dask-array

Dask array with query optimization

Motivation

Dask Array is powerful, but requires expertise to drive effectively.

The dask-array project reimplements dask.array, but represents your calculation at a level where it can be optimized intelligently before executed. This allows the project to reorder and replace calculations to provide the same result but with a more efficient path.

Installation

pip install dask-array

Usage

This project looks and feels like Dask Array

import dask_array as da

x = da.ones((1000, 1000), chunks=(100, 100))

y = x + x.T
result = y[:100, :100]

But when you go to compute, your calculation gets rewritten to be more efficient. This is apparent if you look at the query structure of the underlying array.

>>> result.pprint()
Operation             Shape    Bytes   Chunks
  Getitem          (100, 100)   78 kiB  100×100
  └ Add          (1000, 1000)  7.6 MiB  100×100
    ├ Ones       (1000, 1000)  7.6 MiB  100×100
    └ Transpose  (1000, 1000)  7.6 MiB  100×100
      └ Ones     (1000, 1000)  7.6 MiB  100×100

This calculation starts from a large array expression, then takes a small slice. It is more efficient to push that slice into the expression before building the task graph.

The optimize function rewrites things automatically.

>>> result.optimize().pprint()
Operation            Shape   Bytes   Chunks
  FusedBlockwise  (100, 100)  78 kiB  100×100
  └ Ones          (100, 100)  78 kiB  100×100

You don't need to call optimize though. Dask compute/persist machinery will do this for you. You just need to change your import.

# import dask.array as da
import dask_array as da

Native Frisky acceleration

The normal Dask scheduler path is pure Python and works without a Rust toolchain. On platforms with a native wheel, dask-array also includes a Rust accelerator that Frisky can use to submit compact task records instead of materializing large Python task graphs. This is automatic when computing with a Frisky scheduler; otherwise computations stay on the standard Dask path.

Xarray

Dask-array can replace the default dask.array by registering itself as an Xarray chunk manager.

from dask_array.xarray import register
register()

After this all xarray calculations will benefit from query optimization.

Call register() before you create any chunked arrays. It is the only thing that turns this on: installing or importing dask-array leaves xarray alone, so that adding this package to an environment cannot change how other libraries that use xarray and dask behave.