After my lightning talk at CNG Zürich last Thursday, a friend caught me with a question on the next day: "Okay, but why is Zarr actually useful for us?" He has been working with weather data for more than 20 years and knows NetCDF, so I didn't want to start explaining "cloud-native formats" again.
Explaining via a custom-tailored interactive web page 🔗
It's always better to explain with real world examples instead of reference documentation or abstract tutorials. I couldn't find a "NetCDF to Zarr" tutorial, so that evening I sat down with Claude and built an interactive web page to demonstrate the ease of use:
The page basically explains what chunking means in Zarr by querying different models using weatherzarr.com1. Querying all models for Basel means downloading only 3.9MB instead of 159GB for all datasets!
Narrowing the comprehension gap 🔗
For the geospatial world (or at least for me) there is much more to understand about Zarr and CNG in general. I believe that quick, custom-made explainer pages can help narrow the "comprehension gap" with the great retooling of CNG technologies:
Ian Schuler is the CEO of Development Seed — the team behind a lot of the open source tooling that quietly holds up the geospatial world. He's been at the helm for over a decade, and in this conversation, we dig into what he calls the great retooling: the idea that cloud-native geospatial is about to flip from an emerging pattern to the dominant one, and that AI is the thing tipping it over the edge.
The argument is simple — agents want to discover your data, query it, transform it, and hand back an answer. If your data isn't in a format they can reach, you're simply not part of the conversation anymore.
The point of the explainer page seems very simple: it only shows that Zarr's chunking eases the accessibility of data using simple HTTP requests. This looks small, but it's one core part of why the geospatial world is changing: I agree with Ian: Both the ease of data retrieval and AI will change the geospatial world. We need to make sure that we humans keep up with understanding what's going on.
AI Colophon 🔗
The post was drafted by Stephan and improved with a workshop skill where Claude only pointed out flaws, but did not edit any text directly (other than typos).
For the explainer, I described to Claude what I assumed someone knows about
NetCDF, and Claude built the interactive demos and the page's styling. It even
invented its own miniZarr JS library instead of using zarrita.js or
similar. The cover image is a small diagram I had Claude draw as plain SVG,
rendered to PNG.
-
Weatherzarr is a weather-data service built in Atlanta. It converts forecasts from various models into Zarr, a format that lets tools such as Python’s xarray read just the data they need over the web, without downloading huge GRIB files or requiring API keys. I actually do not know who runs this website and it may be out-of-service soon, but it was the first site I found where I easily could query multiple weather models at once. ↩