# Host ai workflow runner with Starlette: streamed responses

Deploy an AI workflow runner built with Starlette on Ample using the streaming response service pattern. Compute runs the app in an isolated microVM behind a public HTTPS URL, a managed PostgreSQL 16 database is auto-provisioned and injected as DATABASE_URL. Verified on Starlette: a Server-Sent Events endpoint read incrementally through the public HTTPS gateway: five one-second chunks arrived spread over time with the first within seconds (not buffered), a 70-second stream of 36 chunks completed past common 60-second idle timeouts, and a client that disconnected after two chunks was observed and recorded by the server (disconnected=true). The AI workflow runner-specific behavior (your event schema, reconnection strategy and any per-request duration ceiling beyond the 70 seconds measured) is your application code on that verified basis.

Verified November 6, 2026

## What it runs on

- [Computeverified](/content/discover/primitive/compute/index.html)  
  Apps run in isolated x86_64 Firecracker microVMs; sizes are the priced VM sizes.

- [Postgresverified](/content/discover/primitive/postgres/index.html)  
  Managed PostgreSQL 16 with pgvector runs in its own microVM and is auto-provisioned when an app needs a database and no DATABASE_URL is supplied.

## You need

- A Starlette project that builds and starts with the documented commands (pip install into .ample/python from requirements.txt, then uvicorn from run.py reading PORT on the python-3.12 template)

- A PostgreSQL driver reading DATABASE_URL at runtime (auto-provisioned when omitted, or supplied with --env)

- An Ample account token with servers:write, databases:read

## Exact tested configuration

```bash
install python3 -m pip install --target .ample/python -r requirements.txt
runtime python
sizes - 1vcpu - 1gb
start PYTHONPATH = .ample/python:${PYTHONPATH:-} python3 run.py
template python - 3.12
```

## Steps

1.  **Build and start.** pip install into .ample/python from requirements.txt, then uvicorn from run.py reading PORT on the python-3.12 template. The server must bind 0.0.0.0 on PORT.

2.  **Implement the pattern on PostgreSQL.** The fixture's module implements streaming response service: a Server-Sent Events endpoint read incrementally through the public HTTPS gateway: five one-second chunks arrived spread over time with the first within seconds (not buffered), a 70-second stream of 36 chunks completed past common 60-second idle timeouts, and a client that disconnected after two chunks was observed and recorded by the server (disconnected=true). Copy the approach into your schema; keep migrations idempotent and run them with --release-command.

3.  **Deploy.** Run the synchronous deploy once and read the result (exit 0 live, 1 failed, 2 needs an answer first). Re-running with no change is a no-op.

```bash
$ ample deploy . --name <app-name> --public --start "python3 run.py"
```

4.  **Verify.** Fetch the live URL and the pattern self-test route(s) (/p/streaming-response-service) from the example; then run your own checks. On failure read `ample logs <deployment_id> --kind build` then `--kind runtime`.

```bash
$ ample logs <deployment_id> --kind build
```

## How it was checked

✓app responds on its public URL  
✓streaming-response-service self-test

## Evidence

1. November 6, 2026 · canary run  
   Starlette pattern fixture deployed on Ample (pip install into .ample/python from requirements.txt, then uvicorn from run.py reading PORT on the python-3.12 template); checks passed for streaming-response-service and configuration-secrets. Pattern proof: a Server-Sent Events endpoint read incrementally through the public HTTPS gateway: five one-second chunks arrived spread over time with the first within seconds (not buffered), a 70-second stream of 36 chunks completed past common 60-second idle timeouts, and a client that disconnected after two chunks was observed and recorded by the server (disconnected=true).

2. November 5, 2026 · canary run  
   The same Starlette app deployed with a --release-command migration; the marker it created was readable after activation.

## Tested examples

- **Starlette pattern fixture.** Multi-pattern Starlette app whose streaming response service module was checked live; the module is under tests/deploy-canaries/_pattern-modules.

## What it doesn’t cover

- Verified on the python-3.12 template at s-1vcpu-1gb with the example fixture.
- The pattern checks verified the platform side; the AI workflow runner itself (your event schema, reconnection strategy and any per-request duration ceiling beyond the 70 seconds measured) is your application code.
- Managed PostgreSQL 16 with pgvector; apps and their databases are placed together.
