esgf-virtual-zarr-data-access

ESGF working group to enable data access via virtual zarrs.

Motivation

We aim to establish streaming access via zarr as a officially supported access pattern next to local downloading. ESGF/esgf-roadmap#5 provides more justification.

This effort draws heavily from the experience of the Pangeo / ESGF Cloud Data Working Group We aim to do this:

• Without duplicating most/all of the data
• Serving this via official ESGF channels (next to netcdf files).

Guide

1. Install the required dependencies via pip

mamba create -n esgf-virtual-zarr-data-access python=3.11
mamba activate esgf-virtual-zarr-data-access
pip install -r requirements.txt

2. Modify the urls, and the output json filename in virtual-zarr-script.py, and run the script.

python virtual-zarr-script.py

3. Check that the generated JSON file is readable with xarray and average the full dataset (this is also done in the script)

import xarray as xr
ds = xr.open_dataset(
    '<your_filename>.json', 
    engine='kerchunk',
    chunks={},
)
ds.mean().load() # test that all chunks can be accessed.

Goals

On the Tenth Earth System Grid Federation (ESGF) Hybrid Conference we discussed the option to serve virtualized zarr files (kerchunk reference files for demonstration's sake). We saw an excellent demo by @rhysrevans3 who showed how to serve both the virtual zarr and the individual netcdf files as a STAC catalog.

• Our goal here is to show the feasability of streaming access to CMIP data via ESGF infrastructure for CMIP6+ and beyond (CMIP7)

Milestones

• Demonstrate a proof of concept which shows that we can create virtual zarr stores, which point to data on ESGF servers, that work functionally equivalent to loading a native zarr from cloud storage.
  • First working prototype
• Prototype how the generation of reference files is integrated into the future ESGF publishing

Why not Kerchunk?

• I (@jbusecke) strongly suggest to use Virtualizarr to future proof the development here:
  • Is currently backwards compatible with kerchunk.
  • Save out native Zarr. This will enable users to use any Zarr client library to read the data (needs V3?)
  • Reference creation in xarray native, enables cleaner API.

Open Questions

Upstream Requirements