Version 0.17.2 is mainly a patch fix to bugs related to the new ArrayRef implementation.

In addition, ndarray has reduced its packaging footprint to ease supply chain reviews (and shrink the binary size!).
A special thanks to @SwishSwushPow and @weiznich for bringing this to our attention and making the necessary changes.

Added

• Add type aliases for higher-dimensional ArcArrays by @varchasgopalaswamy #1561

Fixed

• Add PartialEq implementations between ArrayRef and ArrayBase by @akern40 #1557
• Implement Sync for ArrayParts by @gaumut #1552
• Clean up clippy allows and unnecessary borrows by @RPG-Alex #1571

Documentation

• fix some typos in comments by @tinyfoolish #1547

This release of ndarray-rand adds compatibility for the new ArrayRef type in ndarray 0.17. It adds the the new RandomRefExt trait, providing sample_axis and sample_axis_using methods on ArrayRef.

This release also bumps the requirements for rand to 0.9.0 and for rand_distr to 0.5.0.

Version 0.17.1 provides a patch to fix the originally-unsound implementation of the new array reference types.

The reference types are now all unsized. Practically speaking, this has one major implication: writing functions and traits that accept RawRef and LayoutRef will now need a + ?Sized bound to work ergonomically with ArrayRef. For example, the release notes for 0.17.0 said

Reading / Writing Shape: LayoutRef<A, D>

LayoutRef lets functions view or modify shape/stride information without touching data.
This replaces verbose signatures like:

fn alter_view<S>(a: &mut ArrayBase<S, Ix1>)
where S: Data<Elem = f64>;

Use AsRef / AsMut for best compatibility:

fn alter_shape<T>(a: &mut T)
where T: AsMut<LayoutRef<f64>>;

However, these functions now need an additional bound to allow for callers to pass in &ArrayRef types:

fn alter_shape<T>(a: &mut T)
where T: AsMut<LayoutRef<f64>> + ?Sized; // Added bound here

A huge thank you to Sarah Quiñones (@sarah-quinones) for catching the original unsound bug and helping to fix it. She does truly excellent work with faer-rs; check it out!

Version 0.17.0 (2025-10-14) [YANKED]

Note: 0.17.0 was yanked due to a bug in the reference type implementation that could cause use-after-free. That bug was fixed in 0.17.1. All the changes listed here are once again available in 0.17.1, with a small caveat for the reference types. See the release notes for 0.17.1 for details.

Version 0.17.0 introduces a new array reference type — the preferred way to write functions and extension traits in ndarray. This release is fully backwards-compatible but represents a major usability improvement. The first section of this changelog explains the change in detail.

It also includes numerous new methods, math functions, and internal improvements — all credited below.

A New Way to Write Functions

TL;DR

ndarray 0.17.0 adds new reference types for writing functions and traits that work seamlessly with owned arrays and views.

When writing functions that accept array arguments:

• Use &ArrayRef<A, D> to read elements from any array.
• Use &mut ArrayRef<A, D> to modify elements.
• Use &T where T: AsRef<LayoutRef<A, D>> to inspect shape/stride only.
• Use &mut T where T: AsMut<LayoutRef<A, D>> to modify shape/stride only.

All existing function signatures continue to work; these new types are fully opt-in.

Background

ndarray has multiple ways to write functions that take arrays (a problem captured well in issue #1059). For example:

fn sum(a: ArrayView1<f64>) -> f64;
fn sum(a: &ArrayView1<f64>) -> f64;
fn sum(a: &Array1<f64>) -> f64;

All of these work, but having several equivalent forms causes confusion. The most general solution, writing generically over storage types:

fn sum<S>(a: &ArrayBase<S, Ix1>) -> f64
where S: Data<Elem = f64>;

is powerful but verbose and often hard to read. Version 0.17.0 introduces a new, simpler pattern that expresses the same flexibility more clearly.

Solution

Three new reference types make it easier to write functions that accept any kind of array while clearly expressing what kind of access (data or layout) they need.

Reading / Writing Elements: ArrayRef<A, D>

ArrayRef is the Deref target of ArrayBase. It behaves like &[T] for Vec<T>, giving access to elements and layout. Mutability is expressed through the reference itself (& vs &mut), not through a trait bound or the type itself. It is used as follows:

fn sum(a: &ArrayRef1<f64>) -> f64;
fn cumsum_mut(a: &mut ArrayRef1<f64>);

(ArrayRef1 is available from the prelude.)

Reading / Writing Shape: LayoutRef<A, D>

LayoutRef lets functions view or modify shape/stride information without touching data. This replaces verbose signatures like:

fn alter_view<S>(a: &mut ArrayBase<S, Ix1>)
where S: Data<Elem = f64>;

Use AsRef / AsMut for best compatibility:

fn alter_shape<T>(a: &mut T)
where T: AsMut<LayoutRef<f64>>;

(Accepting a LayoutRef directly can cause unnecessary copies; see #1440.)

Reading / Writing Unsafe Elements: RawRef<A, D>

RawRef augments RawArrayView and RawArrayViewMut for power users needing unsafe element access (e.g. uninitialized buffers). Like LayoutRef, it is best used via AsRef / AsMut.

Added

• A new "array reference" type by @akern40 #1440
• A diff method for calculating the difference between elements by @johann-cm #1437
• A partition method for partially sorting an array by @NewBornRustacean #1498
• A meshgrid method for building regular grids of values by @akern40 #1477
• A cumprod method for cumulative products by @NewBornRustacean #1491
• More element-wise math functions for floats by @Waterdragen #1507
  • Additions include exp_m1, ln_1p, asin, acos, atan, sinh, cosh, tanh, asinh, acosh, atanh, and hypot
• Dot product support for dynamic arrays by @NewBornRustacean #1483 and @akern40 #1494
• An axis_windows_with_stride method for strided windows by @goertzenator #1460
• In-place methods for permuting (permute_axes) and reversing (reverse_axes) axes by @NewBornRustacean #1505
• Adds into_*_iter functions as lifetime-preserving versions of into-iterator functionality by @akern40 #1510

Changed

• remove_index can now be called on views, in addition to owned arrays by @akern40

Removed

• Removed the serde-1, test, and docs feature flags; by @akern40 #1479
  • Use approx,serde,rayon instead of docs.
  • Use serde instead of serde-1

Fixed

• last_mut() now guarantees that the underlying data is uniquely held by @bluss #1429
• ArrayView is now covariant over lifetime by @akern40 #1480, so that the following code now compiles

fn fn_cov<'a>(x: ArrayView1<'static, f64>) -> ArrayView1<'a, f64> {
    x
}

Documentation

• Filled missing documentation and adds warn(missing_docs) by @akern40
• Fixed a typo in the documentation of select by @Drazhar
• Fixed a typo in the documentation of into_raw_vec_and_offset by @benliepert
• Documented Array::zeros with how to control the return type by @akern40

Other

• Expanded the gitignore file for IDEs by @XXMA16 and @akern40
• Stabilized the MSRV to 1.64 by @akern40
• Switched to nextest for testing by @akern40
• Added Miri to CI by @akern40
• Smoothed out tests that can be run with no_std by @akern40
• Updated to Edition 2021 by @akern40

Version 0.16.1 (2024-08-14)

• Refactor and simplify BLAS gemm call further by @bluss #1421
• Fix infinite recursion and off-by-one error in triu/tril by @akern40 #1418
• Fix using BLAS for all compatible cases of memory layout by @bluss #1419
• Use PR check instead of Merge Queue, and check rustdoc by @bluss #1420
• Make iterators covariant in element type by @bluss #1417

Version 0.16.0 (2024-08-03)

Featured Changes

• Better shape: Deprecate reshape, into_shape by @bluss #1310
  
  .into_shape() is now deprecated.
  Use .into_shape_with_order() or .to_shape() instead, which don't have into_shape's drawbacks.

New Features and Improvements

• Check for aliasing in RawViewMut::from_shape_ptr with a debug assertion by @bluss #1413
• Allow aliasing in ArrayView::from_shape by @bluss #1410
• Remove deprecations from 0.15.x by @bluss #1409
• Make CowArray an owned storage array, require Clone bound for into_shared by @jturner314 #1028
• Change NdProducer::Dim of axis_windows() to Ix1 by @jonasBoss #1305
• Add squeeze() to dynamic dimension arrays by @barakugav #1396
• Add flatten, flatten_with_order and into_flat to arrays by @barakugav #1397
• Make compatible with thumbv6m-none-eabi by @BjornTheProgrammer #1384
• is_unique for ArcArray by @daniellga #1399
• Add triu and tril methods directly to ArrayBase by @akern40 #1386
• Fix styling of the BLAS integration heading. by @adamreichold #1390
• Implement product_axis by @akern40 #1387
• Add reserve method for owned arrays by @ssande7 #1268
• Use inline on spit_at and smaller methods by @bluss #1381
• Update to Approx 0.5 by @bluss #1380
• Add .into_raw_vec_with_offset() and deprecate .into_raw_vec() by @bluss #1379
• Add additional array -> array view conversions by @bluss #1130
• implement DoubleEndedIterator for 1d LanesIter by @Muthsera #1237
• Add Zip::any by @nilgoyette #1228
• Make the aview0, aview1, and aview2 free functions be const fns by @jturner314 #1132
• Add missing safety checks to From<&[[A; N]]> for ArrayView and From<&mut [[A; N]]> for ArrayViewMut by @jturner314 #1131
• derived Debug for Iter and IterMut by @biskwikman #1353
• Fix Miri errors for WindowsIter and ExactChunksIter/Mut by @jturner314 #1142
• Fix Miri failure with -Zmiri-tag-raw-pointers by @jturner314 #1138
• Track-caller panics by @xd009642 #975
• Add slice_axis_move method by @jturner314 #1211
• iterators: Re-export IntoIter by @bluss #1370
• Fix unsafe blocks in s![] macro by @jturner314 #1196
• Fix comparison with NumPy of slicing with negative step by @venkat0791 #1319
• Updated Windows base Computations to be Safer by @LazaroHurtado #1297
• Update README-quick-start.md by @fumseckk #1246
• Added stride support to Windows by @LazaroHurtado #1249
• Added select example to numpy user docs by @WillAyd #1294
• Add both approx features to the readme by @nilgoyette #1289
• Add NumPy examples combining slicing and assignment by @jturner314 #1210
• Fix contig check for single element arrays by @bluss #1362
• Export Linspace and Logspace iterators by @johann-cm #1348
• Use clone_from() in two places by @ChayimFriedman2 #1347
• Update README-quick-start.md by @joelchen #1344
• Provide element-wise math functions for floats by @KmolYuan #1042
• Improve example in doc for columns method by @gkobeaga #1221
• Fix description of stack! in quick start by @jturner314 #1156

Tests, CI and Maintainer tasks

• CI: require rustfmt, nostd by @bluss #1411
• Prepare changelog for 0.16.0 by @bluss #1401
• Organize dependencies with workspace = true (cont.) by @bluss #1407
• Update to use dep: for features by @bluss #1406
• Organize the workspace of test crates a bit better by @bluss #1405
• Add rustfmt commit to ignored revisions for git blame by @lucascolley #1376
• The minimum amount of work required to fix our CI by @adamreichold #1388
• Fixed broke continuous integration badge by @juhotuho10 #1382
• Use mold linker to speed up ci by @bluss #1378
• Add rustformat config and CI by @bluss #1375
• Add docs to CI by @jturner314 #925
• Test using cargo-careful by @bluss #1371
• Further ci updates - numeric tests, and run all tests on PRs by @bluss #1369
• Setup ci so that most checks run in merge queue only by @bluss #1368
• Use merge queue by @bluss #1367
• Try to make the master branch shipshape by @adamreichold #1286
• Update ci - run cross tests only on master by @bluss #1366
• ndarray_for_numpy_users some example to code not pointed out to clippy by @higumachan #1360
• Fix minimum rust version mismatch in lib.rs by @HoKim98 #1352
• Fix MSRV build by pinning crossbeam crates. by @adamreichold #1345
• Fix new rustc lints to make the CI pass. by @adamreichold #1337
• Make Clippy happy and fix MSRV build by [@adamreichold](https://github.com/ada...