minarrow v0.16.0

0.16.0 - 2026-07-16

This version completes Python ecosystem compatibility with DLPack support for landing data (or ChunkedNdArray) data in Rust
and bridging to Python for AI/ML and back.

Moving forward we will aim to minimise breaking changes and stabilise to a 1.0 version in the coming month(s). Additionally, include a Stable Rust build.

The new NdArray, XArray and SuperNdArray may however see breaking changes before then until they have had time for production use.

Added

NdArray, XArray and DLPack Feature:

• NdArray n-dimensional dense array (ndarray feature) over Buffer<T>,
  generic over f32/f64, with NaN null semantics, NdArrayV zero-copy views,
  transpose and axis permutation, and Table/Matrix/Array interop.
• SuperNdArray chunked n-dimensional array with SuperNdArrayV
  chunk-spanning views and rechunking.
• XArray labelled n-dimensional array (xarray feature) with named
  dimensions, coordinate selection via at/between/nearest, and owned,
  view, or chunked storage behind one type.
• DLPack tensor interchange (dlpack feature) over the legacy and 1.x
  versioned ABIs for zero-copy sharing with PyTorch, JAX, and TensorFlow.
• AxisSelection trait for .s() axis slicing, with NdArrayVT,
  SuperNdArrayVT, and XArrayVT view tuple aliases.
• Broadcast kernels for NdArray, SuperNdArray, and XArray, plus
  Value variants, conversions, and concatenation for the new types.
• NdArray in minarrow-py with the DLPack protocol - __dlpack__,
  from_dlpack, and named bridges to NumPy, PyTorch, JAX, TensorFlow,
  and CuPy.
• PyNdArray in minarrow-pyo3 (ndarray feature) for Rust extensions
  bridging tensors to Python, with the DLPack capsule glue hosted in
  its ffi::dlpack and shared by minarrow-py.

Minor Feature Additions

• At value_at to Array and ArrayV for (opt-in) normalised equality checking.
• Added bitmask-driven gather to Array, Table and their view variants.
• Added default SIMD chipset behaviour.
• Added bitwise kernels.
• Added default display trait impl to Scalar.
• Added LBuffer::freeze().
• Added NdArray n-dimensional array with views and chunked variants.
• Added XArray labelled n-dimensional array.
• Added axis selection, broadcasting, and Value support for the n-dimensional types.
• Added DLPack FFI for zero-copy PyTorch, JAX, and TensorFlow interchange, with NdArray in minarrow-py and minarrow-pyo3.
• Added get, get_unchecked, get_str and get_str_unchecked methods to Array
• Added bitmask gather to Array, Table and their view variants.

Bugfixes

• Fixed a kernel branching issue in the arithmetic SIMD paths.
• Fixed out-of-range categorical codes to resolve to the empty string.
• Fixed StringArray length reporting the byte length when MaskedArray was not imported.
• Fixed null handling bug on String and Categorical set_str method.

Modified

• Bumped pyo3 to 0.29.

• Normalised hash_element_at so NaN maps to a dummy/sentinel value

• Added value_eq method to Array and ArrayV where NaN == NaN, -0.0 == 0.0, Null == Null etc.

• Normalised equality checking for Scalar -0.0 == 0.0 and NaN == NaN.
  Note: the above three only affect users opting into the methods and/or Scalar abstraction.
  For users wanting IEEE compatibility for floats using myarr.clone().num().f64() gets to a FloatArray zero-copy.

Minarrow V0.15

Minarrow 0.15.0 accompanies the major release of Minarrow Python.

Added

• _into Datetime kernel variants for efficient non-allocations.
• _into Bitmask kernel variants for efficient non-allocations.
• UnsafeMut utility Buffer for handling cross-thread bitpacked writes.

Changed

• Improved name construction ergonomics.

Minarrow V0.14.0

1. Added _into mutable output buffer datetime variants supporting efficient parallelism.
2. Added TimePeriod enum consistency for datetime periods with automatic deref on string constituents i.e., week becomes TimePeriod::Week at the call-site.
3. Fixed BitmaskV lifetime.
4. Fixed BooleanArray length visibility.

Minarrow 0.13.0

See CHANGELOG.md for details

minarrow-pyo3 0.3.0

Changed

• Bumped minarrow dependency from 0.10.1 to 0.11.0.

Removed

• Unused nightly feature gates (allocator_api, slice_ptr_get,
  portable_simd) from the crate root. The underlying allocator-aware types
  are re-exported from minarrow, so the gates are not required at this
  crate's call sites.

minarrow 0.11.0

[0.11.0] - 2026-05-15

Added

• arrow-rs / polars import bridges (#70): symmetric from_apache_arrow /
  from_polars across Array, FieldArray, Table, SuperArray, and
  SuperTable, each with a try_* sibling returning Result<_, MinarrowError>.
  Adds From<&Series>, From<&RecordBatch>, From<&DataFrame>, and
  From<&[RecordBatch]> for .into() ergonomics.
• New feature-gated bridge modules src/ffi/arrow_rs.rs (cast_arrow) and
  src/ffi/polars.rs (cast_polars) centralising the export / import pairs (#70).
• MinarrowError::BridgeError variant carrying arrow-rs / polars FFI failures,
  with feature-gated From impls for ArrowError and PolarsError (#70).
• Matrix: strided LAPACK matrix getters, improved interoperability, and
  improved docs (#59, #62).
• Cross-tabulate string kernel.
• has_nulls accessor (#63).
• with_capacity for CategoricalArray<T> (#67).
• Filled missing trait impls: SuperTable From, TryFrom<Value> gaps,
  From<BooleanArrayV> for Value, plus other From impls (#67).
• Vec64 arm for the fa! macro.
• Zero-copy to_array escape path.
• Minimal log dependency.

Changed

• View -> owned conversions take a fast path when the view spans its full
  backing storage (offset 0, length matches the backing array): skips
  slice_clone and returns an Arc clone of the underlying allocation (#80).
• Polars Array / FieldArray / Table from_polars now route through
  SuperArray::from_polars / SuperTable::from_polars, correctly handling
  multi-chunked Series / DataFrame inputs (#76, #77).
• Bumped all dependencies to latest semver (#68).
• Breaking: renamed SuperArray::to_apache_arrow_chunks and
  SuperTable::to_apache_arrow_batches to to_apache_arrow (#70).

Removed

• Breaking: removed Array::to_apache_arrow_with_field and
  Array::to_polars_with_field. Wrap in a FieldArray and call its to_*()
  method when an explicit Field is needed (#70).

Fixed

• Per-call Box leak in the to_* export paths — ArrowArray / ArrowSchema
  heap wrappers were not freed after read() (#70).
• Categorical pandas -1 null sentinel handling (#69).
• Zero-sized shared buffer checks: round-trips through SharedBuffer for
  zero-row columns no longer panic on alignment assertions (#73).
• Polars default features (#77).
• Default features and the test harness; stale ref in benches.
• Pinned the toolchain via rust-toolchain.toml to work around an upstream
  issue.