Starpack is tool to package and deploy production-ready packages of machine learning models.

This repository contains the code for the CLI and Python library for interacting with Starpack. To run Starpack locally, you will need to install a Python version listed in the badge above and Docker Desktop installed locally.

Currently, Starpack only supports packaging and deployment in local Docker Desktop or Docker Engine environments, but the generated Docker images can be pushed to remote repositories and deployed manually. It is currently on our roadmap to support deployment through Andromeda 360 down the line.

Table of Contents

• Starpack
  • Table of Contents
  • Quickstart
  • Full Command List
    • Init
    • Upload
    • Package
      • Metadata
      • Artifacts
      • Steps
    • Deploy
    • Engine
      • Start
      • Terminate
    • Plugins
  • Examples
    • Basic Example
  • Troubleshooting

Quickstart

To begin using Starpack, open a terminal and run

pip install starpack

Then, you can initialize an existing or new directory with starter code and files by running the following in a terminal

starpack init ./path/you/want

Reconfigure your starpack.yaml, predict.py, and requirements.txt files to point to your model artifacts, properly ingest your data, and list your project dependencies, respectively. An example of these files can be found in the examples folder.

Finally, run the following terminal command to package and deploy your model locally

starpack deploy ./path/you/want

or alternatively in Python (such as in a notebook), the process can be run as follows:

import starpack
from pathlib import Path

desired_directory = Path("./path/you/want")

starpack.initialize_directory(desired_directory)

# Reconfigure your files that have been initialized for your specific project

starpack.deploy_directory(desired_directory)

Full Command List

A full command list can be found by running

starpack --help

Currently, the following top-level flags are available:

Additionally, we have a variety of commands to manage the Starpack Engine and generate packages.

Init

The command, starpack init, when provided with a local path as an argument, will generate files to help you start developing your package definition. By default, Starpack will prompt for permission to overwrite existing files, unless the --ovewrite/-o flag is given.

Upload

The command, starpack upload, takes in a directory path and uploads its contents to the Starpack Engine. The name of this directory will be the name of the artifacts location within the Engine for when you package and deploy your model.

Package

The command, starpack package, takes in a path, either a directory or starpack.yaml in order to package your model. If given a directory, Starpack will upload the contents to the Starpack Engine, then find any starpack.yaml file in the given directory to send as a payload to the Starpack Engine.

The starpack.yaml file contains two main sections, package and deployment. Furthermore, the YAML packaging section contains three main subsections:

1. Metadata
2. Artifacts
3. Steps

Metadata

Within the metadata section, you can define things such as the name, version, and author of the package. These fields should be generated when creating a directory using starpack init

Artifacts

This section defines the name of your model artifacts location in the Engine system, as well as where you can find files such as your validation data, prediction script, and dependencies.

Steps

This section is an ordered list of your build steps. Each of your build steps uses either built-in or externally written plugins that can be called by name and optionally constrained by version of the plugin. Some plugins may need additional data, which will be included within the step's definition.

Deploy

The command, starpack deploy, takes in a path, either a directory or starpack.yaml in order to deploy your packaged model. If given a path, the assumption is made that the contents should be uploaded to the Engine, the YAML should be parsed for any packaging information, and finally, the deployment should be processed. If given just a starpack.yaml, only the deployment step will be run, with the assumption that the other steps have already been run previously.

Engine

There are several

Start

The command, starpack engine start, is used to ensure the Starpack Engine is running or force the creation of a new container.

Finally, if you pass the --force or -F flags, you will force the deletion of any existing Starpack Engines and create a new Docker container for the Engine.

Terminate

The command, starpack engine terminate, is used to spin down any existing Starpack Engines running on your local machine. Additionally, the --all or -A flag can be passed to additionally delete any existing Docker Volumes and associated data from your machine.

Plugins

The following plugins are available to use in either packaging or deployment:

Examples

Basic Example

Under examples/starpack_basic_example, you can find a Jupyter notebook and associated files for deploying a Starpack package using both Streamlit and FastAPI. By following along with the notebook, then running starpack deploy examples/starpack_basic_example from the root of this repository, you will be able to see how one exports, defines, packages, and deploys a Starpack model.

Troubleshooting

Before troubleshooting any issues, please run the two following commands in your terminal to ensure that you're running the latest version of both the Starpack CLI and Starpack Engine:

pip install starpack --upgrade
starpack engine start --force