Zarr from R: a capability map

There are many ways to reach a Zarr store from R. They fall into a few families:

• GDAL: the Zarr and Icechunk drivers, reached through terra, stars/sf, vapour, gdalraster and GDAL7;
• native R: zarr, pizzarr and Rarr;
• Rust-backed: the zarrs crate inside pizzarr’s r-universe build, plus Rzarrs and zr, and icechunkr;
• Arrow and byte references: zaro, blocklist and altarr;
• netCDF-C’s NCZarr: RNetCDF and ncdf4;
• domain stacks built on these: Bioconductor’s ZarrArray, anndataR, spatialdataR and ImageArray; R-CF’s geozarr; Copernicus Marine clients;
• Python through reticulate.

This repository gives a broad overview of those families. Measurement is kept light. There is one probe: a write-by-read matrix that runs locally (probe/). Everything else is prose with links: landscape.md covers every package, and gdal.md covers the GDAL tendril by GDAL version and by R binding.

Checked on 2026-11-09 with GDAL 3.14.0dev, netCDF-C 4.9.2 and R 4.6.1. It is modelled loosely on rust-proj-capability.

TL;DR: which one?

• Geospatial grids, remote stores, Kerchunk refs, Icechunk, georeferencing: GDAL. GDAL’s Zarr driver is the most complete Zarr reader R can reach:
  • v2 and v3 read and write;
  • sharding (3.13);
  • multiscales as overviews (3.13);
  • spatial:/proj: georeferencing (3.13);
  • Kerchunk JSON and Parquet references (3.11);
  • Icechunk repositories (3.14).

  The driver is the same everywhere. What differs is how much of the multidimensional API each binding exposes:

  • terra: classic API, plus md= for mdim;
  • stars: mdim, through sf;
  • vapour: classic;
  • gdalraster: mdim_* utilities on CRAN, and a full class API on the gdalmultidim-api branch;
  • GDAL7: the full mdim object model, with lazy ALTREP arrays.

  See gdal.md.

• Plain Zarr, no system libraries, CRAN only: zarr (R-CF). It is native R plus Rcpp and has full v3 core, including sharding in our probe. It reads v2, but not v2 zlib.
• Read and write v2 and v3 with an xarray-like API, S3/GCS/HTTP and an Icechunk handoff: pizzarr. Use the r-universe build for the zarrs Rust backend; the CRAN build is pure R.
• Bioconductor, DelayedArray, single-cell or OME-Zarr: Rarr, then ZarrArray, then anndataR, spatialdataR and ImageArray.
• Virtual stores (Kerchunk, VirtualiZarr manifests) without GDAL, or Arrow object stores: zaro, together with blocklist and altarr.
• netCDF-shaped code that already uses RNetCDF or ncdf4: NCZarr works, but only in narrow cases: v2 only, and without codec plugins it returns wrong values for compressed data with no error. See the probe.
• Something only Python has: reticulate with xarray/zarr-python.

The landscape

Versions are the current published ones as of 2026-11-09 (CRAN, Bioc release or devel, r-universe). Details and caveats are in landscape.md.

package	where	tech	Zarr	write	remote	notes
GDAL route
terra	CRAN 1.9-x	GDAL classic; rast(md=) mdim	v2+v3, all GDAL	yes (classic)	/vsi*	2D/3D slabs as SpatRaster
stars (+sf)	CRAN 0.7-3	GDAL mdim (read_mdim, write_mdim)	v2+v3	yes	/vsi*	arrays come back in x,y order (transposed)
vapour	CRAN 0.16	GDAL classic and warper	v2+v3	limited	/vsi*	raw read and warp of ZARR: DSNs
gdalraster	CRAN 2.7.0	GDAL classic + mdim_info, mdim_as_classic, mdim_translate	v2+v3	yes	/vsi*	full GDALMultiDimRaster class on the gdalmultidim-api branch
GDAL7	GitHub rgdal-dev/GDAL7	SWIG-derived cpp11 + S7 bindings	v2+v3	classic only	/vsi*	GDALGroup and GDALMDArray, get_view, advise_read, ALTREP as_altarr
gdalcubes	CRAN 0.7.6	GDAL classic only	per GDAL	no	/vsi*	Zarr as a 2D-band cube source
CopernicusMarine	CRAN 0.5.0	GDAL mdim, through a stars proxy	ARCO	no	HTTP/S3	needs GDAL 3.11+ with BLOSC; no native reader (0.4.6)
cmemsarco	r-universe hypertidy	GDAL classic (ZARR: DSN strings for terra/vapour)	ARCO	no	/vsicurl/, /vsis3/
native R
zarr	CRAN 1.0.1 (R-CF)	R6 + Rcpp, blosc	v3 core, reads v2	yes	memory, file, HTTP, S3	sharding; no consolidated metadata; extensible stores and codecs
geozarr	CRAN 0.3.0 (R-CF)	on zarr	per zarr	yes	per zarr	GeoZarr conventions
pizzarr	CRAN 0.2.1; r-universe 0.2.2	pure R; r-universe adds Rust zarrs	v2+v3	yes	HTTP; S3/GCS (zarrs)	IcechunkStore handoff from icechunkr
Rarr	Bioc 2.0.1 / devel 2.1.57	R + bundled C codecs	v2+v3	yes (local)	S3/HTTP (paws)	DelayedArray layer moved to ZarrArray
rnz	GitHub DOI-USGS/rnz	on pizzarr	v2+v3	no	HTTP	RNetCDF-like API
Rust-backed
Rzarrs	r-universe RGenomicsETL	zarrs via savvy	v3 (+v2)	read-first	file, HTTP, S3	VCF-Zarr, .zarr.zip, sharding
zr	GitHub mdsumner/zr	zarrs via extendr	v2+v3	local only	file, HTTP, S3	experimental
icechunkr	private (mdsumner/icechunk-r), in development	Rust icechunk via extendr	Icechunk (v3)	yes: repos, sessions, commits	local, S3, GCS, Azure, HTTP	DBI interface; chunk refs as tables; sessions handed to pizzarr
Arrow / byte references
zaro	r-universe hypertidy	R + Arrow filesystems; optional gdalraster	v2+v3, sharded	no	HTTP, S3, GCS	Kerchunk JSON/Parquet, VirtualiZarr manifests; lazy via altarr
blocklist	r-universe hypertidy	GDAL mdim mosaic + rhdf5	writes Kerchunk Parquet	refs	-	builds virtual stores from NetCDF4/HDF5
NCZarr
RNetCDF, ncdf4	CRAN	netCDF-C NCZarr	v2 (4.9.2)	yes	S3 if netCDF-C built so	file://...#mode=zarr; codecs need HDF5 filter plugins
Bioconductor / domain
ZarrArray	Bioc 1.0.1	DelayedArray on Rarr	per Rarr	-	per Rarr
anndataR	Bioc 1.2.3	Rarr (Suggests)	per Rarr	yes		AnnData .zarr
spatialdataR, ImageArray	Bioc devel / 1.0.0	Rarr + ZarrArray	per Rarr			SpatialData, OME-NGFF pyramids
romeo, OMERarr, VoltRon, ZarrDataFrame, ZarrSparseMatrix	r-universe / GitHub	mostly on Rarr				OME-Zarr and single-cell; early
lstar	CRAN 0.2.2	C++17	own v3 schema	yes		single-cell; not a general reader
Python
reticulate	CRAN	xarray, zarr-python, icechunk	everything Python does	yes	everything	plus thin wrappers: ezzarr, muschellij2/zarrr, ctreesR
gdalxarray	PyPI 0.5.1 (hypertidy)	xarray backend over GDAL multidim	all GDAL	no	/vsi*	xarray’s model on GDAL’s driver, reachable from R through reticulate; xrdbi sits on xarray too

blosc (CRAN 0.1.2) is the shared codec dependency of zarr and pizzarr.

What the probe shows

We wrote volcano (87 x 61, float64) with 10 writer configurations and read each store with 10 readers. Full grid and caveats: probe/README.md.

• Every GDAL reader read every store. That covers v2, v3, sharded, and the stores written by zarr, pizzarr and Rarr. GDAL is the reference reader in practice.
• The readers disagree on orientation, not on values.
  • stars and GDAL7 return arrays in GDAL’s fastest-dimension-first order (t).
  • terra flips rows for stores without X/Y coordinate arrays.
  • vapour, gdalraster, zarr, pizzarr, Rarr and zaro return R’s [row, col].
• Sharding: zarr, zaro and GDAL read it. pizzarr’s R path and the installed Rarr (2.1.14) refused it, with clear errors.
• v2 zlib: zarr 1.0.1 fails with attempt to set an attribute on NULL. gzip is fine.
• v3 transpose codec with partial edge chunks (as written by Rarr): pizzarr and zaro both read the wrong values.
• NCZarr (RNetCDF, netCDF-C 4.9.2):
  • it reads uncompressed v2;
  • it reads blosc/zlib/zstd v2 as garbage, with no error (no HDF5_PLUGIN_PATH);
  • it cannot see v3 arrays at all.

What no R route does yet (as far as we found)

• write Icechunk from a GDAL binding (GDAL’s Icechunk driver is read-only);
• write sharded v3 from Rarr or pizzarr’s pure-R path;
• give a CRAN-only, non-GDAL path for Kerchunk/virtual references (zaro is r-universe);
• give a single R array type that every reader returns: orientation is the binding’s choice.

Layout

• landscape.md: one entry per package, covering what it is, versions, and what we verified versus what we inferred.
• gdal.md: GDAL Zarr and Icechunk driver features by version, and how each R binding exposes them.
• probe/: probe.R, results.csv, sessioninfo.txt. Fixtures are regenerated into probe/fixtures/, which is gitignored.

To rerun the probe:

Rscript probe/probe.R