Writing a custom Proxy implementation

JupyterHub 0.8 introduced the ability to write a custom implementation of the proxy. This enables deployments with different needs than the default proxy, configurable-http-proxy (CHP). CHP is a single-process nodejs proxy that they Hub manages by default as a subprocess (it can be run externally, as well, and typically is in production deployments).

The upside to CHP, and why we use it by default, is that it’s easy to install and run (if you have nodejs, you are set!). The downsides are that it’s a single process and does not support any persistence of the routing table. So if the proxy process dies, your whole JupyterHub instance is inaccessible until the Hub notices, restarts the proxy, and restores the routing table. For deployments that want to avoid such a single point of failure, or leverage existing proxy infrastructure in their chosen deployment (such as Kubernetes ingress objects), the Proxy API provides a way to do that.

In general, for a proxy to be usable by JupyterHub, it must:

  1. support websockets without prior knowledge of the URL where websockets may occur
  2. support trie-based routing (i.e. allow different routes on /foo and /foo/bar and route based on specificity)
  3. adding or removing a route should not cause existing connections to drop

Optionally, if the JupyterHub deployment is to use host-based routing, the Proxy must additionally support routing based on the Host of the request.

Subclassing Proxy

To start, any Proxy implementation should subclass the base Proxy class, as is done with custom Spawners and Authenticators.

from jupyterhub.proxy import Proxy

class MyProxy(Proxy):
    """My Proxy implementation"""

Starting and stopping the proxy

If your proxy should be launched when the Hub starts, you must define how to start and stop your proxy:

from tornado import gen
class MyProxy(Proxy):
    def start(self):
        """Start the proxy"""

    def stop(self):
        """Stop the proxy"""

These methods may be coroutines.

c.Proxy.should_start is a configurable flag that determines whether the Hub should call these methods when the Hub itself starts and stops.

Purely external proxies

Probably most custom proxies will be externally managed, such as Kubernetes ingress-based implementations. In this case, you do not need to define start and stop. To disable the methods, you can define should_start = False at the class level:

class MyProxy(Proxy):
    should_start = False


At its most basic, a Proxy implementation defines a mechanism to add, remove, and retrieve routes. A proxy that implements these three methods is complete. Each of these methods may be a coroutine.

Definition: routespec

A routespec, which will appear in these methods, is a string describing a route to be proxied, such as /user/name/. A routespec will:

  1. always end with /
  2. always start with / if it is a path-based route /proxy/path/
  3. precede the leading / with a host for host-based routing, e.g. host.tld/proxy/path/

Adding a route

When adding a route, JupyterHub may pass a JSON-serializable dict as a data argument that should be attacked to the proxy route. When that route is retrieved, the data argument should be returned as well. If your proxy implementation doesn’t support storing data attached to routes, then your Python wrapper may have to handle storing the data piece itself, e.g in a simple file or database.

def add_route(self, routespec, target, data):
    """Proxy `routespec` to `target`.

    Store `data` associated with the routespec
    for retrieval later.

Adding a route for a user looks like this:

proxy.add_route('/user/pgeorgiou/', '',
                {'user': 'pgeorgiou'})

Removing routes

delete_route() is given a routespec to delete. If there is no such route, delete_route should still succeed, but a warning may be issued.

def delete_route(self, routespec):
    """Delete the route"""

Retrieving routes

For retrieval, you only need to implement a single method that retrieves all routes. The return value for this function should be a dictionary, keyed by routespect, of dicts whose keys are the same three arguments passed to add_route (routespec, target, data)

def get_all_routes(self):
    """Return all routes, keyed by routespec"""
  '/proxy/path/': {
    'routespec': '/proxy/path/',
    'target': 'http://...',
    'data': {},

Note on activity tracking

JupyterHub can track activity of users, for use in services such as culling idle servers. As of JupyterHub 0.8, this activity tracking is the responsibility of the proxy. If your proxy implementation can track activity to endpoints, it may add a last_activity key to the data of routes retrieved in .get_all_routes(). If present, the value of last_activity should be an ISO8601 UTC date string:

  '/user/pgeorgiou/': {
    'routespec': '/user/pgeorgiou/',
    'target': '',
    'data': {
      'user': 'pgeourgiou',
      'last_activity': '2017-10-03T10:33:49.570Z',

If the proxy does not track activity, then only activity to the Hub itself is tracked, and services such as cull-idle will not work.

Now that notebook-5.0 tracks activity internally, we can retrieve activity information from the single-user servers instead, removing the need to track activity in the proxy. But this is not yet implemented in JupyterHub 0.8.0.