Create custom kernel via post-startup script in Vertex AI User Managed notebook

user6135514 0
Viewed 85

I am trying to use a post-startup script to create a Vertex AI User Managed Notebook whose Jupyter Lab has a dedicated virtual environment and corresponding computing kernel when first launched. I have had success creating the instance and then, as a second manual step from within the Jupyter Lab > Terminal, running a bash script like so:

#!/bin/bash
cd /home/jupyter
mkdir -p env
cd env
python3 -m venv envName --system-site-packages
source envName/bin/activate
envName/bin/python3 -m pip install --upgrade pip
python -m ipykernel install --user --name=envName
pip3 install geemap --user 
pip3 install earthengine-api --user 
pip3 install ipyleaflet --user 
pip3 install folium --user 
pip3 install voila --user 
pip3 install jupyterlab_widgets
deactivate
jupyter labextension install --no-build @jupyter-widgets/jupyterlab-manager jupyter-leaflet
jupyter lab build --dev-build=False --minimize=False
jupyter labextension enable @jupyter-widgets/jupyterlab-manager

However, I have not had luck using this code as a post-startup script (being supplied through the console creation tools, as opposed to command line, thus far). When I open Jupyter Lab and look at the relevant structures, I find that there is no environment or kernel. Could someone please provide a working example that accomplishes my aim, or otherwise describe the order of build steps that one would follow?

gcp-ai-platform-notebook jupyter-lab google-cloud-vertex-ai google-cloud-platform
2 Answers

Post startup scripts run as root. When you run:

python -m ipykernel install --user --name=envName

Notebook is using current user which is root vs when you use Terminal, which is running as jupyter user.

Option 1) Have 2 scripts:

  • Script A. Contents specified in original post. Example: gs://newsml-us-central1/so73649262.sh
  • Script B. Downloads script and execute it as jupyter. Example: gs://newsml-us-central1/so1.sh and use it as post-startup script.
#!/bin/bash

set -x

gsutil cp gs://newsml-us-central1/so73649262.sh /home/jupyter
chown jupyter /home/jupyter/so73649262.sh
chmod a+x /home/jupyter/so73649262.sh
su -c '/home/jupyter/so73649262.sh' jupyter

Option 2) Create a file in bash using EOF. Write the contents into a single file and execute it as mentioned above.

gogasca0

This is being posted as support context for the accepted solution from @gogasca.

@gogasca's suggestion (I'm using Option 1) works great, if you are patient. Through many attempts, I discovered that inconsistent behavior was based on timing of access. Using Option 1, the User Managed Notebook appears available for use in Vertex AI Workbench (green check and clickable "OPEN JUPYTERLAB" link) before the installation script(s) have finished.

If you open the Notebook too soon, you will find two things: (1) you will be prompted for a recommended Jupyter Lab build, for instance:

Build Recommended
JupyterLab build is suggested:
@jupyter-widgets/jupyterlab-manager changed from file:../extensions/jupyter-widgets-jupyterlab-manager-3.1.1.tgz to file:../extensions/jupyter-widgets-jupyterlab-manager-5.0.3.tgz

and (2) while the custom environment/kernel is present and accessible, if you try to use ipyleaflet or ipywidget tools, you will see one of several JavaScript errors, depending on how quickly you try to use the kernel, relative to the build that is (apparently) continuing to take place in the background: Error displaying widget: model not found, and/or a broken page icon with a JavaScript error, that, if clicked, will show you something like:

[Open Browser Console for more detailed log - Double click to close this message]
Failed to load model class 'LeafletMapModel' from module 'jupyter-leaflet'
Error: No version of module jupyter-leaflet is registered
at f.loadClass (https://someURL.notebooks.googleusercontent.com/lab/extensions/@jupyter-widgets/jupyterlab-manager/static/134.bcbea9feb6e7c4da7530.js?v=bcbea9feb6e7c4da7530:1:74856)
at f.loadModelClass (https://someURL.notebooks.googleusercontent.com/lab/extensions/@jupyter-widgets/jupyterlab-manager/static/150.3e1e5adfd821b9b96340.js?v=3e1e5adfd821b9b96340:1:10729)
at f._make_model (https://someURL.notebooks.googleusercontent.com/lab/extensions/@jupyter-widgets/jupyterlab-manager/static/150.3e1e5adfd821b9b96340.js?v=3e1e5adfd821b9b96340:1:7517)
at f.new_model (https://someURL.notebooks.googleusercontent.com/lab/extensions/@jupyter-widgets/jupyterlab-manager/static/150.3e1e5adfd821b9b96340.js?v=3e1e5adfd821b9b96340:1:5137)
at https://someURL.notebooks.googleusercontent.com/lab/extensions/@jupyter-widgets/jupyterlab-manager/static/150.3e1e5adfd821b9b96340.js?v=3e1e5adfd821b9b96340:1:6385
at Array.map ()
at f._loadFromKernel (https://someURL.notebooks.googleusercontent.com/lab/extensions/@jupyter-widgets/jupyterlab-manager/static/150.3e1e5adfd821b9b96340.js?v=3e1e5adfd821b9b96340:1:6278)
at async f.restoreWidgets (https://someURL.notebooks.googleusercontent.com/lab/extensions/@jupyter-widgets/jupyterlab-manager/static/134.bcbea9feb6e7c4da7530.js?v=bcbea9feb6e7c4da7530:1:77764)

The solution here is to keep waiting. In my demo script, I transfer a file at the end of the build process. If I wait long enough for this file to actually appear in the Instance directories, the recommendation for a rebuild is absent and the extensions work properly.

user61355140
Related