jhsingle-native-proxy

Wrap an arbitrary webapp so it can be used in place of jupyter-singleuser in a JupyterHub setting.

Within JupyterHub this allows similar operation to jupyter-server-proxy except it also removes the Jupyter notebook itself, so is working directly with the arbitrary web service.

OAuth authentication is enforced based on JUPYTERHUB_* environment variables.

This project is used in ContainDS Dashboards, which is a user-friendly way to launch Jupyter notebooks as shareable dashboards inside JupyterHub. Also works with Streamlit and other visualization frameworks.

Install and Run

Install using pip.

pip install jhsingle-native-proxy

The process to start is specified on the command line, for example a streamlit web app:

jhsingle-native-proxy streamlit hello

By default the jhsingle-native-proxy server will listen on port 8888, forwarding to port 8500.

But you will normally need to tell jhsingle-native-proxy which port the end process will run in, and maybe tell the end process which port you want it to use (which you can do with the substitution variable {port}).

Note the use of -- to signal the end of command line options to jhsingle-native-proxy. Then the third party command line itself can contain options starting with dashes. An alternative is to use the substitution {--}

jhsingle-native-proxy -- streamlit hello --server.port {port} --server.headless True --server.enableCORS False

To run jhsingle-native-proxy itself listening on a different port use:

jhsingle-native-proxy --port 8000 streamlit hello

To run jhsingle-native-proxy on port 8000, and the end process on 8505:

jhsingle-native-proxy --port 8000 --destport 8505 -- streamlit hello --server.port {port} --server.headless True --server.enableCORS False

Use the JUPYTERHUB_SERVICE_PREFIX env var to specify the first part of the URL to listen to (and then strip before forwarding). E.g. JUPYTERHUB_SERVICE_PREFIX=/user/dan will mean requests on http://localhost:8888/user/dan/something will forward to http://localhost:8500/something

You can also specify --ip 0.0.0.0 for the address to listen on.

Below we use the substitution {--} for the command to run, allowing us to specify --ip to jhsingle-native-proxy instead of the command being run.

jhsingle-native-proxy --port 8000 --destport 8505 streamlit hello {--}server.port {port} {--}server.headless True {--}server.enableCORS False --ip 0.0.0.0 

Similarly, use e.g. {-}m to represent -m in the final command.

Voila example:

Running voila at the subfolder URL e.g. /user/dan/:

python -m jhsingle_native_proxy.main --destport 0 voila ./Presentation.ipynb {--}port={port} {--}no-browser {--}Voila.server_url=/ {--}Voila.base_url={base_url}/ {--}debug

'destport 0' above instructs jhsingle-native-proxy to choose a random free port on which to run the sub-process (Voila), and of course substitutes that as {port} in the Voila command line so it knows which port to listen on. destport 0 is the default anyway.

Or specify presentation_path as a substitution instead of hard-coding, which is sometimes easier in your wrapper code:

python -m jhsingle_native_proxy.main --destport 0 voila {presentation_path} {--}port={port} {--}no-browser {--}Voila.server_url=/ {--}Voila.base_url={base_url}/ {--}debug --presentation_path=./Presentation.ipynb

In addition, if presentation_path is provided, two further substitution variables are available: presentation_dirname and presentation_basename. These are computed using Python's os.path.dirname and os.path.basename functions on presentation_path.

Authentication

The above examples all assume OAuth will be enforced, as per the JUPYTERHUB_* env vars.

Alternatives can be specified via the authtype flag:

Same as default:

jhsingle-native-proxy --authtype=oauth streamlit hello

No auth required at all:

jhsingle-native-proxy --authtype=none streamlit hello

Specifying Authorized Users

The env vars JUPYTERHUB_USER and JUPYTERHUB_GROUP can be used, as typical for any JupyterHub single server, to specify user/groups of JupyterHub that should be allowed access via OAuth. There is an additional bespoke env var called JUPYTERHUB_ANYONE which can be set to 1 to allow any authenticated user access. (i.e. anyone who has an account on the JupyterHub)

Extra Arguments

--request-timeout=300 specifies the timeout in seconds that it waits for the underlying subprocess to return when proxying normal requests. Default is 300.

{origin_host} in the command argument will be replaced with the first 'host' seen in any request to the jhsingle-native-proxy server.

--last-activity-interval=300 specifies how often in seconds to update the hub to provide the last time any traffic passed through the proxy (default 300). Specify 0 to never update. --force-keep-alive or --no-force-keep-alive: the former (default) ensures that the hub is notified of recent activity even if there wasn't any - only works if last-activity-interval is not 0.

--ready-check-path (default /) to change the URL on the subprocess used to poll with an HTTP request to check for readiness.

--repo - use git to check out a repo before running the sub process

--repofolder - the path of a folder (to be created if necessary) to contain the git repo contents

--forward-user-info - forward to the underlying process an X-CDSDASHBOARDS-JH-USER HTTP header containing JupyterHub user info as JSON-encoded string

--query-user-info - add a GET query param named CDSDASHBOARDS_JH_USER when calling underlying process containing JupyterHub user info as JSON-encoded string

--ready-timeout - integer timeout for period of checking the process is running at startup (default 10). Increase if your process is not able to return anything at --ready-check-path until a longer time after it first starts up. Be aware that the process must (once ready) return its HTTP response within 1 second. Note this argument is different from --request-timeout which applies to individual HTTP proxy calls during normal operation (not just at startup).

--websocket-max-message-size - message size in bytes allowed by websocket connections made to the underlying process (default is to rely on the tornado library defaults).

Changelog

v0.7.5 released 20 Apr 2021

• New command-line options --ready-timeout and --websocket-max-message-size

v0.7.3 released 9 Apr 2021

• New command-line option --progressive to flush buffer from underlying service whenever chunks appear (this is useful to see results from Voila sooner)
• oauth_callback URL now accessible when running with JUPYTERHUB_BASE_URL of /

v0.7.1 released 22 Feb 2021

• New command-line option --query-user-info to add a CDSDASHBOARDS_JH_USER GET query param to the http request to the underlying service.

v0.7.0 released 12 Feb 2021

• New command-line option --forward-user-info to add a X-CDSDASHBOARDS-JH-USER header to the http request to the underlying service. The header value is a JSON-encoded dict containing kind, name, admin, groups fields from the logged-in JupyterHub user if available.

v0.6.1 released 6 Jan 2021

• Require simpervisor >= 0.4 to ensure Python 3.9 compat.

v0.6.0 released 20 Nov 2020

• Displays INFO level logs by default, which includes output of the subprocess (turn off with --no-logs) Issue #7
• Logs from subprocess written out at different level depending on source (stderr -> error, stdout -> info)
• Long subprocess logs are handled and truncated instead of throwing an error cdsdashboards issue #44
• Different handling of branch checkout when using git repo source, when switching brances compared to what was checked out before

v0.5.6 released 18 Sep 2020

• Always convert presentation_path to an absolute path (based on CWD) before passing to the sub-command.

v0.5.5 released 10 Sep 2020

• Also accept URLs at the URL-encoded equivalent of the prefix and redirect to the regular version of the URL.

v0.5.4 released 3 Sep 2020

• Change working folder to repofolder when specified

v0.5.2 released 17 Aug 2020

• Require tornado 6.0.4+

v0.5.1 released 17 Aug 2020

• Fix to ensure both websockets are opened at the same time, to avoid writing to a websocket that's not yet open.

v0.5.0 released 17 Aug 2020

• Open up underlying process' websocket before connecting our own with the client. This ensures any other GET headers can be passed back to the client. (Fix for Streamlit XSRF problems.)

v0.4.3 released 30 July 2020

• Added --allow-root option (currently ignored) to avoid errors if this flag is usually passed to jupyter-singleuser

v0.4.2 released 23 July 2020

• Switch to a Conda env before running subprocess by specifying --conda-env option

v0.4.1 released 20 July 2020

• fix because subprocess sometimes blocked if too much output generated

v0.4.0 released 15 July 2020

• repo and repofolder optional arguments added

v0.3.2 released 25 June 2020

v0.3.1 released 18 June 2020

• Defaults presentation_path to empty str ('') if not supplied, avoiding error

v0.3.0 released 17 June 2020

• presentation_path can be provided as a command line argument to become a substitution variable.
• presentation_basename and presentation_dirname are also available when presentation_path is supplied.

v0.2.0 released 11 June 2020

• Better websocket handling (subprotocols)
• {origin_host} variable added

v0.1.3 released 1 June 2020

• request-timeout added to the proxy call, and the default set to 300 (20 seconds was the httpclient's default previously)

v0.1.2 released 29 May 2020

• Now allows single-dash placeholder, e.g. {-}m translates to -m in the final subprocess command.

Development install

git clone https://github.com/ideonate/jhsingle-native-proxy.git
cd jhsingle-native-proxy

pip install -e .

To run directly in python: python -m jhsingle_native_proxy.main <rest of command line>

Testing git puller:

python -m jhsingle_native_proxy.main --authtype=none --destport=0 --port=8888 voila ./sincosfolder/Presentation.ipynb {--}port={port} {--}no-browser {--}Voila.server_url=/ {--}Voila.base_url={base_url}/ --repo=https://github.com/danlester/binder-sincos --repofolder=sincosfolder