Probe: write by read

probe.R writes R’s volcano matrix (87 x 61, float64) with each writer, then reads every resulting store back with each reader. Fixtures are written to probe/fixtures/ (gitignored) and rebuilt on every run. Output:

• results.csv: the grid below;
• sessioninfo.txt: versions.

Rscript probe/probe.R

Run on 2026-11-09 with:

• R 4.6.1, GDAL 3.14.0dev and netCDF-C 4.9.2;
• terra 1.9-57, stars 0.7-3, vapour 0.16.0.9002, gdalraster 2.7.0.9002 and GDAL7 0.1.0;
• zarr 1.0.1, pizzarr 0.2.2 (the r-universe build, with zarrs), Rarr 2.1.14, zaro 0.0.1.9030 and RNetCDF 2.11-2.

Rarr and zaro trail their current releases (devel 2.1.57 and 0.0.1.9036).

Writers

writer	layout
gdal v2 none / blosc / zlib	gdal_translate -of ZARR -co FORMAT=ZARR_V2 -co COMPRESS=…; C order, single chunk, X/Y coordinate arrays, .zmetadata
gdal v3 zstd	FORMAT=ZARR_V3 COMPRESS=ZSTD
gdal v3 sharded	FORMAT=ZARR_V3 COMPRESS=GZIP BLOCKSIZE=32,32 SHARD_CHUNK_SHAPE=16,16
zarr v3 default	zarr::as_zarr(volcano, name = "volcano"): transpose + bytes + blosc(zstd, bitshuffle), single chunk
pizzarr v2 / v3 default	zarr_create_array(data = volcano, zarr_format = 2L/3L): v2 order: "F" + zstd; v3 bytes + zstd
Rarr v2 zlib	write_zarr_array(zarr_version = 2L, use_zlib()), 32x32 chunks, order: "F"
Rarr v3 zstd	write_zarr_array(zarr_version = 3L, use_zstd()), 32x32 chunks, transpose codec

Reading a cell

• exact: the reader returned volcano as is.
• t, flip rows, …: the values are right once you apply that transform. This is the reader’s orientation convention, not an error.
• WRONG: the reader returned values without an error, and no transpose or flip recovers volcano. WRONG: shuffled means the right values are present in some other arrangement.
• ERR: the reader raised an error. The message is truncated.

Results

writer	terra	stars	vapour	gdalraster mdim	GDAL7	zarr	pizzarr	Rarr	zaro	RNetCDF (NCZarr)
gdal v2 none	exact	t	exact	exact	t	exact	exact	exact	exact	t
gdal v2 blosc	exact	t	exact	exact	t	exact	exact	exact	exact	WRONG
gdal v2 zlib	exact	t	exact	exact	t	ERR: attempt to set an attribute on NULL…	exact	exact	exact	WRONG
gdal v3 zstd	exact	t	exact	exact	t	exact	exact	exact	exact	ERR: NetCDF: Variable not found…
gdal v3 sharded	exact	t	exact	exact	t	exact	ERR: Unsupported V3 codec: ‘sharding_indexe…	ERR: The following codecs are not supported…	exact	ERR: NetCDF: Variable not found…
zarr v3 default	flip rows	t	exact	exact	t	exact	exact	exact	WRONG: shuffled	ERR: NetCDF: Variable not found…
pizzarr v2 default	flip rows	t	exact	exact	t	exact	exact	exact	exact	WRONG
pizzarr v3 default	flip rows	t	exact	exact	t	exact	exact	exact	exact	ERR: NetCDF: Variable not found…
Rarr v2 zlib	flip rows	t	exact	exact	t	ERR: attempt to set an attribute on NULL…	exact	exact	exact	WRONG
Rarr v3 zstd	flip rows	t	exact	exact	t	exact	WRONG	exact	WRONG	ERR: NetCDF: Variable not found…

What to take from it

1. GDAL reads everything here, whichever binding you use. The GDAL columns differ only in orientation:
   • stars and GDAL7 keep GDAL’s fastest-first dimension order;
   • terra flips rows when there are no Y coordinates.
2. Sharding is read by GDAL, zarr and zaro. pizzarr’s R codec path and Rarr 2.1.14 refuse it, with clear messages.
3. zarr (R-CF) cannot decode the v2 zlib compressor. gzip is fine. The error message does not say what is missing.
4. The v3 transpose codec is a trap:
   • pizzarr reads it correctly from a single-chunk store, but reads wrong values from Rarr’s 32x32-chunked store, which has partial edge chunks;
   • zaro reads wrong values from both.

   Both of these are silent. They are worth reporting upstream.

5. NCZarr (RNetCDF on netCDF-C 4.9.2):
   • it reads uncompressed v2;
   • compressed v2 comes back as garbage with no error: with no HDF5_PLUGIN_PATH, the codecs are not loaded;
   • v3 arrays are invisible.

Not covered

The probe only uses local stores and float64 data. It does not cover:

• remote stores (HTTP, S3, GCS);
• Kerchunk and VirtualiZarr references;
• Icechunk repositories;
• fill values and NA handling;
• strings and other data types;
• consolidated metadata on its own;
• Rzarrs, zr, the gdalraster gdalmultidim-api branch, and the reticulate route.

Each of these would be a natural next column or row.