwrite=True)
>>> z = root.zeros('foo/bar', shape=(1000, 1000), chunks=(100, 100), dtype='i4')
>>> z[:] = 42
>>> store.close()

In Zarr version 2.3 is the zarr.storage.SQLiteStore class which enables the SQLite database to be used for storing an array or group (requires Python is built with SQLite support):

>>> store = zarr.SQLiteStore('data/example.sqldb')
>>> root = zarr.group(store=store, overwrite=True)
>>> z = root.zeros('foo/bar', shape=(1000, 1000), chunks=(100, 100), dtype='i4')
>>> z[:] = 42
>>> store.close()

Also added in Zarr version 2.3 are two storage classes for interfacing with server-client databases. The zarr.storage.RedisStore class interfaces Redis (an in memory data structure store), and the zarr.storage.MongoDB class interfaces with MongoDB (an object oriented NoSQL database). These stores respectively require the redis-py and pymongo packages to be installed.

For compatibility with the N5 data format, Zarr also provides an N5 backend (this is currently an experimental feature). Similar to the ZIP storage class, an zarr.n5.N5Store can be instantiated directly:

>>> store = zarr.N5Store('data/example.n5')
>>> root = zarr.group(store=store)
>>> z = root.zeros('foo/bar', shape=(1000, 1000), chunks=(100, 100), dtype='i4')
>>> z[:] = 42

For convenience, the N5 backend will automatically be chosen when the filename ends with .n5:

>>> root = zarr.open('data/example.n5', mode='w')

Distributed/cloud storage#

It is also possible to use distributed storage systems. The Dask project has implementations of the MutableMapping interface for Amazon S3 (S3Map), Hadoop Distributed File System (HDFSMap) and Google Cloud Storage (GCSMap), which can be used with Zarr.

Here is an example using S3Map to read an array created previously:

>>> import s3fs
>>> import zarr
>>> s3 = s3fs.S3FileSystem(anon=True, client_kwargs=dict(region_name='eu-west-2'))
>>> store = s3fs.S3Map(root='zarr-demo/store', s3=s3, check=False)
>>> root = zarr.group(store=store)
>>> z = root['foo/bar/baz']
>>> z
<zarr.core.Array '/foo/bar/baz' (21,) |S1>
>>> z.info
Name               : /foo/bar/baz
Type               : zarr.core.Array
Data type          : |S1
Shape              : (21,)
Chunk shape        : (7,)
Order              : C
Read-only          : False
Compressor         : Blosc(cname='lz4', clevel=5, shuffle=SHUFFLE, blocksize=0)
Store type         : zarr.storage.KVStore
No. bytes          : 21
No. bytes stored   : 382
Storage ratio      : 0.1
Chunks initialized : 3/3
>>> z[:]
array([b'H', b'e', b'l', b'l', b'o', b' ', b'f', b'r', b'o', b'm', b' ',
       b't', b'h', b'e', b' ', b'c', b'l', b'o', b'u', b'd', b'!'],
      dtype='|S1')
>>> z[:].tobytes()
b'Hello from the cloud!'

Zarr now also has a builtin storage backend for Azure Blob Storage. The class is zarr.storage.ABSStore (requires azure-storage-blob to be installed):

>>> import azure.storage.blob
>>> container_client = azure.storage.blob.ContainerClient(...)  
>>> store = zarr.ABSStore(client=container_client, prefix='zarr-testing')  
>>> root = zarr.group(store=store, overwrite=True)  
>>> z = root.zeros('foo/bar', shape=(1000, 1000), chunks=(100, 100), dtype='i4')  
>>> z[:] = 42  

When using an actual storage account, provide account_name and account_key arguments to zarr.storage.ABSStore, the above client is just testing against the emulator. Please also note that this is an experimental feature.

Note that retrieving data from a remote service via the network can be significantly slower than retrieving data from a local file system, and will depend on network latency and bandwidth between the client and server systems. If you are experiencing poor performance, there are several things you can try. One option is to increase the array chunk size, which will reduce the number of chunks and thus reduce the number of network round-trips required to retrieve data for an array (and thus reduce the impact of network latency). Another option is to try to increase the compression ratio by changing compression options or trying a different compressor (which will reduce the impact of limited network bandwidth).

As of version 2.2, Zarr also provides the zarr.storage.LRUStoreCache which can be used to implement a local in-memory cache layer over a remote store. E.g.:

>>> s3 = s3fs.S3FileSystem(anon=True, client_kwargs=dict(region_name='eu-west-2'))
>>> store = s3fs.S3Map(root='zarr-demo/store', s3=s3, check=False)
>>> cache = zarr.LRUStoreCache(store, max_size=2**28)
>>> root = zarr.group(store=cache)
>>> z = root['foo/bar/baz']
>>> from timeit import timeit
>>> # first data access is relatively slow, retrieved from store
... timeit('print(z[:].tobytes())', number=1, globals=globals())  
b'Hello from the cloud!'
0.1081731989979744
>>> # second data access is faster, uses cache
... timeit('print(z[:].tobytes())', number=1, globals=globals())  
b'Hello from the cloud!'
0.0009490990014455747

If you are still experiencing poor performance with distributed/cloud storage, please raise an issue on the GitHub issue tracker with any profiling data you can provide, as there may be opportunities to optimise further either within Zarr or within the mapping interface to the storage.

IO with fsspec#

As of version 2.5, zarr supports passing URLs directly to fsspec, and having it create the “mapping” instance automatically. This means, that for all of the backend storage implementations supported by fsspec, you can skip importing and configuring the storage explicitly. For example:

>>> g = zarr.open_group("s3://zarr-demo/store", storage_options={'anon': True})   
>>> g['foo/bar/baz'][:].tobytes()  
b'Hello from the cloud!'

The provision of the protocol specifier “s3://” will select the correct backend. Notice the kwargs storage_options, used to pass parameters to that backend.

As of version 2.6, write mode and complex URLs are also supported, such as:

>>> g = zarr.open_group("simplecache::s3://zarr-demo/store",
...                     storage_options={"s3": {'anon': True}})  
>>> g['foo/bar/baz'][:].tobytes()  # downloads target file  
b'Hello from the cloud!'
>>> g['foo/bar/baz'][:].tobytes()  # uses cached file  
b'Hello from the cloud!'

The second invocation here will be much faster. Note that the storage_options have become more complex here, to account for the two parts of the supplied URL.

It is also possible to initialize the filesystem outside of Zarr and then pass it through. This requires creating an zarr.storage.FSStore object explicitly. For example:

>>> import s3fs  
>>> fs = s3fs.S3FileSystem(anon=True)  
>>> store = zarr.storage.FSStore('/zarr-demo/store', fs=fs)  
>>> g = zarr.open_group(store)  

This is useful in cases where you want to also use the same fsspec filesystem object separately from Zarr.

Accessing ZIP archives on S3#

The built-in zarr.storage.ZipStore will only work with paths on the local file-system; however it is possible to access ZIP-archived Zarr data on the cloud via the ZipFileSystem class from fsspec. The following example demonstrates how to access a ZIP-archived Zarr group on s3 using s3fs and ZipFileSystem:

>>> s3_path = "s3://path/to/my.zarr.zip"
>>>
>>> s3 = s3fs.S3FileSystem()
>>> f = s3.open(s3_path)
>>> fs = ZipFileSystem(f, mode="r")
>>> store = FSMap("", fs, check=False)
>>>
>>> # caching may improve performance when repeatedly reading the same data
>>> cache = zarr.storage.LRUStoreCache(store, max_size=2**28)
>>> z = zarr.group(store=cache)

This store can also be generated with fsspec’s handler chaining, like so:

>>> store = zarr.storage.FSStore(url=f"zip::{s3_path}",  mode="r")

This can be especially useful if you have a very large ZIP-archived Zarr array or group on s3 and only need to access a small portion of it.

Consolidating metadata#

Since there is a significant overhead for every connection to a cloud object store such as S3, the pattern described in the previous section may incur significant latency while scanning the metadata of the array hierarchy, even though each individual metadata object is small. For cases such as these, once the data are static and can be regarded as read-only, at least for the metadata/structure of the array hierarchy, the many metadata objects can be consolidated into a single one via zarr.convenience.consolidate_metadata(). Doing this can greatly increase the speed of reading the array metadata, e.g.:

>>> zarr.consolidate_metadata(store)  

This creates a special key with a copy of all of the metadata from all of the metadata objects in the store.

Later, to open a Zarr store with consolidated metadata, use zarr.convenience.open_consolidated(), e.g.:

>>> root = zarr.open_consolidated(store)  

This uses the special key to read all of the metadata in a single call to the backend storage.

Note that, the hierarchy could still be opened in the normal way and altered, causing the consolidated metadata to become out of sync with the real state of the array hierarchy. In this case, zarr.convenience.consolidate_metadata() would need to be called again.

To protect against consolidated metadata accidentally getting out of sync, the root group returned by zarr.convenience.open_consolidated() is read-only for the metadata, meaning that no new groups or arrays can be created, and arrays cannot be resized. However, data values with arrays can still be updated.

Copying/migrating data#

If you have some data in an HDF5 file and would like to copy some or all of it into a Zarr group, or vice-versa, the zarr.convenience.copy() and zarr.convenience.copy_all() functions can be used. Here’s an example copying a group named ‘foo’ from an HDF5 file to a Zarr group:

>>> import h5py
>>> import zarr
>>> import numpy as np
>>> source = h5py.File('data/example.h5', mode='w')
>>> foo = source.create_group('foo')
>>> baz = foo.create_dataset('bar/baz', data=np.arange(100), chunks=(50,))
>>> spam = source.create_dataset('spam', data=np.arange(100, 200), chunks=(30,))
>>> zarr.tree(source)
/
 ├── foo
 │   └── bar
 │       └── baz (100,) int64
 └── spam (100,) int64
>>> dest = zarr.open_group('data/example.zarr', mode='w')
>>> from sys import stdout
>>> zarr.copy(source['foo'], dest, log=stdout)
copy /foo
copy /foo/bar
copy /foo/bar/baz (100,) int64
all done: 3 copied, 0 skipped, 800 bytes copied
(3, 0, 800)
>>> dest.tree()  # N.B., no spam
/
 └── foo
     └── bar
         └── baz (100,) int64
>>> source.close()

If rather than copying a single group or array you would like to copy all groups and arrays, use zarr.convenience.copy_all(), e.g.:

>>> source = h5py.File('data/example.h5', mode='r')
>>> dest = zarr.open_group('data/example2.zarr', mode='w')
>>> zarr.copy_all(source, dest, log=stdout)
copy /foo
copy /foo/bar
copy /foo/bar/baz (100,) int64
copy /spam (100,) int64
all done: 4 copied, 0 skipped, 1,600 bytes copied
(4, 0, 1600)
>>> dest.tree()
/
 ├── foo
 │   └── bar
 │       └── baz (100,) int64
 └── spam (100,) int64

If you need to copy data between two Zarr groups, the zarr.convenience.copy() and zarr.convenience.copy_all() functions can be used and provide the most flexibility. However, if you want to copy data in the most efficient way possible, without changing any configuration options, the zarr.convenience.copy_store() function can be used. This function copies data directly between the underlying stores, without any decompression or re-compression, and so should be faster. E.g.:

>>> import zarr
>>> import numpy as np
>>> store1 = zarr.DirectoryStore('data/example.zarr')
>>> root = zarr.group(store1, overwrite=True)
>>> baz = root.create_dataset('foo/bar/baz', data=np.arange(100), chunks=(50,))
>>> spam = root.create_dataset('spam', data=np.arange(100, 200), chunks=(30,))
>>> root.tree()
/
 ├── foo
 │   └── bar
 │       └── baz (100,) int64
 └── spam (100,) int64
>>> from sys import stdout
>>> store2 = zarr.ZipStore('data/example.zip', mode='w')
>>> zarr.copy_store(store1, store2, log=stdout)
copy .zgroup
copy foo/.zgroup
copy foo/bar/.zgroup
copy foo/bar/baz/.zarray
copy foo/bar/baz/0
copy foo/bar/baz/1
copy spam/.zarray
copy spam/0
copy spam/1
copy spam/2
copy spam/3
all done: 11 copied, 0 skipped, 1,138 bytes copied
(11, 0, 1138)
>>> new_root = zarr.group(store2)
>>> new_root.tree()
/
 ├── foo
 │   └── bar
 │       └── baz (100,) int64
 └── spam (100,) int64
>>> new_root['foo/bar/baz'][:]
array([ 0,  1,  2,  ..., 97, 98, 99])
>>> store2.close()  # ZIP stores need to be closed

String arrays#

There are several options for storing arrays of strings.

If your strings are all ASCII strings, and you know the maximum length of the string in your array, then you can use an array with a fixed-length bytes dtype. E.g.:

>>> z = zarr.zeros(10, dtype='S6')
>>> z
<zarr.core.Array (10,) |S6>
>>> z[0] = b'Hello'
>>> z[1] = b'world!'
>>> z[:]
array([b'Hello', b'world!', b'', b'', b'', b'', b'', b'', b'', b''],
      dtype='|S6')

A fixed-length unicode dtype is also available, e.g.:

>>> greetings = ['¡Hola mundo!', 'Hej Världen!', 'Servus Woid!', 'Hei maailma!',
...              'Xin chào thế giới', 'Njatjeta Botë!', 'Γεια σου κόσμε!',
...              'こんにちは世界', '世界,你好!', 'Helló, világ!', 'Zdravo svete!',
...              'เฮลโลเวิลด์']
>>> text_data = greetings * 10000
>>> z = zarr.array(text_data, dtype='U20')
>>> z
<zarr.core.Array (120000,) <U20>
>>> z[:]
array(['¡Hola mundo!', 'Hej Världen!', 'Servus Woid!', ...,
       'Helló, világ!', 'Zdravo svete!', 'เฮลโลเวิลด์'],
      dtype='<U20')

For variable-length strings, the object dtype can be used, but a codec must be provided to encode the data (see also Object arrays below). At the time of writing there are four codecs available that can encode variable length string objects: numcodecs.vlen.VLenUTF8, numcodecs.json.JSON, numcodecs.msgpacks.MsgPack. and numcodecs.pickles.Pickle. E.g. using VLenUTF8:

>>> import numcodecs
>>> z = zarr.array(text_data, dtype=object, object_codec=numcodecs.VLenUTF8())
>>> z
<zarr.core.Array (120000,) object>
>>> z.filters
[VLenUTF8()]
>>> z[:]
array(['¡Hola mundo!', 'Hej Världen!', 'Servus Woid!', ...,
       'Helló, világ!', 'Zdravo svete!', 'เฮลโลเวิลด์'], dtype=object)

As a convenience, dtype=str (or dtype=unicode on Python 2.7) can be used, which is a short-hand for dtype=object, object_codec=numcodecs.VLenUTF8(), e.g.:

>>> z = zarr.array(text_data, dtype=str)
>>> z
<zarr.core.Array (120000,) object>
>>> z.filters
[VLenUTF8()]
>>> z[:]
array(['¡Hola mundo!', 'Hej Världen!', 'Servus Woid!', ...,
       'Helló, világ!', 'Zdravo svete!', 'เฮลโลเวิลด์'], dtype=object)

Variable-length byte strings are also supported via dtype=object. Again an object_codec is required, which can be one of numcodecs.vlen.VLenBytes or numcodecs.pickles.Pickle. For convenience, dtype=bytes (or dtype=str on Python 2.7) can be used as a short-hand for dtype=object, object_codec=numcodecs.VLenBytes(), e.g.:

>>> bytes_data = [g.encode('utf-8') for g in greetings] * 10000
>>> z = zarr.array(bytes_data, dtype=bytes)
>>> z
<zarr.core.Array (120000,) object>
>>> z.filters
[VLenBytes()]
>>> z[:]
array([b'\xc2\xa1Hola mundo!', b'Hej V\xc3\xa4rlden!', b'Servus Woid!',
       ..., b'Hell\xc3\xb3, vil\xc3\xa1g!', b'Zdravo svete!',
       b'\xe0\xb9\x80\xe0\xb8\xae\xe0\xb8\xa5\xe0\xb9\x82\xe0\xb8\xa5\xe0\xb9\x80\xe0\xb8\xa7\xe0\xb8\xb4\xe0\xb8\xa5\xe0\xb8\x94\xe0\xb9\x8c'], dtype=object)

If you know ahead of time all the possible string values that can occur, you could also use the numcodecs.categorize.Categorize codec to encode each unique string value as an integer. E.g.:

>>> categorize = numcodecs.Categorize(greetings, dtype=object)
>>> z = zarr.array(text_data, dtype=object, object_codec=categorize)
>>> z
<zarr.core.Array (120000,) object>
>>> z.filters
[Categorize(dtype='|O', astype='|u1', labels=['¡Hola mundo!', 'Hej Världen!', 'Servus Woid!', ...])]
>>> z[:]
array(['¡Hola mundo!', 'Hej Världen!', 'Servus Woid!', ...,
       'Helló, világ!', 'Zdravo svete!', 'เฮลโลเวิลด์'], dtype=object)

Object arrays#

Zarr supports arrays with an “object” dtype. This allows arrays to contain any type of object, such as variable length unicode strings, or variable length arrays of numbers, or other possibilities. When creating an object array, a codec must be provided via the object_codec argument. This codec handles encoding (serialization) of Python objects. The best codec to use will depend on what type of objects are present in the array.

At the time of writing there are three codecs available that can serve as a general purpose object codec and support encoding of a mixture of object types: numcodecs.json.JSON, numcodecs.msgpacks.MsgPack. and numcodecs.pickles.Pickle.

For example, using the JSON codec:

>>> z = zarr.empty(5, dtype=object, object_codec=numcodecs.JSON())
>>> z[0] = 42
>>> z[1] = 'foo'
>>> z[2] = ['bar', 'baz', 'qux']
>>> z[3] = {'a': 1, 'b': 2.2}
>>> z[:]
array([42, 'foo', list(['bar', 'baz', 'qux']), {'a': 1, 'b': 2.2}, None], dtype=object)

Not all codecs support encoding of all object types. The numcodecs.pickles.Pickle codec is the most flexible, supporting encoding any type of Python object. However, if you are sharing data with anyone other than yourself, then Pickle is not recommended as it is a potential security risk. This is because malicious code can be embedded within pickled data. The JSON and MsgPack codecs do not have any security issues and support encoding of unicode strings, lists and dictionaries. MsgPack is usually faster for both encoding and decoding.

Ragged arrays#

If you need to store an array of arrays, where each member array can be of any length and stores the same primitive type (a.k.a. a ragged array), the numcodecs.vlen.VLenArray codec can be used, e.g.:

>>> z = zarr.empty(4, dtype=object, object_codec=numcodecs.VLenArray(int))
>>> z
<zarr.core.Array (4,) object>
>>> z.filters
[VLenArray(dtype='<i8')]
>>> z[0] = np.array([1, 3, 5])
>>> z[1] = np.array([4])
>>> z[2] = np.array([7, 9, 14])
>>> z[:]
array([array([1, 3, 5]), array([4]), array([ 7,  9, 14]),
       array([], dtype=int64)], dtype=object)

As a convenience, dtype='array:T' can be used as a short-hand for dtype=object, object_codec=numcodecs.VLenArray('T'), where ‘T’ can be any NumPy primitive dtype such as ‘i4’ or ‘f8’. E.g.:

>>> z = zarr.empty(4, dtype='array:i8')
>>> z
<zarr.core.Array (4,) object>
>>> z.filters
[VLenArray(dtype='<i8')]
>>> z[0] = np.array([1, 3, 5])
>>> z[1] = np.array([4])
>>> z[2] = np.array([7, 9, 14])
>>> z[:]
array([array([1, 3, 5]), array([4]), array([ 7,  9, 14]),
       array([], dtype=int64)], dtype=object)

Chunk optimizations#

Chunk size and shape#

In general, chunks of at least 1 megabyte (1M) uncompressed size seem to provide better performance, at least when using the Blosc compression library.

The optimal chunk shape will depend on how you want to access the data. E.g., for a 2-dimensional array, if you only ever take slices along the first dimension, then chunk across the second dimension. If you know you want to chunk across an entire dimension you can use None or -1 within the chunks argument, e.g.:

>>> z1 = zarr.zeros((10000, 10000), chunks=(100, None), dtype='i4')
>>> z1.chunks
(100, 10000)

Alternatively, if you only ever take slices along the second dimension, then chunk across the first dimension, e.g.:

>>> z2 = zarr.zeros((10000, 10000), chunks=(None, 100), dtype='i4')
>>> z2.chunks
(10000, 100)

If you require reasonable performance for both access patterns then you need to find a compromise, e.g.:

>>> z3 = zarr.zeros((10000, 10000), chunks=(1000, 1000), dtype='i4')
>>> z3.chunks
(1000, 1000)

If you are feeling lazy, you can let Zarr guess a chunk shape for your data by providing chunks=True, although please note that the algorithm for guessing a chunk shape is based on simple heuristics and may be far from optimal. E.g.:

>>> z4 = zarr.zeros((10000, 10000), chunks=True, dtype='i4')
>>> z4.chunks
(625, 625)

If you know you are always going to be loading the entire array into memory, you can turn off chunks by providing chunks=False, in which case there will be one single chunk for the array:

>>> z5 = zarr.zeros((10000, 10000), chunks=False, dtype='i4')
>>> z5.chunks
(10000, 10000)

Chunk memory layout#

The order of bytes within each chunk of an array can be changed via the order keyword argument, to use either C or Fortran layout. For multi-dimensional arrays, these two layouts may provide different compression ratios, depending on the correlation structure within the data. E.g.:

>>> a = np.arange(100000000, dtype='i4').reshape(10000, 10000).T
>>> c = zarr.array(a, chunks=(1000, 1000))
>>> c.info
Type               : zarr.core.Array
Data type          : int32
Shape              : (10000, 10000)
Chunk shape        : (1000, 1000)
Order              : C
Read-only          : False
Compressor         : Blosc(cname='lz4', clevel=5, shuffle=SHUFFLE, blocksize=0)
Store type         : zarr.storage.KVStore
No. bytes          : 400000000 (381.5M)
No. bytes stored   : 6696010 (6.4M)
Storage ratio      : 59.7
Chunks initialized : 100/100
>>> f = zarr.array(a, chunks=(1000, 1000), order='F')
>>> f.info
Type               : zarr.core.Array
Data type          : int32
Shape              : (10000, 10000)
Chunk shape        : (1000, 1000)
Order              : F
Read-only          : False
Compressor         : Blosc(cname='lz4', clevel=5, shuffle=SHUFFLE, blocksize=0)
Store type         : zarr.storage.KVStore
No. bytes          : 400000000 (381.5M)
No. bytes stored   : 4684636 (4.5M)
Storage ratio      : 85.4
Chunks initialized : 100/100

In the above example, Fortran order gives a better compression ratio. This is an artificial example but illustrates the general point that changing the order of bytes within chunks of an array may improve the compression ratio, depending on the structure of the data, the compression algorithm used, and which compression filters (e.g., byte-shuffle) have been applied.

Empty chunks#

As of version 2.11, it is possible to configure how Zarr handles the storage of chunks that are “empty” (i.e., every element in the chunk is equal to the array’s fill value). When creating an array with write_empty_chunks=False, Zarr will check whether a chunk is empty before compression and storage. If a chunk is empty, then Zarr does not store it, and instead deletes the chunk from storage if the chunk had been previously stored.

This optimization prevents storing redundant objects and can speed up reads, but the cost is added computation during array writes, since the contents of each chunk must be compared to the fill value, and these advantages are contingent on the content of the array. If you know that your data will form chunks that are almost always non-empty, then there is no advantage to the optimization described above. In this case, creating an array with write_empty_chunks=True (the default) will instruct Zarr to write every chunk without checking for emptiness.

The following example illustrates the effect of the write_empty_chunks flag on the time required to write an array with different values.:

>>> import zarr
>>> import numpy as np
>>> import time
>>> from tempfile import TemporaryDirectory
>>> def timed_write(write_empty_chunks):
...     """
...     Measure the time required and number of objects created when writing
...     to a Zarr array with random ints or fill value.
...     """
...     chunks = (8192,)
...     shape = (chunks[0] * 1024,)
...     data = np.random.randint(0, 255, shape)
...     dtype = 'uint8'
...
...     with TemporaryDirectory() as store:
...         arr = zarr.open(store,
...                         shape=shape,
...                         chunks=chunks,
...                         dtype=dtype,
...                         write_empty_chunks=write_empty_chunks,
...                         fill_value=0,
...                         mode='w')
...         # initialize all chunks
...         arr[:] = 100
...         result = []
...         for value in (data, arr.fill_value):
...             start = time.time()
...             arr[:] = value
...             elapsed = time.time() - start
...             result.append((elapsed, arr.nchunks_initialized))
...
...         return result
>>> for write_empty_chunks in (True, False):
...     full, empty = timed_write(write_empty_chunks)
...     print(f'\nwrite_empty_chunks={write_empty_chunks}:\n\tRandom Data: {full[0]:.4f}s, {full[1]} objects stored\n\t Empty Data: {empty[0]:.4f}s, {empty[1]} objects stored\n')

write_empty_chunks=True:
        Random Data: 0.1252s, 1024 objects stored
        Empty Data: 0.1060s, 1024 objects stored


write_empty_chunks=False:
        Random Data: 0.1359s, 1024 objects stored
        Empty Data: 0.0301s, 0 objects stored

In this example, writing random data is slightly slower with write_empty_chunks=True, but writing empty data is substantially faster and generates far fewer objects in storage.

Changing chunk shapes (rechunking)#

Sometimes you are not free to choose the initial chunking of your input data, or you might have data saved with chunking which is not optimal for the analysis you have planned. In such cases it can be advantageous to re-chunk the data. For small datasets, or when the mismatch between input and output chunks is small such that only a few chunks of the input dataset need to be read to create each chunk in the output array, it is sufficient to simply copy the data to a new array with the desired chunking, e.g.

>>> a = zarr.zeros((10000, 10000), chunks=(100,100), dtype='uint16', store='a.zarr')
>>> b = zarr.array(a, chunks=(100, 200), store='b.zarr')

If the chunk shapes mismatch, however, a simple copy can lead to non-optimal data access patterns and incur a substantial performance hit when using file based stores. One of the most pathological examples is switching from column-based chunking to row-based chunking e.g.

>>> a = zarr.zeros((10000,10000), chunks=(10000, 1), dtype='uint16', store='a.zarr')
>>> b = zarr.array(a, chunks=(1,10000), store='b.zarr')

which will require every chunk in the input data set to be repeatedly read when creating each output chunk. If the entire array will fit within memory, this is simply resolved by forcing the entire input array into memory as a numpy array before converting back to zarr with the desired chunking.

>>> a = zarr.zeros((10000,10000), chunks=(10000, 1), dtype='uint16', store='a.zarr')
>>> b = a[...]
>>> c = zarr.array(b, chunks=(1,10000), store='c.zarr')

For data sets which have mismatched chunks and which do not fit in memory, a more sophisticated approach to rechunking, such as offered by the rechunker package and discussed here may offer a substantial improvement in performance.

Parallel computing and synchronization#

Zarr arrays have been designed for use as the source or sink for data in parallel computations. By data source we mean that multiple concurrent read operations may occur. By data sink we mean that multiple concurrent write operations may occur, with each writer updating a different region of the array. Zarr arrays have not been designed for situations where multiple readers and writers are concurrently operating on the same array.

Both multi-threaded and multi-process parallelism are possible. The bottleneck for most storage and retrieval operations is compression/decompression, and the Python global interpreter lock (GIL) is released wherever possible during these operations, so Zarr will generally not block other Python threads from running.

When using a Zarr array as a data sink, some synchronization (locking) may be required to avoid data loss, depending on how data are being updated. If each worker in a parallel computation is writing to a separate region of the array, and if region boundaries are perfectly aligned with chunk boundaries, then no synchronization is required. However, if region and chunk boundaries are not perfectly aligned, then synchronization is required to avoid two workers attempting to modify the same chunk at the same time, which could result in data loss.

To give a simple example, consider a 1-dimensional array of length 60, z, divided into three chunks of 20 elements each. If three workers are running and each attempts to write to a 20 element region (i.e., z[0:20], z[20:40] and z[40:60]) then each worker will be writing to a separate chunk and no synchronization is required. However, if two workers are running and each attempts to write to a 30 element region (i.e., z[0:30] and z[30:60]) then it is possible both workers will attempt to modify the middle chunk at the same time, and synchronization is required to prevent data loss.

Zarr provides support for chunk-level synchronization. E.g., create an array with thread synchronization:

>>> z = zarr.zeros((10000, 10000), chunks=(1000, 1000), dtype='i4',
...                 synchronizer=zarr.ThreadSynchronizer())
>>> z
<zarr.core.Array (10000, 10000) int32>

This array is safe to read or write within a multi-threaded program.

Zarr also provides support for process synchronization via file locking, provided that all processes have access to a shared file system, and provided that the underlying file system supports file locking (which is not the case for some networked file systems). E.g.:

>>> synchronizer = zarr.ProcessSynchronizer('data/example.sync')
>>> z = zarr.open_array('data/example', mode='w', shape=(10000, 10000),
...                     chunks=(1000, 1000), dtype='i4',
...                     synchronizer=synchronizer)
>>> z
<zarr.core.Array (10000, 10000) int32>

This array is safe to read or write from multiple processes.

When using multiple processes to parallelize reads or writes on arrays using the Blosc compression library, it may be necessary to set numcodecs.blosc.use_threads = False, as otherwise Blosc may share incorrect global state amongst processes causing programs to hang. See also the section on Configuring Blosc below.

Please note that support for parallel computing is an area of ongoing research and development. If you are using Zarr for parallel computing, we welcome feedback, experience, discussion, ideas and advice, particularly about issues related to data integrity and performance.

Pickle support#

Zarr arrays and groups can be pickled, as long as the underlying store object can be pickled. Instances of any of the storage classes provided in the zarr.storage module can be pickled, as can the built-in dict class which can also be used for storage.

Note that if an array or group is backed by an in-memory store like a dict or zarr.storage.MemoryStore, then when it is pickled all of the store data will be included in the pickled data. However, if an array or group is backed by a persistent store like a zarr.storage.DirectoryStore, zarr.storage.ZipStore or zarr.storage.DBMStore then the store data are not pickled. The only thing that is pickled is the necessary parameters to allow the store to re-open any underlying files or databases upon being unpickled.

E.g., pickle/unpickle an in-memory array:

>>> import pickle
>>> z1 = zarr.array(np.arange(100000))
>>> s = pickle.dumps(z1)
>>> len(s) > 5000  # relatively large because data have been pickled
True
>>> z2 = pickle.loads(s)
>>> z1 == z2
True
>>> np.all(z1[:] == z2[:])
True

E.g., pickle/unpickle an array stored on disk:

>>> z3 = zarr.open('data/walnuts.zarr', mode='w', shape=100000, dtype='i8')
>>> z3[:] = np.arange(100000)
>>> s = pickle.dumps(z3)
>>> len(s) < 200  # small because no data have been pickled
True
>>> z4 = pickle.loads(s)
>>> z3 == z4
True
>>> np.all(z3[:] == z4[:])
True

Datetimes and timedeltas#

NumPy’s datetime64 (‘M8’) and timedelta64 (‘m8’) dtypes are supported for Zarr arrays, as long as the units are specified. E.g.:

>>> z = zarr.array(['2007-07-13', '2006-01-13', '2010-08-13'], dtype='M8[D]')
>>> z
<zarr.core.Array (3,) datetime64[D]>
>>> z[:]
array(['2007-07-13', '2006-01-13', '2010-08-13'], dtype='datetime64[D]')
>>> z[0]
numpy.datetime64('2007-07-13')
>>> z[0] = '1999-12-31'
>>> z[:]
array(['1999-12-31', '2006-01-13', '2010-08-13'], dtype='datetime64[D]')

Usage tips#

Copying large arrays#

Data can be copied between large arrays without needing much memory, e.g.:

>>> z1 = zarr.empty((10000, 10000), chunks=(1000, 1000), dtype='i4')
>>> z1[:] = 42
>>> z2 = zarr.empty_like(z1)
>>> z2[:] = z1

Internally the example above works chunk-by-chunk, extracting only the data from z1 required to fill each chunk in z2. The source of the data (z1) could equally be an h5py Dataset.

Configuring Blosc#

The Blosc compressor is able to use multiple threads internally to accelerate compression and decompression. By default, Blosc uses up to 8 internal threads. The number of Blosc threads can be changed to increase or decrease this number, e.g.:

>>> from numcodecs import blosc
>>> blosc.set_nthreads(2)  
8

When a Zarr array is being used within a multi-threaded program, Zarr automatically switches to using Blosc in a single-threaded “contextual” mode. This is generally better as it allows multiple program threads to use Blosc simultaneously and prevents CPU thrashing from too many active threads. If you want to manually override this behaviour, set the value of the blosc.use_threads variable to True (Blosc always uses multiple internal threads) or False (Blosc always runs in single-threaded contextual mode). To re-enable automatic switching, set blosc.use_threads to None.

Please note that if Zarr is being used within a multi-process program, Blosc may not be safe to use in multi-threaded mode and may cause the program to hang. If using Blosc in a multi-process program then it is recommended to set blosc.use_threads = False.