Services#

Definition of a Service#

When working with JupyterHub, a Service is defined as a process that interacts with the Hub’s REST API. A Service may perform a specific action or task. For example, the following tasks can each be a unique Service:

  • shutting down individuals’ single user notebook servers that have been idle for some time

  • registering additional web servers which should use the Hub’s authentication and be served behind the Hub’s proxy.

Two key features help define a Service:

  • Is the Service managed by JupyterHub?

  • Does the Service have a web server that should be added to the proxy’s table?

Currently, these characteristics distinguish two types of Services:

  • A Hub-Managed Service which is managed by JupyterHub

  • An Externally-Managed Service which runs its own web server and communicates operation instructions via the Hub’s API.

Properties of a Service#

A Service may have the following properties:

  • name: str - the name of the service

  • admin: bool (default - false) - whether the service should have administrative privileges

  • url: str (default - None) - The URL where the service is/should be. If a url is specified for where the Service runs its own web server, the service will be added to the proxy at /services/:name

  • api_token: str (default - None) - For Externally-Managed Services you need to specify an API token to perform API requests to the Hub

  • display: bool (default - True) - When set to true, display a link to the service’s URL under the ‘Services’ dropdown in user’s hub home page.

If a service is also to be managed by the Hub, it has a few extra options:

  • command: (str/Popen list) - Command for JupyterHub to spawn the service. - Only use this if the service should be a subprocess. - If command is not specified, the Service is assumed to be managed externally. - If a command is specified for launching the Service, the Service will be started and managed by the Hub.

  • environment: dict - additional environment variables for the Service.

  • user: str - the name of a system user to manage the Service. If unspecified, run as the same user as the Hub.

Hub-Managed Services#

A Hub-Managed Service is started by the Hub, and the Hub is responsible for the Service’s actions. A Hub-Managed Service can only be a local subprocess of the Hub. The Hub will take care of starting the process and restarts it if it stops.

While Hub-Managed Services share some similarities with notebook Spawners, there are no plans for Hub-Managed Services to support the same spawning abstractions as a notebook Spawner.

If you wish to run a Service in a Docker container or other deployment environments, the Service can be registered as an Externally-Managed Service, as described below.

Launching a Hub-Managed Service#

A Hub-Managed Service is characterized by its specified command for launching the Service. For example, a ‘cull idle’ notebook server task configured as a Hub-Managed Service would include:

  • the Service name,

  • admin permissions, and

  • the command to launch the Service which will cull idle servers after a timeout interval

This example would be configured as follows in jupyterhub_config.py:

c.JupyterHub.load_roles = [
    {
        "name": "idle-culler",
        "scopes": [
            "read:users:activity", # read user last_activity
            "servers", # start and stop servers
            # 'admin:users' # needed if culling idle users as well
        ]
    }
]

c.JupyterHub.services = [
    {
        'name': 'idle-culler',
        'command': [sys.executable, '-m', 'jupyterhub_idle_culler', '--timeout=3600']
    }
]

A Hub-Managed Service may also be configured with additional optional parameters, which describe the environment needed to start the Service process:

  • environment: dict - additional environment variables for the Service.

  • user: str - name of the user to run the server if different from the Hub. Requires Hub to be root.

  • cwd: path directory in which to run the Service, if different from the Hub directory.

The Hub will pass the following environment variables to launch the Service:

JUPYTERHUB_SERVICE_NAME:   The name of the service
JUPYTERHUB_API_TOKEN:      API token assigned to the service
JUPYTERHUB_API_URL:        URL for the JupyterHub API (default, http://127.0.0.1:8080/hub/api)
JUPYTERHUB_BASE_URL:       Base URL of the Hub (https://mydomain[:port]/)
JUPYTERHUB_SERVICE_PREFIX: URL path prefix of this service (/services/:service-name/)
JUPYTERHUB_SERVICE_URL:    Local URL where the service is expected to be listening.
                           Only for proxied web services.
JUPYTERHUB_OAUTH_SCOPES:   JSON-serialized list of scopes to use for allowing access to the service.

For the previous ‘cull idle’ Service example, these environment variables would be passed to the Service when the Hub starts the ‘cull idle’ Service:

JUPYTERHUB_SERVICE_NAME: 'idle-culler'
JUPYTERHUB_API_TOKEN: API token assigned to the service
JUPYTERHUB_API_URL: http://127.0.0.1:8080/hub/api
JUPYTERHUB_BASE_URL: https://mydomain[:port]
JUPYTERHUB_SERVICE_PREFIX: /services/idle-culler/

See the GitHub repo for additional information about the jupyterhub_idle_culler.

Externally-Managed Services#

You may prefer to use your own service management tools, such as Docker or systemd, to manage a JupyterHub Service. These Externally-Managed Services, unlike Hub-Managed Services, are not subprocesses of the Hub. You must tell JupyterHub which API token the Externally-Managed Service is using to perform its API requests. Each Externally-Managed Service will need a unique API token, because the Hub authenticates each API request and the API token is used to identify the originating Service or user.

A configuration example of an Externally-Managed Service with admin access and running its own web server is:

c.JupyterHub.services = [
    {
        'name': 'my-web-service',
        'url': 'https://10.0.1.1:1984',
        # any secret >8 characters, you'll use api_token to
        # authenticate api requests to the hub from your service
        'api_token': 'super-secret',
    }
]

In this case, the url field will be passed along to the Service as JUPYTERHUB_SERVICE_URL.

Writing your own Services#

When writing your own services, you have a few decisions to make (in addition to what your service does!):

  1. Does my service need a public URL?

  2. Do I want JupyterHub to start/stop the service?

  3. Does my service need to authenticate users?

When a Service is managed by JupyterHub, the Hub will pass the necessary information to the Service via the environment variables described above. A flexible Service, whether managed by the Hub or not, can make use of these same environment variables.

When you run a service that has a url, it will be accessible under a /services/ prefix, such as https://myhub.horse/services/my-service/. For your service to route proxied requests properly, it must take JUPYTERHUB_SERVICE_PREFIX into account when routing requests. For example, a web service would normally service its root handler at '/', but the proxied service would need to serve JUPYTERHUB_SERVICE_PREFIX.

Note that JUPYTERHUB_SERVICE_PREFIX will contain a trailing slash. This must be taken into consideration when creating the service routes. If you include an extra slash you might get unexpected behavior. For example if your service has a /foo endpoint, the route would be JUPYTERHUB_SERVICE_PREFIX + foo, and /foo/bar would be JUPYTERHUB_SERVICE_PREFIX + foo/bar.

Hub Authentication and Services#

JupyterHub provides some utilities for using the Hub’s authentication mechanism to govern access to your service.

Requests to all JupyterHub services are made with OAuth tokens. These can either be requests with a token in the Authorization header, or url parameter ?token=..., or browser requests which must complete the OAuth authorization code flow, which results in a token that should be persisted for future requests (persistence is up to the service, but an encrypted cookie confined to the service path is appropriate, and provided by default).

Changed in version 2.0: The shared jupyterhub-services cookie is removed. OAuth must be used to authenticate browser requests with services.

JupyterHub includes a reference implementation of Hub authentication that can be used by services. You may go beyond this reference implementation and create custom hub-authenticating clients and services. We describe the process below.

The reference, or base, implementation is the HubAuth class, which implements the API requests to the Hub that resolve a token to a User model.

There are two levels of authentication with the Hub:

  • HubAuth - the most basic authentication, for services that should only accept API requests authorized with a token.

  • HubOAuth - For services that should use oauth to authenticate with the Hub. This should be used for any service that serves pages that should be visited with a browser.

To use HubAuth, you must set the .api_token, either programmatically when constructing the class, or via the JUPYTERHUB_API_TOKEN environment variable.

Most of the logic for authentication implementation is found in the HubAuth.user_for_token() methods, which makes a request of the Hub, and returns:

  • None, if no user could be identified, or

  • a dict of the following form:

    {
      "name": "username",
      "groups": ["list", "of", "groups"],
      "scopes": [
          "access:users:servers!server=username/",
      ],
    }
    

You are then free to use the returned user information to take appropriate action.

HubAuth also caches the Hub’s response for a number of seconds, configurable by the cookie_cache_max_age setting (default: five minutes).

If your service would like to make further requests on behalf of users, it should use the token issued by this OAuth process. If you are using tornado, you can access the token authenticating the current request with HubAuth.get_token().

Changed in version 2.2: HubAuth.get_token() adds support for retrieving tokens stored in tornado cookies after completion of OAuth. Previously, it only retrieved tokens from URL parameters or the Authorization header. Passing get_token(handler, in_cookie=False) preserves this behavior.

Flask Example#

For example, you have a Flask service that returns information about a user. JupyterHub’s HubAuth class can be used to authenticate requests to the Flask service. See the service-whoami-flask example in the JupyterHub GitHub repo for more details.

#!/usr/bin/env python3
"""
whoami service authentication with the Hub
"""
import json
import os
import secrets
from functools import wraps

from flask import Flask, Response, make_response, redirect, request, session

from jupyterhub.services.auth import HubOAuth

prefix = os.environ.get('JUPYTERHUB_SERVICE_PREFIX', '/')

auth = HubOAuth(api_token=os.environ['JUPYTERHUB_API_TOKEN'], cache_max_age=60)

app = Flask(__name__)
# encryption key for session cookies
app.secret_key = secrets.token_bytes(32)


def authenticated(f):
    """Decorator for authenticating with the Hub via OAuth"""

    @wraps(f)
    def decorated(*args, **kwargs):
        token = session.get("token")

        if token:
            user = auth.user_for_token(token)
        else:
            user = None

        if user:
            return f(user, *args, **kwargs)
        else:
            # redirect to login url on failed auth
            state = auth.generate_state(next_url=request.path)
            response = make_response(redirect(auth.login_url + '&state=%s' % state))
            response.set_cookie(auth.state_cookie_name, state)
            return response

    return decorated


@app.route(prefix)
@authenticated
def whoami(user):
    return Response(
        json.dumps(user, indent=1, sort_keys=True), mimetype='application/json'
    )


@app.route(prefix + 'oauth_callback')
def oauth_callback():
    code = request.args.get('code', None)
    if code is None:
        return 403

    # validate state field
    arg_state = request.args.get('state', None)
    cookie_state = request.cookies.get(auth.state_cookie_name)
    if arg_state is None or arg_state != cookie_state:
        # state doesn't match
        return 403

    token = auth.token_for_code(code)
    # store token in session cookie
    session["token"] = token
    next_url = auth.get_next_url(cookie_state) or prefix
    response = make_response(redirect(next_url))
    return response

Authenticating tornado services with JupyterHub#

Since most Jupyter services are written with tornado, we include a mixin class, [HubOAuthenticated][huboauthenticated], for quickly authenticating your own tornado services with JupyterHub.

Tornado’s authenticated() decorator calls a Handler’s get_current_user() method to identify the user. Mixing in HubAuthenticated defines get_current_user() to use HubAuth. If you want to configure the HubAuth instance beyond the default, you’ll want to define an initialize() method, such as:

class MyHandler(HubOAuthenticated, web.RequestHandler):

    def initialize(self, hub_auth):
        self.hub_auth = hub_auth

    @web.authenticated
    def get(self):
        ...

The HubAuth class will automatically load the desired configuration from the Service environment variables.

Changed in version 2.0: Access scopes are used to govern access to services. Prior to 2.0, sets of users and groups could be used to grant access by defining .hub_groups or .hub_users on the authenticated handler. These are ignored if the 2.0 .hub_scopes is defined.

Implementing your own Authentication with JupyterHub#

If you don’t want to use the reference implementation (e.g. you find the implementation a poor fit for your Flask app), you can implement authentication via the Hub yourself. JupyterHub is a standard OAuth2 provider, so you can use any OAuth 2 client implementation appropriate for your toolkit. See the FastAPI example for an example of using JupyterHub as an OAuth provider with FastAPI, without using any code imported from JupyterHub.

On completion of OAuth, you will have an access token for JupyterHub, which can be used to identify the user and the permissions (scopes) the user has authorized for your service.

You will only get to this stage if the user has the required access:services!service=$service-name scope.

To retrieve the user model for the token, make a request to GET /hub/api/user with the token in the Authorization header. For example, using flask:

#!/usr/bin/env python3
"""
whoami service authentication with the Hub
"""
import json
import os
import secrets
from functools import wraps

from flask import Flask, Response, make_response, redirect, request, session

from jupyterhub.services.auth import HubOAuth

prefix = os.environ.get('JUPYTERHUB_SERVICE_PREFIX', '/')

auth = HubOAuth(api_token=os.environ['JUPYTERHUB_API_TOKEN'], cache_max_age=60)

app = Flask(__name__)
# encryption key for session cookies
app.secret_key = secrets.token_bytes(32)


def authenticated(f):
    """Decorator for authenticating with the Hub via OAuth"""

    @wraps(f)
    def decorated(*args, **kwargs):
        token = session.get("token")

        if token:
            user = auth.user_for_token(token)
        else:
            user = None

        if user:
            return f(user, *args, **kwargs)
        else:
            # redirect to login url on failed auth
            state = auth.generate_state(next_url=request.path)
            response = make_response(redirect(auth.login_url + '&state=%s' % state))
            response.set_cookie(auth.state_cookie_name, state)
            return response

    return decorated


@app.route(prefix)
@authenticated
def whoami(user):
    return Response(
        json.dumps(user, indent=1, sort_keys=True), mimetype='application/json'
    )


@app.route(prefix + 'oauth_callback')
def oauth_callback():
    code = request.args.get('code', None)
    if code is None:
        return 403

    # validate state field
    arg_state = request.args.get('state', None)
    cookie_state = request.cookies.get(auth.state_cookie_name)
    if arg_state is None or arg_state != cookie_state:
        # state doesn't match
        return 403

    token = auth.token_for_code(code)
    # store token in session cookie
    session["token"] = token
    next_url = auth.get_next_url(cookie_state) or prefix
    response = make_response(redirect(next_url))
    return response

We recommend looking at the [HubOAuth][huboauth] class implementation for reference, and taking note of the following process:

  1. retrieve the token from the request.

  2. Make an API request GET /hub/api/user, with the token in the Authorization header.

    For example, with requests:

    r = requests.get(
        "http://127.0.0.1:8081/hub/api/user",
        headers = {
            'Authorization' : f'token {api_token}',
        },
    )
    r.raise_for_status()
    user = r.json()
    
  3. On success, the reply will be a JSON model describing the user:

    {
      "name": "inara",
      # groups may be omitted, depending on permissions
      "groups": ["serenity", "guild"],
      # scopes is new in JupyterHub 2.0
      "scopes": [
        "access:services",
        "read:users:name",
        "read:users!user=inara",
        "..."
      ]
    }
    

The scopes field can be used to manage access. Note: a user will have access to a service to complete oauth access to the service for the first time. Individual permissions may be revoked at any later point without revoking the token, in which case the scopes field in this model should be checked on each access. The default required scopes for access are available from hub_auth.oauth_scopes or $JUPYTERHUB_OAUTH_SCOPES.

An example of using an Externally-Managed Service and authentication is in nbviewer README section on securing the notebook viewer, and an example of its configuration is found here. nbviewer can also be run as a Hub-Managed Service as described nbviewer README section on securing the notebook viewer.