Virtual Icechunk
A short tutorial for the 2026 IOOS Hackathon
This tutorial gives you a quick overview of virtual Icechunk stores: what they are, how one gets built, and what to check before building one. During the hackathon you will mostly hand the actual work to a coding agent that has the ocean-icechunks-skills installed. Each chapter is a short slide deck. Click through it, or press S for the speaker notes.
Before you start: get set up
New to the hub? Work through JupyterHub Skills first (about 30 minutes):
- Log in to the JupyterHub
- Find your way around JupyterLab
- Git and GitHub, with an agent: log in once with
gh, then let the agent run git for you - The scratch bucket: cloud storage for test stores
Then start your coding agent: Claude Code, Codex or Antigravity, with your own subscription.
Lectures
| Deck | What it covers | |
|---|---|---|
| 1 | Why virtual Icechunk | A product, not a pile of files; both quick wins |
| 2 | Working with an agent | One task per session; installing and using ocean-icechunks-skills |
| 3 | The basic pipeline | Six steps from NetCDF files to a store anyone can open |
| 4 | Auditing your files | What to check before anyone writes build code |
| 5 | Building a store | What the skill needs from you, what it works out, and how to prompt it |
| 6 | Visualizing | xarray, then Zarr in the browser: zarrita, icechunk-js and gridlook |
Examples
Each example opens a large ocean dataset as a virtual Icechunk store and answers one question with it in a few seconds or minutes. You download nothing in advance, even when the dataset is hundreds of terabytes or thousands of files. Use them to see how a store opens, how much a read costs, and what kind of analysis becomes easy. Each one has an Open in Colab button.
See Resources for links to everything mentioned in the decks.