Internal Design
===============
Overview
--------
.. image:: images/array.svg
:width: 40 %
:align: right
:alt: 12 rectangular blocks arranged as a 4-row, 3-column layout. Each block includes 'x' and its location in the table starting with ('x',0,0) in the top-left, and a size of 5x8.
Dask arrays define a large array with a grid of blocks of smaller arrays.
These arrays may be actual arrays or functions that produce arrays. We
define a Dask array with the following components:
* A Dask graph with a special set of keys designating blocks
such as ``('x', 0, 0), ('x', 0, 1), ...`` (See :doc:`Dask graph
documentation ` for more details)
* A sequence of chunk sizes along each dimension called ``chunks``,
for example ``((5, 5, 5, 5), (8, 8, 8))``
* A name to identify which keys in the Dask graph refer to this array, like
``'x'``
* A NumPy dtype
Example
~~~~~~~
.. code-block:: python
>>> import dask.array as da
>>> x = da.arange(0, 15, chunks=(5,))
>>> x.name
'arange-539766a'
>>> x.__dask_graph__()
>>> dict(x.__dask_graph__()) # somewhat simplified
{('arange-539766a', 0): (np.arange, 0, 5),
('arange-539766a', 1): (np.arange, 5, 10),
('arange-539766a', 2): (np.arange, 10, 15)}
>>> x.chunks
((5, 5, 5),)
>>> x.dtype
dtype('int64')
Keys of the Dask graph
----------------------
By special convention, we refer to each block of the array with a tuple of the
form ``(name, i, j, k)``, with ``i, j, k`` being the indices of the block
ranging from ``0`` to the number of blocks in that dimension. The Dask graph
must hold key-value pairs referring to these keys. Moreover, it likely also
holds other key-value pairs required to eventually compute the desired values
(usually organised in a :doc:`HighLevelGraph `, but shown
in a flattened form here for illustration):
.. code-block:: python
{
('x', 0, 0): (add, 1, ('y', 0, 0)),
('x', 0, 1): (add, 1, ('y', 0, 1)),
...
('y', 0, 0): (getitem, dataset, (slice(0, 1000), slice(0, 1000))),
('y', 0, 1): (getitem, dataset, (slice(0, 1000), slice(1000, 2000)))
...
}
The name of an ``Array`` object can be found in the ``name`` attribute. One
can get a nested list of keys with the ``.__dask_keys__()`` method. Additionally,
one can flatten down this list with ``dask.array.core.flatten()``. This is sometimes
useful when building new dictionaries.
Chunks
------
We also store the size of each block along each axis. This is composed of
a tuple of tuples such that the length of the outer tuple is equal to the
number of dimensions of the array, and the lengths of the inner tuples are
equal to the number of blocks along each dimension. In the example illustrated
above this value is as follows::
chunks = ((5, 5, 5, 5), (8, 8, 8))
Note that these numbers do not necessarily need to be regular. We often create
regularly sized grids but blocks change shape after complex slicing. Beware
that some operations do expect certain symmetries in the block-shapes. For
example, matrix multiplication requires that blocks on each side have
anti-symmetric shapes.
Some ways in which ``chunks`` reflects properties of our array:
1. ``len(x.chunks) == x.ndim``: the length of chunks is the number of dimensions
2. ``tuple(map(sum, x.chunks)) == x.shape``: the sum of each internal chunk is the
length of that dimension
3. The length of each internal chunk is the number of keys in that dimension.
For instance, for ``chunks == ((a, b), (d, e, f))`` and name == ``'x'``
our array has tasks with the following keys::
('x', 0, 0), ('x', 0, 1), ('x', 0, 2)
('x', 1, 0), ('x', 1, 1), ('x', 1, 2)
Metadata
--------
Many Array operations rely on knowing the dtype (int, float,..) and type (numpy, cupy,...). To
keep track of this information, all Dask Array objects have a ``_meta``
attribute which contains an empty Numpy object with the same dtypes.
For example:
.. code-block:: python
>>> np_array = np.arange(15).reshape(3, 5)
>>> da_array = da.from_array(np_array, npartitions=2)
>>> da_array._meta
Empty Array
Shape: (0, 0)
dtype: int64
array([], shape=(0, 0), dtype=int64)
>>> ddf._meta.dtype
dtype: int64
Internally, Dask Array does its best to propagate this information
through all operations, so most of the time a user shouldn't have to worry
about this.
Create an Array Object
----------------------
In order to create an ``da.Array`` object we need a graph with these special
keys::
layer = {('x', 0, 0): ...}
dsk = HighLevelGraph.from_collections('x', layer, dependencies=())
a name specifying which keys this array refers to::
name = 'x'
and a chunks tuple::
chunks = ((5, 5, 5, 5), (8, 8, 8))
Then, using these elements, one can construct an array::
x = da.Array(dsk, name, chunks)
In short, ``dask.array`` operations update Dask graphs, update dtypes, and track chunk
shapes.
Example - ``eye`` function
--------------------------
As an example, let's build the ``np.eye`` function for ``dask.array`` to make the
identity matrix:
.. code-block:: python
def eye(n, blocksize):
chunks = ((blocksize,) * (n // blocksize),
(blocksize,) * (n // blocksize))
name = 'eye' + next(tokens) # unique identifier
layer = {(name, i, j): (np.eye, blocksize)
if i == j else
(np.zeros, (blocksize, blocksize))
for i in range(n // blocksize)
for j in range(n // blocksize)}
dsk = dask.highlevelgraph.HighLevelGraph.from_collections(name, layer, dependencies=())
dtype = np.eye(0).dtype # take dtype default from numpy
return dask.array.Array(dsk, name, chunks, dtype)