Introduction to Cloud Tasks

Introduction

Cloud Tasks (contained in the rms-cloud-tasks package) is a framework for running independent tasks on cloud providers with automatic compute instance and task queue management. It is specifically designed for running the same code multiple times in a batch environment to process a series of different inputs. For example, the program could be an image processing program that takes the image filename as an argument, downloads the image from the cloud, performs some manipulations, and writes the result to a cloud-based location. It is very important that the tasks are completely independent; no communication between them is supported. Also, the processing happens entirely in a batch mode: a certain number of compute instances are created, they all process tasks in parallel, and then the compute instances are destroyed.

rms-cloud-tasks is a product of the PDS Ring-Moon Systems Node.

Features

Cloud Tasks is extremely easy to use with a simple command line interface and straightforward configuration file. It supports AWS and GCP compute instances and queues along with the ability to run jobs on a local workstation, all using a provider-independent API. Although each cloud provider has implemented similar functionality as part of their offering (e.g. GCP’s Cloud Batch), Cloud Tasks is unique in that it unifies all supported providers into a single, simple, universal system that does not require learning the often-complicated details of the official full-featured services.

Cloud Tasks consists of four primary components:

  • A Python module to make parallel execution simple

    • Allows conversion of an existing Python program to a parallel task with only a few lines of code

    • Supports both cloud compute instance and local machine environments

    • Executes each task in its own process for complete isolation

    • Reads task information from a cloud-based task queue or directly from a local file

    • Monitors the state of spot instances to notify tasks of upcoming preemption

  • A command line interface to manage the task queue system, that allows

    • Loading of tasks from a JSON or YAML file

    • Checking the status of a queue

    • Purging a queue of remaining tasks

    • Deleting a queue entirely

  • A command line interface to query the cloud about available resources, given certain constraints

    • Types of compute instances available, including price (both demand and spot instances)

    • VM boot images available

    • Regions and zones

  • A command line interface to manage a pool of compute instances optimized for price, given certain constraints

    • Automatically finds the optimal compute instance type given pricing and other constraints

    • Automatically determines the number of simultaneous instances to use

    • Creates new instances and runs a specified startup script to execute the task manager

    • Monitors instances for failure or preemption and creates new instances as needed to keep the compute pool full

    • Detects when all jobs are complete and terminates the instances

Installation

cloud_tasks consists of a command line interface (called cloud_tasks) and a Python module (also called cloud_tasks). They are available via the rms-cloud-tasks package on PyPI and can be installed with:

pip install rms-cloud-tasks

Note that this will install cloud_tasks into your current system Python, or into your currently activated virtual environment (venv), if any.

If you already have the rms-cloud-tasks package installed but wish to upgrade to a more recent version, you can use:

pip install --upgrade rms-cloud-tasks

You may also install cloud_tasks using pipx, which will isolate the installation from your system Python without requiring the creation of a virtual environment. To install pipx, please see the installation instructions. Once pipx is available, you may install cloud_tasks with:

pipx install rms-cloud-tasks

If you already have the rms-cloud-tasks package installed with pipx, you may upgrade to a more recent version with:

pipx upgrade rms-cloud-tasks

Using pipx is only useful if you want to use the command line interface and not access the Python module; however, it does not require you to worry about the Python version, setting up a virtual environment, etc.

Basic Examples

The cloud_tasks command line program supports many useful commands that control the task queue, compute instance pool, and retrieve general information about the cloud in a provider-indepent manner. A few examples are given below.

To get a list of available commands:

cloud_tasks --help

To get help on a particular command:

cloud_tasks load_queue --help

To list all ARM64-based compute instance types that have 2 to 4 vCPUs and at most 4 GB memory per vCPU.

cloud_tasks list_instance_types \
  --provider gcp --region us-central1 \
  --min-cpu 2 --max-cpu 4 --arch ARM64 --max-memory-per-cpu 4

To load a JSON file containing task descriptions into the task queue:

cloud_tasks load_queue \
  --provider gcp --region us-central1 --project-id my-project \
  --job-id my-job --task-file mytasks.json

To start automatic creation and management of a compute instance pool:

cloud_tasks manage_pool --provider gcp --config myconfig.yaml

Contributing

Information on contributing to this package can be found in the Contributing Guide.

Licensing

This code is licensed under the Apache License v2.0.

Contents:

Indices and tables