Landscape: every R route to Zarr we found

Checked on 2026-11-09. Each entry gives where the package lives, what it is built on, and what it does with Zarr. Facts are marked:

• (V) verified, from a DESCRIPTION, README or NEWS, the package’s exports, or the probe;
• (I) inferred, not checked on a page.

Sources:

• the r-universe search API (q=zarr, 34 hits);
• CRAN available.packages();
• the Bioconductor 3.23 and 3.24 VIEWS files;
• GitHub repositories;
• packages installed locally.

Versions in the probe can trail the published ones. Those are noted where they matter.

1. GDAL route

These packages all go through the same GDAL Zarr and Icechunk drivers, so their spec coverage is GDAL’s (see gdal.md). They differ in two things:

• whether they use the classic raster API (2D bands, with extra dimensions flattened into bands or subdatasets) or the multidimensional API (groups, arrays, dimensions, attributes);
• the orientation of the arrays they return.

terra

• CRAN; 1.9-57 was used locally.
• rast() uses the classic API by default. Its signature includes md and dims (V), which open through the multidim API and pick dimensions.
• sds() lists subdatasets. writeRaster() writes Zarr through GDAL CreateCopy (I).
• Probe: reads every store (V). The orientation is right when X/Y coordinate arrays exist, and flipped in rows when they don’t (V).

stars, with sf

• CRAN stars 0.7-3 and sf 1.1-3.
• stars::read_mdim() and write_mdim() sit on sf::gdal_read_mdim() and gdal_write_mdim() (V, exports).
• read_mdim() uses the GDAL multidim API and supports offset, count, step and proxy (V, signature).
• Arrays come back in x,y (GDAL fastest-first) order: transposed relative to R’s [row, col] (V, probe).

vapour

• CRAN 0.16.
• Classic API through vapour_read_raster(), vapour_raster_info() and vapour_sds_names(), plus the warper (gdal_raster_data(), vapour_warp_raster()).
• Probe: reads every store in R orientation (V).
• gdal_raster_data() refused stores without georeferencing (they “could not be processed with GDALWarp api”). Use vapour_read_raster() for those (V).

gdalraster

• CRAN 2.7.0, from the firelab maintainers. The probe used 2.7.0.9002.
• Classic GDALRaster plus multidim utilities: mdim_info(), mdim_as_classic() and mdim_translate() (V, exports).
• mdim_as_classic(dsn, array_name, idx_xdim, idx_ydim) gives a classic GDALRaster view of any 2D slice of an MDArray.
• Probe: reads every store in R orientation (V).
• The gdalmultidim-api branch (mdsumner/gdalraster, last merged with main 2026-06-16) adds a full Rcpp-module class GDALMultiDimRaster (V, R/gdalmultidim.R) and about 6,100 lines (V, diff stat):
  • supporting classes GDALGroupR, GDALMDArrayR, GDALDimensionR, GDALAttributeR and GDALExtendedDataTypeR;
  • methods $getRootGroup(), $getArrayNames(), $openArray(), $openArrayFromFullname(), $openSubGroup(), $createGroup(), metadata and $info();
  • functional wrappers mdim_read(), mdim_read_info(), mdim_read_dim_values(), mdim_read_coord_info(), mdim_read_array_attrs() and mdim_read_group_attrs().

  Related branches: multidimnew, as-classic-mdim-factory, multidim. The branch is not installed here and was not probed.

GDAL7

• GitHub rgdal-dev/GDAL7, 0.1.0, active (last merge 2026-11-08).
• Bindings generated from GDAL’s own SWIG interface files, bound with cpp11 and S7 (V, README).
• The multidim surface is gdal_open(path, multidim = TRUE), get_root_group(), open_group(), open_mdarray(), mdarray_info(), read_mdarray(x, start, count, step), get_view(arr, "[0,:,:]") and as_mdarray() (V).
• advise_read() sends GDAL a batch read hint, so Zarr decodes the chunks it touches on GDAL_NUM_THREADS threads (V, docs).
• as_altarr() gives a lazy, chunk-cached ALTREP array (V).
• It ships a multidim.zarr example.
• Probe: reads every store. It returns GDAL dimension order (t) (V).

gdalcubes

• CRAN 0.7.6.
• Classic API only. Cubes are built from GDAL image collections of 2D bands, so a Zarr store is just another classic raster source (I).

Copernicus Marine clients

• CopernicusMarine (CRAN 0.5.0; 0.4.6.0002 checked locally) uses the multidim API, through stars. It has no Zarr reader of its own in 0.4.6. Evidence (V, function bodies):
  • cms_download_subset() and cms_zarr_proxy() open a stars mdim proxy on a /vsi* URI, subset by dimension values, then call st_as_stars();
  • it sets GDAL_NUM_THREADS=ALL_CPUS, GDAL_HTTP_MULTICURL=YES and GDAL_DISABLE_READDIR_ON_OPEN=EMPTY_DIR;
  • it requires GDAL 3.11 or later with BLOSC, and has_blosc() checks for it.

  The same author wrote the CRAN blosc package. Earlier versions may have decoded chunks themselves (I, not checked). CRAN 0.5.0 was not checked.

• cmemsarco (hypertidy, r-universe 0.1.0) is classic API in practice. It only builds ZARR:"/vsicurl/<url>":/<array> and /vsis3/ strings, and suggests terra and vapour to read them (V, source).

2. Native R

zarr (R-CF)

• CRAN 1.0.1 (2026-11-08); the probe used 1.0.1.
• It is a “native and extensible R driver”, written in R6 and Rcpp, with blosc (V).
• Full v3 core (V, DESCRIPTION). It reads v2.
• Stores: zarr_localstore, zarr_memorystore and zarr_httpstore, plus S3 (V, exports and agent).
• Conventions API: zarr_conventions(), zarr_convention_uom() and others (V).
• No consolidated metadata (V, the print method shows ✖ Consolidated metadata).
• Probe:
  • reads v2 and v3 from every writer, including GDAL-sharded stores (V);
  • fails on v2 zlib with attempt to set an attribute on NULL. Stores from both GDAL and Rarr hit this (V).
• Default write layout: v3, with transpose + bytes + blosc(zstd, bitshuffle), chunk keys c.0.0, and fill value 9.969e36, which is netCDF’s default (V, fixture).
• Reverse dependency: geozarr.

geozarr (R-CF)

• CRAN 0.3.0 (2026-10-30), built on zarr.
• Implements the GeoZarr domain conventions (V, repo description).
• Not probed.

pizzarr

• From zarr-developers. CRAN 0.2.1 (2026-08-22); r-universe 0.2.2 (V).
• Two builds:
  • the CRAN build is pure R;
  • the r-universe build links the Rust zarrs crate via extendr. It adds parallel decode, more codecs, object stores, S3, GCS and Icechunk (V, README and pizzarr_compiled_features()).
• API: zarr_open(), zarr_create_group(), zarr_create_array(), DirectoryStore, HttpStore, S3Store, GcsStore, IcechunkStore and slice() (V).
• Probe, using the r-universe 0.2.2 build:
  • it reads v2 and v3 from every writer, except:
  • sharded v3 gives Unsupported V3 codec: 'sharding_indexed' (the R codec path, even with zarrs compiled in);
  • Rarr’s v3 store (transpose with 32x32 chunks over 87x61) reads wrong values;
  • zarr’s single-chunk transpose store reads correctly (V).
• Default v2 write layout: order: "F" and zstd. Every reader handled the F order except NCZarr, which needs codec plugins anyway (V).
• Its CRAN reverse dependencies: none on CRAN. rnz uses it (V).

Rarr

• Bioconductor, from the Huber group at EMBL. Release 3.23 is 2.0.1, devel 3.24 is 2.1.57, and r-universe has 2.1.59. The probe used 2.1.14, which is older than devel.
• R with bundled C codecs: blosc, zlib, gzip, bz2, lzma, lz4 and zstd (use_*()) (V).
• read_zarr_array(), write_zarr_array(zarr_version = 2L|3L), zarr_overview(), zarr_consolidate_metadata() and attribute helpers (V).
• Remote: S3 and HTTP via paws.storage and curl. No S3 writes and no sharded writes yet (V, agent).
• The DelayedArray layer moved to ZarrArray (V).
• Probe:
  • Rarr 2.1.14 reads every store except GDAL-sharded, where it says The following codecs are not supported (V). Newer versions may differ.
• Bioc reverse dependencies (devel): anndataR, ImageArray, spatialdataR, spicyR, ZarrArray.

rnz

• GitHub DOI-USGS/rnz, forked from mdsumner/rnz.
• An RNetCDF-style API (z_demo() and similar) over pizzarr.
• v2 and v3 (NZ-1.0), over HTTP (V, README/agent).

3. Rust-backed

Rzarrs

• r-universe RGenomicsETL, 0.1.0 (2026-09-07).
• The zarrs crate via savvy.
• v3 plus compatible v2, read first.
• Stores: file, HTTP and S3, with GCS and Azure optional.
• Codecs: bytes, gzip, zstd, crc32c, blosc, sharding and transpose.
• Aimed at .zarr.zip VCF-Zarr (V, README/agent).

zr

• GitHub mdsumner/zr, 0.0.1.
• zarrs via extendr: v2 and v3 read over file, HTTP and S3, with local write only (V, README).
• Last commit (2026-03-19): “hit up against icechunk limitation”.
• The README install line points at hypertidy/zr, which returns 404 (V).
• The locally installed copy is an older stub, so it was not probed.

icechunkr

• In development in a private repo, mdsumner/icechunk-r (0.0.0.9000).
• Bindings to the Icechunk Rust library: storage (local, S3, GCS, Azure, HTTP, memory), repos, branches, tags, sessions, commits, rebase/merge, and virtual chunk refs with explicit container authorisation (V, DESCRIPTION and exports).
• A DBI interface (dbConnect() and friends).
• ic_chunk_refs() and ic_refs_arrow() read a snapshot’s chunk references as a table. ic_to_byterefs() and ic_repo_from_byterefs() round-trip Kerchunk-style refs.
• Sessions are handed to pizzarr::IcechunkStore as bytes (ic_session_bytes()).
• This is the only R route that writes Icechunk without Python.

4. Arrow and byte references (hypertidy)

zaro

• r-universe hypertidy, 0.0.1.9036; the probe used 0.0.1.9030.
• “Zarr via Arrow”: Arrow filesystems and codecs, with optional gdalraster and nanoparquet.
• v2 and v3 with automatic detection, consolidated metadata and sharding (V, README and probe).
• Stores: HTTP, S3 and GCS.
• Virtual stores: Kerchunk JSON and Parquet, and VirtualiZarr manifests (V).
• API: zaro(), zaro_read(), zaro_array(), zaro_plan(), zaro_chunks() and others. Laziness comes from altarr.
• Probe:
  • reads every store in R orientation, including sharded (V);
  • except transpose-codec stores (both zarr’s and Rarr’s), which read wrong values (V).

altarr, ndr

• altarr (hypertidy, 0.1.1) provides lazy chunked arrays as ALTREP vectors that carry a dim but no class.
• ndr adds named-dimension arrays with coordinate indexing.
• Neither is a Zarr reader. They are the lazy array layer that zaro and GDAL7 both use.

blocklist, rustycogs, sds

• blocklist builds Kerchunk Parquet reference stores from NetCDF4/HDF5 collections, using gdal mdim mosaic and rhdf5 (V, README).
• rustycogs virtualises TIFF/COG tiles into byte refs with Rust (V, README).
• sds is a catalogue of spatial data sources.

5. NCZarr: RNetCDF and ncdf4

• netCDF-C has had NCZarr since 4.8. RNetCDF::open.nc() and ncdf4::nc_open() open "file:///abs/path.zarr#mode=zarr,file". Use mode=nczarr for netCDF-flavoured stores, and s3 instead of file when netCDF-C was built with S3.
• Probe (RNetCDF 2.11-2, netCDF-C 4.9.2, no HDF5_PLUGIN_PATH) (V):
  • uncompressed v2 reads correctly. Dimensions arrive reversed, as usual for netCDF in R;
  • compressed v2 (blosc, zlib, zstd) reads garbage with no error: the filter plugins are not loaded;
  • v3 arrays are not found at all;
  • pure-Zarr arrays without _ARRAY_DIMENSIONS get synthetic _zdim_N dimensions.
• Binary builds that bundle netCDF-C and plugins may behave better. We did not check them.

6. Bioconductor and domain stacks

• ZarrArray (Bioc 1.0.1; devel 1.1.8) is a DelayedArray backend on Rarr (V).
• anndataR (Bioc 1.2.3, from scverse) has a Zarr-backed AnnData through Rarr, in Suggests (V).
• spatialdataR (Bioc devel 0.99.47) handles SpatialData/OME-NGFF using Rarr, ZarrArray and duckdb. Companions: SpatialData.data and SpatialData.plot (V).
• ImageArray (Bioc 1.0.0) provides pyramidal images over Rarr, ZarrArray and rhdf5 (V).
• OME-Zarr:
  • romeo (huber-group-embl r-universe, 0.99.2) reads, writes and validates multiscale OME-Zarr (V);
  • OMERarr (GitHub BIMSBbioinfo) is based on ome-zarr-models.py;
  • VoltRon (BIMSB, 0.3.2) is not on Bioc and uses ImageArray (I).
• ZarrDataFrame (BIMSB r-universe) is undocumented.
• ZarrSparseMatrix (GitHub GabrielHoffman) holds CSR/CSC AnnData X on Rarr. Its README says Bioc, but it is in neither VIEWS file.
• lstar (CRAN 0.2.2) is a C++17 single-cell store with its own v3 schema. It is not a general Zarr reader.
• Not Zarr, though often assumed:
  • zellkonverter (H5AD through Python);
  • RBioFormats;
  • rhdf5/HDF5Array, which are siblings;
  • ncdfCF and CFtime, which belong to the same R-CF org as zarr but have no Zarr in their DESCRIPTION (V);
  • openeo and rstac.

7. Python via reticulate

• xarray.open_zarr(), zarr-python 3, icechunk, VirtualiZarr and kerchunk all work through reticulate. Expect numpy C-order to R F-order conversion and orientation bookkeeping.
• Thin wrappers include ezzarr (ez-zarr), muschellij2/zarrr and hypertidy/zarrr (a 2022 skeleton).
• ctreesR wraps Arraylake.
• The dynverse anndata package (CRAN 0.8.0) is superseded by anndataR.
• Bioconductor/ZarrExperiment is stale (2024).
• gdalxarray (hypertidy; PyPI 0.5.1, 2026-10-20) is an xarray backend powered directly by GDAL: xr.open_dataset(path, engine="gdalxarray") (V, README). From R, through reticulate, it gives xarray’s data model over GDAL’s multidim API. That covers every GDAL Zarr feature above, including Kerchunk and Icechunk, plus non-Zarr multidim sources.
• xrdbi (not detailed yet) is a DBI-style layer that works only through xarray.