Template for deploying apps and endpoints on Outerbounds

Quickstart

1. Start a new workstation using the following image:

006988687827.dkr.ecr.us-west-2.amazonaws.com/obp-apps/vanilla:master-81c04052-1738296769

2. Clone this repo on the workstation

git clone https://github.com/outerbounds/ob-example-app.git

3. Deploy an example app

cd ob-example-app
./deployapp

It will output two URLs:

• A browser URL prefixed with https://ui.
• An API URL prefixed with https://api.

Click the browser URL to test the endpoint in the browser. If you see a JSON output, your endpoint works! 🎉

5. Test API access

You can access your freshly deployed endpoint in any environment that has a valid Outerbounds access token installed, either associated with a user or a machine token. For starters, you can test it on the workstation:

• Install a simple auth wrapper library: pip install -U outerbounds-app-client

• Run

python client_example.py https://api...

replacing the argument with the API URL output by the deployment flow. If you see a JSON output, you can access the endpoint programmatically from any authenticated environment! 🎉

To access the endpoint in another (micro)service, you should mint a new machine user for programmatic access, following the instructions here.

Tip

Reach out to your support Slack if you need help configuring machine users

Developing your own apps and endpoints

Your endpoint needs to be implemented as a Python package, such as myapp included here as an example. Your endpoint package can implement a Streamlit app, a FastAPI server, or any other service that serves HTTP requests at a given port.

Your package needs to implement an entrypoint script at __main__.py which starts the server. Importantly, the server needs to start at a port defined in appconfig.toml, which also specifes the package to be deployed with the app key.

List any Python (Pypi) dependencies that your service requires in the [dependencies] section in appconfig.toml. These libraries are installed automatically during deployment.

Serving models and other data through endpoints

You can expose any data, Metaflow artifacts in particular, to your deployments. For instance, you can deploy the latest model produced by a separate training flow, stored as an artifact.

To load arbitrary data in your deployments, implement a app_init module in your package. This module needs to include a method

def init(deploy_dir):
    ...

which is called at deploy time to make any data available for your endpoint. Typically, this method uses Metaflow client to retrieve artifacts and store them in the given deploy_dir so they are accessible by the deployment.

To see this in action, run the ScalableFlow example in the Getting Started section. The flow will produce a dummy model, stored as an artifact best. Deploy myapp again to invoke myapp.app_init which loads the artifact and serves it through the endpoint.