Changelog

Source: NEWS.md

dafrjulia (development version)

New Features (DataAxesFormats 0.3.0 catch-up)

• New storage backends: zarr_daf(), zip_daf(), and read-only http_daf().
• packed = TRUE for chunked+compressed storage on files_daf(), h5df(), zarr_daf(), zip_daf(), open_daf(), and complete_daf().
• open_daf() now delegates to the Julia dispatch, routing .daf.zarr, .daf.zip, http(s)://, and .h5df paths to the correct backend (it previously opened .daf.zarr/.daf.zip as plain files repositories).
• Format conversion: files_to_zarr() and zarr_to_files().
• Axis reordering: reorder_axes() and reset_reorder_axes().
• Configuration: daf_packed_options() and enforce_contracts().

Bug Fixes

• Reading a vector or matrix from packed/chunked/disk-backed storage (Zarr/Zip/Http or packed = TRUE) no longer segfaults the R process: the zero-copy jlview path is now restricted to genuine in-memory arrays, and lazy arrays are materialized via a copying fallback.

dafrjulia 0.1.1

Tracks DataAxesFormats.jl 0.3.0. DataAxesFormats 0.3.0 requires Julia 1.12 and changed the empty-buffer builder protocol in writers.jl, which broke the wrapper’s filled_empty_sparse_vector / filled_empty_sparse_matrix (they hit MethodErrors on the new signatures). Fixed:

• get_empty_sparse_vector / get_empty_sparse_matrix capture the cache_group the builder now returns and hold it until the paired filled_* call. filled_empty_sparse_vector / filled_empty_sparse_matrix thread that cache_group to the Julia finalizer and release the data write lock the builder opened (the old _b binding helpers a non-Julia wrapper used to rely on were removed in 0.3.0).
• get_empty_dense_vector / get_empty_dense_matrix unwrap the buffer from the new (buffer, cache_group) return tuple and release the write lock (dense buffers have no separate finalizer step).

The public R API (function names, arguments, return shapes) is unchanged. Full test suite passes against DataAxesFormats 0.3.0 on Julia 1.12.

dafrjulia 0.1.0

First CRAN release. Renamed from dafr — the dafr name has been freed for a native-R reimplementation in a separate repository. This package is the JuliaCall-based wrapper around DataAxesFormats.jl and has been renamed accordingly.

Migration from dafr

• library(dafr) becomes library(dafrjulia).
• options(dafr.JULIA_HOME = ...) becomes options(dafrjulia.JULIA_HOME = ...); same for dafr.julia_environment and other dafr.* options.
• Public API (function names, arguments, return shapes) is unchanged.

New Features

AnnData-like facade

• as_anndata(daf) returns a DafAnnData R6 object exposing $X, $obs, $var, $layers, $uns, $obs_names, $var_names, $n_obs, $n_vars, and $shape. Read-only and live: accesses go through to the underlying Daf on demand, served by the R-side cache on hits. obs_axis/var_axis auto-detect cell/metacell and gene; x_name is validated at construction.

R-side caching

• get_vector, get_matrix, and axis_vector now cache results on the R side, keyed by Daf-level version counters so that modifications via set_vector/set_matrix/add_axis invalidate entries automatically.
• The cache env is stored directly on each Daf object, so there is no global registry and no leak. Two R-level copies of the same Daf list share the cache (reference semantics); two wrappers of the same Julia object get separate caches.
• empty_cache(daf, clear = ..., keep = ...) now always purges the R-side cache (regardless of selectors) and validates its arguments with match.arg.

Contracts

• New CreatedOutput expectation type, matching DataAxesFormats.jl v0.2.0. Like GuaranteedOutput, a CreatedOutput item is expected to be produced by the computation and warns if a value is already present; unlike OptionalOutput, which never warns.

Performance

• Consolidated from_julia_array from 9 sequential Julia↔︎R bridge calls to 4 (5–6 for named arrays). The helper is re-entrant and uses structured Tuple return (no Main-level globals, no string delimiters).
• _prepare_for_r eagerly materialises dense non-zero-copy types (Bool, String, etc.) in Julia, saving a collect() round-trip on the R side.
• Version counters are now fetched as strings, preventing UInt32 overflow when converting to R integers. axis_version_counter, vector_version_counter, and matrix_version_counter return character strings; compare with identical() or ==.
• _prepare_for_r flattens type dispatch to reduce cold-start JIT cost.

CRAN compliance

• DESCRIPTION: updated Title, Description, URL, BugReports.
• cran-comments.md documents Julia dependency handling, \dontrun{} examples, vignette eval-off, and the JULIA_AVAILABLE test guard.
• Vignette chunks are explicitly eval = FALSE, purl = FALSE so R CMD check’s tangle step does not try to initialise Julia.
• inst/WORDLIST updated for spelling checks.
• All Julia-dependent tests guarded with skip_if(!JULIA_AVAILABLE).

Safety

• jl_R_to_julia_type (internal) now validates string inputs against a whitelist regex and rejects anything that isn’t a plausible Julia type name. Prevents arbitrary Julia code execution via the type-conversion path.
• Daf(jl_obj) rejects non-JuliaObject inputs with a clear error instead of passing them through silently.

Breaking changes vs. dafr

• Package name. All the options and library() calls above.
• get_vector / get_matrix return values for zero-copy types are now live views over Julia memory (via jlview / jlview_sparse). Copy explicitly (as.numeric(v), as.matrix(m)) if stability across subsequent Daf modifications is needed.

---