intake-virtual-icechunk

An intake plugin for building and reading Icechunk stores via VirtualiZarr and intake-esm.

Concept

The goal is a pipeline that takes a pre-built intake-esm datastore and produces a single virtual Icechunk store that mirrors its structure:

1. Open a pre-built intake-esm datastore with intake-esm.
2. For each dataset in the catalog, open the constituent files with VirtualiZarr to create virtual references — no data is copied.
3. Write each dataset as a named Zarr group inside one Icechunk store, using the catalog's groupby_attrs to derive the group name.
4. Expose the result through an intake driver (virtual_icechunk) that hides all Icechunk-specific complexity (sessions, stores, branches) behind an interface that feels like a hybrid of an esm-datastore and an xarray.Dataset — defaulting to Xarray semantics wherever possible, and falling back to esm-datastore conventions only where necessary (e.g. catalog search and group selection).

The end result is one Icechunk store, one group per dataset, fully virtual (no data duplication), and accessible via intake.open_virtual_icechunk().

This package provides two things

1. Building (IcechunkStoreBuilder) — given a pre-built intake-esm catalog, creates virtual references with VirtualiZarr and writes each dataset as a named Zarr group inside a single Icechunk store.
2. Reading (IcechunkSource) — an intake driver for opening a group from an Icechunk store as an xarray.Dataset via intake.open_virtual_icechunk().

Installation

pip install intake-virtual-icechunk

License

Apache-2.0. See LICENSE for details.