Dask Tutorial

This tutorial was last given at SciPy 2020 which was a virtual conference. A video of the SciPy 2020 tutorial is available online.

Dask provides multi-core execution on larger-than-memory datasets.

We can think of dask at a high and a low level

• High level collections: Dask provides high-level Array, Bag, and DataFrame collections that mimic NumPy, lists, and Pandas but can operate in parallel on datasets that don't fit into main memory. Dask's high-level collections are alternatives to NumPy and Pandas for large datasets.
• Low Level schedulers: Dask provides dynamic task schedulers that execute task graphs in parallel. These execution engines power the high-level collections mentioned above but can also power custom, user-defined workloads. These schedulers are low-latency (around 1ms) and work hard to run computations in a small memory footprint. Dask's schedulers are an alternative to direct use of threading or multiprocessing libraries in complex cases or other task scheduling systems like Luigi or IPython parallel.

Different users operate at different levels but it is useful to understand both. This tutorial will interleave between high-level use of dask.array and dask.dataframe (even sections) and low-level use of dask graphs and schedulers (odd sections.)

Prepare

1. You should clone this repository

git clone http://github.com/dask/dask-tutorial

and then install necessary packages. There are three different ways to achieve this, pick the one that best suits you, and only pick one option. They are, in order of preference:

2a) Create a conda environment (preferred)

In the main repo directory

conda env create -f binder/environment.yml
conda activate dask-tutorial
jupyter labextension install @jupyter-widgets/jupyterlab-manager
jupyter labextension install @bokeh/jupyter_bokeh

2b) Install into an existing environment

You will need the following core libraries

conda install numpy pandas h5py pillow matplotlib scipy toolz pytables snakeviz scikit-image dask distributed -c conda-forge

You may find the following libraries helpful for some exercises

conda install python-graphviz -c conda-forge

Note that these options will alter your existing environment, potentially changing the versions of packages you already have installed.

2c) Use Dockerfile

You can build a docker image from the provided Dockerfile.

$ docker build . # This will build using the same env as in a)

Run a container, replacing the ID with the output of the previous command

$ docker run -it -p 8888:8888 -p 8787:8787 <container_id_or_tag>

The above command will give an URL (Like http://(container_id or 127.0.0.1):8888/?token=<sometoken>) which can be used to access the notebook from browser. You may need to replace the given hostname with "localhost" or "127.0.0.1".

You should follow only one of the options above!

Launch notebook

From the repo directory

jupyter notebook

Or

jupyter lab

This was already done for method c) and does not need repeating.

Links

• Reference
  • Docs
  • Examples
  • Code
  • Blog
• Ask for help
  • dask tag on Stack Overflow, for usage questions
  • github issues for bug reports and feature requests
  • discourse forum for general, non-bug, questions and discussion
  • Attend a live tutorial

Outline

0. Overview - dask's place in the universe.

1. Delayed - the single-function way to parallelize general python code.

1x. Lazy - some of the principles behind lazy execution, for the interested.

2. Bag - the first high-level collection: a generalized iterator for use with a functional programming style and to clean messy data.

3. Array - blocked numpy-like functionality with a collection of numpy arrays spread across your cluster.

4. Dataframe - parallelized operations on many pandas dataframes spread across your cluster.

5. Distributed - Dask's scheduler for clusters, with details of how to view the UI.

6. Advanced Distributed - further details on distributed computing, including how to debug.

7. Dataframe Storage - efficient ways to read and write dataframes to disc.

8. Machine Learning - applying dask to machine-learning problems.