These are the major changes made in each release. For details of the changes see the commit log at http://github.com/kwgoodman/bottleneck

*Release date: 2016-10-20*

This release is a complete rewrite of Bottleneck.

**Port to C**

- Bottleneck is now written in C
- Cython is no longer a dependency
- Source tarball size reduced by 80%
- Build time reduced by 66%
- Install size reduced by 45%

**Redesign**

- Besides porting to C, much of bottleneck has been redesigned to be simpler and faster. For example, bottleneck now uses its own N-dimensional array iterators, reducing function call overhead.

**New features**

- The new function bench_detailed runs a detailed performance benchmark on a single bottleneck function.
- Bottleneck can be installed on systems that do not yet have NumPy installed. Previously that only worked on some systems.

**Beware**

- Functions partsort and argpartsort have been renamed to partition and argpartition to match NumPy. Additionally the meaning of the input arguments have changed: bn.partsort(a, n) is now equivalent to bn.partition(a, kth=n-1). Similarly for bn.argpartition.
- The keyword for array input has been changed from arr to a in all functions. It now matches NumPy.

**Thanks**

- Moritz E. Beber: continuous integration with AppVeyor
- Christoph Gohlke: Windows compatibility
- Jennifer Olsen: comments and suggestions
- A special thanks to the Cython developers. The quickest way to appreciate their work is to remove Cython from your project. It is not easy.

Release notes from past releases.

*Release date: 2016-06-22*

This release makes Bottleneck more robust, releases GIL, adds new functions.

**More Robust**

- move_median can now handle NaNs and min_count parameter
- move_std is slower but numerically more stable
- Bottleneck no longer crashes on byte-swapped input arrays

**Faster**

- All Bottleneck functions release the GIL
- median is faster if the input array contains NaN
- move_median is faster for input arrays that contain lots of NaNs
- No speed penalty for median, nanmedian, nanargmin, nanargmax for Fortran ordered input arrays when axis is None
- Function call overhead cut in half for reduction along all axes (axis=None) if the input array satisfies at least one of the following properties: 1d, C contiguous, F contiguous
- Reduction along all axes (axis=None) is more than twice as fast for long, narrow input arrays such as a (1000000, 2) C contiguous array and a (2, 1000000) F contiguous array

**New Functions**

- move_var
- move_argmin
- move_argmax
- move_rank
- push

**Beware**

- median now returns NaN for a slice that contains one or more NaNs
- Instead of using the distutils default, the ‘-O2’ C compiler flag is forced
- move_std output changed when mean is large compared to standard deviation
- Fixed: Non-accelerated moving window functions used min_count incorrectly
- move_median is a bit slower for float input arrays that do not contain NaN

**Thanks**

Alphabeticaly by last name

- Alessandro Amici worked on setup.py
- Pietro Battiston modernized bottleneck installation
- Moritz E. Beber set up continuous integration with Travis CI
- Jaime Frio improved the numerical stability of move_std
- Christoph Gohlke revived Windows compatibility
- Jennifer Olsen added NaN support to move_median

*Release date: 2015-02-06*

This release is a complete rewrite of Bottleneck.

**Faster**

- “python setup.py build” is 18.7 times faster
- Function-call overhead cut in half—a big speed up for small input arrays
- Arbitrary ndim input arrays accelerated; previously only 1d, 2d, and 3d
- bn.nanrankdata is twice as fast for float input arrays
- bn.move_max, bn.move_min are faster for int input arrays
- No speed penalty for reducing along all axes when input is Fortran ordered

**Smaller**

- Compiled binaries 14.1 times smaller
- Source tarball 4.7 times smaller
- 9.8 times less C code
- 4.3 times less Cython code
- 3.7 times less Python code

**Beware**

- Requires numpy 1.9.1
- Single API, e.g.: bn.nansum instead of bn.nansum and nansum_2d_float64_axis0
- On 64-bit systems bn.nansum(int32) returns int32 instead of int64
- bn.nansum now returns 0 for all NaN slices (as does numpy 1.9.1)
- Reducing over all axes returns, e.g., 6.0; previously np.float64(6.0)
- bn.ss() now has default axis=None instead of axis=0
- bn.nn() is no longer in bottleneck

**min_count**

- Previous releases had moving window function pairs: move_sum, move_nansum
- This release only has half of the pairs: move_sum
- Instead a new input parameter, min_count, has been added
- min_count=None same as old move_sum; min_count=1 same as old move_nansum
- If # non-NaN values in window < min_count, then NaN assigned to the window
- Exception: move_median does not take min_count as input

**Bug Fixes**

- Can now install bottleneck with pip even if numpy is not already installed
- bn.move_max, bn.move_min now return float32 for float32 input

*Release date: 2014-01-21*

This version of Bottleneck requires NumPy 1.8.

**Breaks from 0.7.0**

- This version of Bottleneck requires NumPy 1.8
- nanargmin and nanargmax behave like the corresponding functions in NumPy 1.8

**Bug fixes**

- nanargmax/nanargmin wrong for redundant max/min values in 1d int arrays

*Release date: 2013-09-10*

**Enhancements**

- bn.rankdata() is twice as fast (with input a = np.random.rand(1000000))
- C files now included in github repo; cython not needed to try latest
- C files are now generated with Cython 0.19.1 instead of 0.16
- Test bottleneck across multiple python/numpy versions using tox
- Source tarball size cut in half

**Bug fixes**

- #50 move_std, move_nanstd return inappropriate NaNs (sqrt of negative #)
- #52 make test fails on some computers
- #57 scipy optional yet some unit tests depend on scipy
- #49, #55 now works on Mac OS X 10.8 using clang compiler
- #60 nanstd([1.0], ddof=1) and nanvar([1.0], ddof=1) crash

*Release date: 2012-06-04*

Thanks to Dougal Sutherland, Bottleneck now runs on Python 3.2.

**New functions**

- replace(arr, old, new), e.g, replace(arr, np.nan, 0)
- nn(arr, arr0, axis) nearest neighbor and its index of 1d arr0 in 2d arr
- anynan(arr, axis) faster alternative to np.isnan(arr).any(axis)
- allnan(arr, axis) faster alternative to np.isnan(arr).all(axis)

**Enhancements**

- Python 3.2 support (may work on earlier versions of Python 3)
- C files are now generated with Cython 0.16 instead of 0.14.1
- Upgrade numpydoc from 0.3.1 to 0.4 to support Sphinx 1.0.1

**Breaks from 0.5.0**

- Support for Python 2.5 dropped
- Default axis for benchmark suite is now axis=1 (was 0)

**Bug fixes**

- #31 Confusing error message in partsort and argpartsort
- #32 Update path in MANIFEST.in
- #35 Wrong output for very large (2**31) input arrays

*Release date: 2011-06-13*

The fifth release of bottleneck adds four new functions, comes in a single source distribution instead of separate 32 and 64 bit versions, and contains bug fixes.

J. David Lee wrote the C-code implementation of the double heap moving window median.

**New functions**

- move_median(), moving window median
- partsort(), partial sort
- argpartsort()
- ss(), sum of squares, faster version of scipy.stats.ss

**Changes**

- Single source distribution instead of separate 32 and 64 bit versions
- nanmax and nanmin now follow Numpy 1.6 (not 1.5.1) when input is all NaN

**Bug fixes**

- #14 Support python 2.5 by importing with statement
- #22 nanmedian wrong for particular ordering of NaN and non-NaN elements
- #26 argpartsort, nanargmin, nanargmax returned wrong dtype on 64-bit Windows
- #29 rankdata and nanrankdata crashed on 64-bit Windows

*Release date: 2011-03-17*

This is a bug fix release.

**Bug fixes**

- #11 median and nanmedian modified (partial sort) input array
- #12 nanmedian wrong when odd number of elements with all but last a NaN

**Enhancement**

- Lazy import of SciPy (rarely used) speeds Bottleneck import 3x

*Release date: 2011-03-08*

This is a bug fix release.

Same bug fixed in Bottleneck 0.4.1 for nanstd() was fixed for nanvar() in this release. Thanks again to Christoph Gohlke for finding the bug.

*Release date: 2011-03-08*

This is a bug fix release.

The low-level functions nanstd_3d_int32_axis1 and nanstd_3d_int64_axis1, called by bottleneck.nanstd(), wrote beyond the memory owned by the output array if arr.shape[1] == 0 and arr.shape[0] > arr.shape[2], where arr is the input array.

Thanks to Christoph Gohlke for finding an example to demonstrate the bug.

*Release date: 2011-03-08*

The fourth release of Bottleneck contains new functions and bug fixes. Separate source code distributions are now made for 32 bit and 64 bit operating systems.

**New functions**

- rankdata()
- nanrankdata()

**Enhancements**

- Optionally specify the shapes of the arrays used in benchmark
- Can specify which input arrays to fill with one-third NaNs in benchmark

**Breaks from 0.3.0**

- Removed group_nanmean() function
- Bump dependency from NumPy 1.4.1 to NumPy 1.5.1
- C files are now generated with Cython 0.14.1 instead of 0.13

**Bug fixes**

- #6 Some functions gave wrong output dtype for some input dtypes on 32 bit OS
- #7 Some functions choked on size zero input arrays
- #8 Segmentation fault with Cython 0.14.1 (but not 0.13)

*Release date: 2010-01-19*

The third release of Bottleneck is twice as fast for small input arrays and contains 10 new functions.

**Faster**

- All functions are faster (less overhead in selector functions)

**New functions**

- nansum()
- move_sum()
- move_nansum()
- move_mean()
- move_std()
- move_nanstd()
- move_min()
- move_nanmin()
- move_max()
- move_nanmax()

**Enhancements**

- You can now specify the dtype and axis to use in the benchmark timings
- Improved documentation and more unit tests

**Breaks from 0.2.0**

- Moving window functions now default to axis=-1 instead of axis=0
- Low-level moving window selector functions no longer take window as input

**Bug fix**

- int input array resulted in call to slow, non-cython version of move_nanmean

*Release date: 2010-12-27*

The second release of Bottleneck is faster, contains more functions, and supports more dtypes.

**Faster**

- All functions faster (less overhead) when output is not a scalar
- Faster nanmean() for 2d, 3d arrays containing NaNs when axis is not None

**New functions**

- nanargmin()
- nanargmax()
- nanmedian()

**Enhancements**

- Added support for float32
- Fallback to slower, non-Cython functions for unaccelerated ndim/dtype
- Scipy is no longer a dependency
- Added support for older versions of NumPy (1.4.1)
- All functions are now templated for dtype and axis
- Added a sandbox for prototyping of new Bottleneck functions
- Rewrote benchmarking code

*Release date: 2010-12-10*

Initial release. The three categories of Bottleneck functions:

- Faster replacement for NumPy and SciPy functions
- Moving window functions
- Group functions that bin calculations by like-labeled elements