Overview
Most Tornado apps are run as single processes. For production, this usually means a fairly straight-forward combination of external process management and proxying. Here are some gathered best practices/resources.
Development
When debug mode is enabled, templates are not cached and the app will automatically restart during development. This will fail if a Python syntax error occurs, however. (This can be worked around w/ some additional code or by using Supervisor in development)
You might want to run your app from a terminal multiplexer like screen or tmux for more flexibility in leaving things running and tracing fatal errors.
Production
Typically in production, multiple tornado app processes are run (at least one per core) with a frontend proxy. Tornado developer bdarnell has a tornado-production-skeleton illustrating this using Supervisor (process management) and nginx (proxying).
Process Management
Traditionally, Tornado apps are single-process and require an external process manager, however HTTPServer can be run with multiple processes. Additionally, there are a couple additional helpers for helping out with managing multiple processes.
Supervisor
- Managing multiple Pylons apps with Supervisor - an excellent tutorial for getting started with Supervisor
- Deploy tornado application - short tutorial w/ Supervisor and nginx setup walkthrough
- supervisord-example.conf - this is an example cfg for running 4 tornado instances
Daemonizing
- start-stop-daemon example - if you are running a standard Linux system this is an easy way to daemonize your Tornado app
- Upstart example - Upstart is built into Ubuntu and can respawn crashed instances.
Tornado Multi-Process
As mentioned above, Tornado's HTTPServer can be configured for both multiple processes on a single or multiple sockets.
Proxying
The official docs includes an example for running nginx as a load balancing proxy and for serving static files.
Deployment
- Deploying Python web (Tornado) applications to multiple servers - short discussion on using a Load Balancer for live migration w/o downtime
- Rolling Deployment w/ Fabric
- Python Deployment Anti-Patterns
- buedafab - a nice collection of Fabric deployment scripts w/ EC2