Deploying Clusters

Deploying Clusters

The dask.distributed scheduler works well on a single machine and scales to many machines in a cluster. We recommend using dask.distributed clusters at all scales for the following reasons:

  1. It provides access to asynchronous API, notably Futures

  2. It provides a diagnostic dashboard that can provide valuable insight on performance and progress

  3. It handles data locality with sophistication, and so can be more efficient than the multiprocessing scheduler on workloads that require multiple processes

This page describes various ways to set up Dask clusters on different hardware, either locally on your own machine or on a distributed cluster. If you are just getting started then you can save this page for later as Dask runs perfectly well on a single machine without a distributed scheduler. But once you start using Dask in anger you’ll find a lot of benefit both in terms of scaling and debugging by using the distirbuted scheduler.

You can continue reading or watch the screencast below:

If you import Dask, set up a computation, and call compute, then you will use the single-machine scheduler by default. To use the dask.distributed scheduler you must set up a Client.

import dask.dataframe as dd
df = dd.read_csv(...)
df.x.sum().compute()  # This uses the single-machine scheduler by default
from dask.distributed import Client
client = Client(...)  # Connect to distributed cluster and override default
df.x.sum().compute()  # This now runs on the distributed system

There are many ways to start the distributed scheduler and worker components that your client needs to connect to. You can run them manually using command line tools but often the most straight forward way is to use a cluster manager utility class.

from dask.distributed import Client, LocalCluster
cluster = LocalCluster()  # Launches a scheduler and workers locally
client = Client(cluster)  # Connect to distributed cluster and override default
df.x.sum().compute()  # This now runs on the distributed system

There are a number of different cluster managers available, so you can use Dask distributed with a range of platforms. These cluster managers deploy a scheduler and the necessary workers as determined by communicating with the resource manager. All cluster managers follow the same interface but have platform specific configuration options. This makes it convenient to switch from your local machine to a remote multi-node cluster without sacrificing the flexibility of the platform you are deploying on.

Dask Jobqueue, for example, is a set of cluster managers for HPC users and works with job queueing systems (in this case, the resource manager) such as PBS, Slurm, and SGE. Those workers are then allocated physical hardware resources.

from dask.distributed import Client
from dask_jobqueue import PBSCluster
cluster = PBSCluster()  # Launches a scheduler and workers on HPC via PBS
client = Client(cluster)  # Connect to distributed cluster and override default
df.x.sum().compute()  # This now runs on the distributed system

An overview of cluster management with Dask distributed.

To summarize, you can use the default, single-machine scheduler to use Dask on your local machine. If you’d like use a cluster or simply take advantage of the extensive diagnostics, you can use Dask distributed. The following resources explain in more detail how to set up Dask on a variety of local and distributed hardware:

  • Single Machine:
    • Default Scheduler: The no-setup default. Uses local threads or processes for larger-than-memory processing

    • dask.distributed: The sophistication of the newer system on a single machine. This provides more advanced features while still requiring almost no setup.

  • Distributed computing:
    • Beginner’s Guide to Configuring a Dask distributed Cluster

    • Overview of cluster management options

    • Manual Setup: The command line interface to set up dask-scheduler and dask-worker processes. Useful for IT or anyone building a deployment solution.

    • SSH: Use SSH to set up Dask across an un-managed cluster.

    • High Performance Computers: How to run Dask on traditional HPC environments using tools like MPI, or job schedulers like SLURM, SGE, TORQUE, LSF, and so on.

    • Kubernetes: Deploy Dask with the popular Kubernetes resource manager using either Helm or a native deployment.

    • YARN / Hadoop: Deploy Dask on YARN clusters, such as are found in traditional Hadoop installations.

    • Dask Gateway provides a secure, multi-tenant server for managing Dask clusters and allows users to launch and use Dask clusters in a shared cluster environment.

    • Python API (advanced): Create Scheduler and Worker objects from Python as part of a distributed Tornado TCP application. This page is useful for those building custom frameworks.

    • Docker images are available and may be useful in some of the solutions above.

    • Cloud for current recommendations on how to deploy Dask and Jupyter on common cloud providers like Amazon, Google, or Microsoft Azure.

  • Hosted / managed Dask clusters (listed in alphabetical order):
    • Coiled handles the creation and management of Dask clusters on cloud computing environments (AWS, Azure, and GCP).

    • Domino Data Lab lets users create Dask clusters in a hosted platform.

    • Saturn Cloud lets users create Dask clusters in a hosted platform or within their own AWS accounts.