Dask

Dask provides multi-core execution on larger-than-memory datasets using blocked algorithms and task scheduling. It maps high-level NumPy and list operations on large datasets on to graphs of many operations on small in-memory datasets. It then executes these graphs in parallel on a single machine. Dask lets us use traditional NumPy and list programming while operating on inconveniently large data in a small amount of space.

dask is a specification to describe task dependency graphs.
dask.array is a drop-in NumPy replacement (for a subset of NumPy) that encodes blocked algorithms in dask dependency graphs.
dask.bag encodes blocked algorithms on Python lists of arbitrary Python objects.
dask.async is a shared-memory asynchronous scheduler efficiently execute dask dependency graphs on multiple cores.

Dask does not currently have a distributed memory scheduler.

See full documentation at http://dask.pydata.org or read developer-focused blogposts about dask's development.

Install

Dask is easily installable through your favorite Python package manager:

conda install dask

or

pip install dask[array]
or
pip install dask[bag]
or
pip install dask[complete]

Dask Graphs

Consider the following simple program:

def inc(i):
    return i + 1

def add(a, b):
    return a + b

x = 1
y = inc(x)
z = add(y, 10)

We encode this as a dictionary in the following way:

d = {'x': 1,
     'y': (inc, 'x'),
     'z': (add, 'y', 10)}

While less aesthetically pleasing this dictionary may now be analyzed, optimized, and computed on by other Python code, not just the Python interpreter.

Dask Arrays

The dask.array module creates these graphs from NumPy-like operations

>>> import dask.array as da
>>> x = da.random.random((4, 4), blockshape=(2, 2))
>>> x.T[0, 3].dask
{('x', 0, 0): (np.random.random, (2, 2)),
 ('x', 0, 1): (np.random.random, (2, 2)),
 ('x', 1, 0): (np.random.random, (2, 2)),
 ('x', 1, 1): (np.random.random, (2, 2)),
 ('y', 0, 0): (np.transpose, ('x', 0, 0)),
 ('y', 0, 1): (np.transpose, ('x', 1, 0)),
 ('y', 1, 0): (np.transpose, ('x', 0, 1)),
 ('y', 1, 1): (np.transpose, ('x', 1, 1)),
 ('z',): (getitem, ('y', 0, 1), (0, 1))}

Finally, a scheduler executes these graphs to achieve the intended result. The dask.async module contains a shared memory scheduler that efficiently leverages multiple cores.

Dependencies

dask.core supports Python 2.6+ and Python 3.3+ with a common codebase. It is pure Python and requires no dependencies beyond the standard library. It is a light weight dependency.

dask.array depends on numpy.

dask.bag depends on toolz and dill.

LICENSE

New BSD. See License File.

Related Work

Task Scheduling

One might ask why we didn't use one of these other fine libraries:

Luigi
Joblib
mrjob
Any of the fine schedulers in numeric analysis (DAGue, ...)
Any of the fine high-throughput schedulers (Condor, Pegasus, Swiftlang, ...)

The answer is because we wanted all of the following:

Fine-ish grained parallelism (latencies around 1ms)
In-memory communication of intermediate results
Dependency structures more complex than map
Good support for numeric data
First class Python support
Trivial installation

Most task schedulers in the Python ecosystem target long-running batch jobs, often for processing large amounts of text and aren't appropriate for executing multi-core numerics.

Arrays

There are many "Big NumPy Array" or general distributed array solutions all with fine characteristics. Some projects in the Python ecosystem include the following:

There is a rich history of distributed array computing. An incomplete sampling includes the following projects:

Elemental
Plasma
Arrays in MLlib

marianotepper/dask