Document running the RAPIDS Notebooks container as a Snowpark Service

[jacobtomlinson]

The Snowflake documentation has an example of running a container with Jupyter in as a service in Snowpark and accessing the Jupyter UI via a Service endpoint.

It would be great to add a documentation page that walks through the same steps but using the RAPIDS notebooks container on a GPU compute pool.

I imagine the steps will be something along the lines of:

• Build an image based on the RAPIDS notebooks image and install the snowflake Python library
• Push the image to the Snowflake container registry
• Create a service specification with the RAPIDS container
• Launch the service
• Expose/access the Jupyter UI
• Show reading data from Snowflake into a cudf dataframe

[ncclementi]

I looked a bit more into this, and it looks like the steps are pretty much what's detailed. Couple of notes and things to figure out are:

Here is a walkthrough video that also covers part of the rpocess: https://youtu.be/izeG3pvvy9M?list=PLavJpcg8cl1GN7AwaY95PXqwUw4qj8zLo&t=411

Snowflake setup:

TODO:

• Access to non-trial Snowflake account. It looks like to do most of this, it's needed A Snowflake account login with a role that has the ACCOUNTADMIN
• Setup snowflake env https://quickstarts.snowflake.com/guide/intro_to_snowpark_container_services/#1

local env setup/config:

• configure snowflake cli and create a connection to the container
• start docker and check login into image repo registry
  ref: https://quickstarts.snowflake.com/guide/intro_to_snowpark_container_services/#2

Build an image based on the RAPIDS notebooks image and install the snowflake Python library

TODO:

• Determine what extra dependencies need to be added (probably "snowflake-snowpark-python[pandas]" snowflake-connector-python ? )
• Create docker file that is based on the RAPIDS notebook image, create the image

For what I could gather from the demos, it seems like you build the image locally and then push it to the container registry. I was hoping one would be able to pull from let's say dockerhub or similar directly into snowpark registry but I can't find docs on that. Which it might be a bit limiting for a doc example

Push the image to the Snowflake container registry

From the demo docs steps are the image is build locally and they use snowpark CLI + docker to push it to the registry.

Create a service specification with the RAPIDS container
Launch the service
Expose/access the Jupyter UI

This should be almost identical to the process here https://quickstarts.snowflake.com/guide/intro_to_snowpark_container_services/#3 except some rapids caveats that we will find out.

Show reading data from Snowflake into a cudf dataframe

Do we need to create a snowflake table then, is there any snowflake table we can "copy"? what do we have in mind here?

[jacobtomlinson]

Closed by #493