NEWS

---

pizzarr 0.2.2

(developed with the assistance of Claude Opus 5.5)

Released on r-universe only. CRAN stays at 0.2.1; the pure-R fixes below reach CRAN with the next CRAN release.

String data

• Zarr V3 arrays with data_type: "string" and the vlen-utf8 codec, as written by zarr-python 3, could not be read: the codec pipeline stopped with Unsupported V3 codec: 'vlen-utf8' (#192). They now read as character arrays, and zarr_create(dtype = "|O", object_codec = VLenUtf8Codec$new(), zarr_format = 3L) writes the same layout, with fill_value: "".
• VLenUtf8Codec failed to encode an empty string (replacement has length zero), which also hit any partially filled boundary chunk, and decoded an empty string as two stray bytes. This affected V2 as well as V3. Decoded strings are now marked UTF-8, and strings in another encoding are converted to UTF-8 before writing.
• Writing fixed-length S or U arrays whose shape is not a multiple of the chunk shape failed with replacement has length zero, because the padded boundary chunks hold empty strings (#112). Empty and NA strings now write as null-filled slots and read back as "".
• The default fill_value of 0 is now "" for S and U dtypes, as in zarr-python. Creating such an array without a fill_value previously stopped with fill_value must be a character string for string dtype.
• A U string longer than its dtype now raises an error. Strings up to four times too long previously overwrote the neighbouring element.
• dtype strings are written the way numpy spells them: | for S, O, and one-byte b, i, u (<S20 becomes |S20, <i1 becomes |i1), and < in place of | for multi-byte types. zarr-python 3 refused to open arrays with the other spellings. Existing stores still read; their metadata takes the numpy spelling if pizzarr rewrites it.

Rust backend

• extendr-api now comes from crates.io (0.8.2) rather than a pin to extendr's git master branch. 0.8.2 carries the R >= 4.5 OBJSXP fix that the pin was for, and a crates.io dependency can be vendored.
• Cargo.lock refreshed within semver: zarrs 0.23.10 -> 0.23.14, tokio 1.53. This clears RustSec advisories in h2 (unbounded empty DATA frames), rustls (TLS 1.3 handshake messages across encryption levels), crossbeam-epoch, event-listener, and anyhow. Advisories remain in lru (fixed upstream in 0.18; zarrs still requires 0.16) and quick-xml (needs object_store 0.14, planned separately).

Documentation

• zarrs_get_key(), exported in 0.2.1, now has a help page.
• The package description names Zarr V3 alongside V2.
• Help pages regenerated with roxygen2 8.1.0.

pizzarr 0.2.1 (2026-08-22)

Cloud storage

• New zarrs_get_key() reads a single key from any store the zarrs backend can open, returning raw bytes or NULL when the key is absent. Cloud stores delegate all I/O to object_store and expose no key-level read, so consolidated metadata (.zmetadata) was unreachable from R for an s3:// or gs:// URL. Reading that one key is enough to enumerate a store's arrays and recover their dimension names and attributes, which is what group-level inquiry over s3:// needs.

Cloud storage fixes

• S3 requests were always sent unsigned, so configured AWS credentials were ignored and private buckets returned 403. Signing is now decided from the environment: requests are unsigned when no credential source is present, so public buckets still work without configuration, and signed as soon as AWS_ACCESS_KEY_ID, AWS_SESSION_TOKEN, AWS_PROFILE, or another credential variable is set. AWS_SKIP_SIGNATURE forces either behaviour.
• zarr_open() on an s3:// or gs:// string fell through to DirectoryStore, which tried to create a directory named after the URL. These URLs now resolve to S3Store and GcsStore, and cloud stores default to read mode.
• S3Store and GcsStore inherited no-op get_item(), contains_item(), and listdir() methods that silently returned NULL. They now raise an error naming the supported read paths.
• HttpStore$listdir() took no arguments, so the path argument every other store accepts raised an error that callers swallowed --- ZarrGroup's print method reported No. members : 0 for every HTTP-backed group. It now accepts path and lists within it.

Documentation

• New vignettes/remote-stores.Rmd covering remote and cloud access: the mapping from fsspec/xarray storage_options to pizzarr's environment variables, the HTTPS and s3:// routes, alternate endpoints such as MinIO, Ceph, and Open Storage Network pods, and the current limits.
• Alternate S3 endpoints are configured with AWS_ENDPOINT; AWS_ALLOW_HTTP and AWS_VIRTUAL_HOSTED_STYLE_REQUEST cover plain-HTTP and path-style addressing. This worked before but was undocumented.
• Anonymous access to public GCS buckets over gs:// is documented, and now tested. It needs GOOGLE_SKIP_SIGNATURE=true, since GCS does not infer anonymous access from absent credentials the way S3 does.
• S3Store, GcsStore, and HttpStore gained detailed documentation of credentials, endpoints, and which distribution tier each requires.
• zarr_open()'s store argument now documents every accepted form: a path, an http(s):// URL, an s3:// or gs:// URL, or a Store object.
• zarr_open_group()'s storage_options argument is documented as ignored rather than as an fsspec passthrough.
• Vignettes reorganized around data conventions rather than transport, now that connection details have their own article. remote-ome-ngff.Rmd merged into ome-ngff.Rmd, which covers the OME-NGFF multiscales convention against both a local and a remote store; remote-anndata.Rmd renamed anndata.Rmd.
• The OME-NGFF and AnnData vignettes now evaluate their code at build time instead of showing static listings, so the examples are checked. Both require the blosc package and network access, and degrade to unevaluated when either is missing.
• README gained a "Remote and cloud data" section.

pizzarr 0.2.0 (2026-04-20)

Two-tier distribution

• CRAN ships pure R (no Rust compilation). r-universe ships pre-built binaries with the zarrs Rust backend compiled in. No Rust toolchain needed for end users on either tier.
• New functions: is_zarrs_available(), pizzarr_compiled_features(), pizzarr_upgrade().
• tools/cran-build.sh produces a CRAN tarball with src/ stripped out.

zarrs Rust backend (r-universe tier)

• Added zarrs Rust backend via extendr for parallel decompression, codec execution, and store abstraction. 9 #[extendr] functions, 13 Rust source files.
• Transparent dispatch: ZarrArray$get_item() and $set_item() automatically use the zarrs fast path when eligible (contiguous slices, supported dtype, filesystem or HTTP store). Unsupported selections fall through to the R-native path silently.
• Read path: two-step dtype dispatch retrieves data at stored precision, then widens to R-compatible types (f32→f64, i16→i32, etc.) in Rust before crossing the FFI boundary.
• Write path: symmetric narrowing with range checks (R double→f32, R integer→i16, etc.). Out-of-range values produce errors, not silent truncation.
• Array creation via zarrs: zarr_create() dispatches to the zarrs backend when available. Supports V2 and V3 formats with four codec presets (none, gzip, blosc, zstd).
• C↔F order transpose moved from R (aperm) to Rust, eliminating two full-array copies per read and up to five per write. Cache-blocked tiling for 2D, output-order iteration for nD.
• HTTP/HTTPS reads via zarrs_http crate. HttpStore-backed arrays use zarrs for parallel chunk decode on remote data. R-native crul-based path remains as fallback.
• S3 reads via object_store crate. New S3Store R6 class for s3:// URLs. Public buckets work without credentials (unsigned requests).
• GCS reads via object_store crate. New GcsStore R6 class for gs:// URLs. Public GCS data accessible via HTTPS endpoint without credentials.
• Process-global store handle cache with explicit lifecycle management via zarrs_close_store().

Configuration

• New pizzarr_config() function for viewing and setting concurrency options. Three settings: nthreads (rayon thread pool size), concurrent_target (codec concurrency level), and http_batch_range_requests (multipart range toggle for HTTP stores). All three backed by environment variables (PIZZARR_NTHREADS, PIZZARR_CONCURRENT_TARGET, PIZZARR_HTTP_BATCH_RANGE_REQUESTS).
• zarrs_runtime_info() now includes nthreads (rayon thread pool size).

Breaking changes

• Removed R-native parallel chunk I/O infrastructure. The pizzarr.parallel_backend, pizzarr.parallel_write_enabled, and pizzarr.progress_bar options are no longer used. On r-universe builds, zarrs handles parallelism internally via its Rust thread pool. On CRAN (pure R), chunk I/O is sequential.
• Dropped pbapply, parallel, future, and future.apply from Suggests.
• Removed vignettes/parallel.Rmd. Replaced by vignettes/zarrs-backend.Rmd.
• Minimum R version raised from 4.1.0 to 4.2.

Other changes

• New vignettes/zarrs-backend.Rmd documenting the zarrs backend, dispatch behavior, store types, concurrency tuning, and cloud access.
• R-native chunk loop simplified from parallel dispatch infrastructure to sequential for loop.

pizzarr 0.1.3

• Added set_dimension_names() / get_dimension_names() methods to ZarrArray (#170).
• pkgdown site and documentation improvements (#171).

pizzarr 0.1.2 (2026-04-06)

• Added get_dimension_names() method to ZarrArray for V3 dimension names support.
• OME-NGFF vignettes now use base R rasterImage() instead of the raster package (#161).

pizzarr 0.1.1 (2026-03-25)

Fixes a flaky test failing on CRAN -- no functional changes.

pizzarr 0.1.0 (2026-03-24)

This is the first release of pizzarr to CRAN. The package has significant testing and validation but has not been vetted by a large and diverse user community. As a result, bugs and undesirable behavior may be found. If you find issues please report them at:

https://github.com/zarr-developers/pizzarr/issues

• Initial release of pizzarr.
• Read and write Zarr V2 and Zarr V3 stores.
• Storage backends: DirectoryStore, MemoryStore, HttpStore.
• Compression codecs: Zstd, LZ4, Blosc, zlib, gzip, bzip2, LZMA.
• VLenUTF8 object codec for variable-length strings.
• Parallel read/write support via pbapply, parallel, and future.
• R-like one-based and Python-like zero-based slicing.