• MultiCell now requires a cell count of one or more; the argument has no default by @henryiii in #1179
• Filling an integer axis with a value that does not fit in an int raises ValueError instead of wrapping around by @henryiii in #1181
• boost_histogram.numpy matches NumPy more closely: empty input, 1D histogramdd input, a single edge array in histogram2d, and range validation by @henryiii in #1178

• Encode UTF-8 for non-ASCII strings filled into StrCategory from a list or array by @henryiii in #1177
• Avoid C++ aborts in axis metadata compare, MultiCell fill, and Boolean.value(); correct empty() for MultiCell by @henryiii in #1179
• Keep growth axes valid after a fill or merge, and raise worker exceptions from a threaded fill by @henryiii in #1182
• Correct UHI indexing with flow locators, disabled flow bins, and sliced group rebin by @henryiii in #1183
• Give each axis copy its own metadata dict, and correct serialization, register(), integer-storage multiplication, transform equality, and repr by @henryiii in #1184
• Avoid Boost indexed-range UB in rank-0 histograms and add Windows 3.14t testing by @ikrommyd in #1175

• Cheaper fill argument variant, faster dtype dispatch, and move support for MultiCell by @henryiii in #1181
• Release the GIL on copy, reset, compare, and in-place histogram operations by @henryiii in #1187

• Document that view() can be invalidated by a growing fill by @henryiii in #1182
• Restore the Histogram API page on Read the Docs and correct the user guide and wheel table by @henryiii in #1180

• Hold the GIL when copying or destroying Python objects in axes by @henryiii in #1172
• Remove the dead regular_numpy Python axis, factor out pybind11 helpers, and correct stub drift and weak tests by @henryiii in #1185
• Config and CI fixes: uv cache keys, workflow permissions and timeouts, and AGENTS.md commands by @henryiii in #1180
• Test Python 3.15 now that NumPy wheels are on PyPI by @henryiii in #1170
• Fix the scheduled wheel and weekly-update failures by @henryiii in #1173
• Bump astral-sh/setup-uv from 9.0.0 to 10.0.1 by @dependabot in #1174

• Python 3.15 wheels, including free-threading and Pyodide, by @henryiii in #1163
• Support flow=False in histogram projections by @Rishabh-git10 in #1109
• Add Histogram.allclose for approximate histogram comparison by @gohil-jay in #778
• Add a pytest plugin with rich assertion reporting for histogram comparisons by @gohil-jay in #763

• Support subtraction for the MultiCell storage by @henryiii in #1131
• Support vectorized NumPy-array indexing for getting and setting bins by @henryiii in #1134
• Support addition for Mean/WeightedMean views by @henryiii in #1137
• Add reflected __rsub__ and __rtruediv__ operators by @henryiii in #1155
• Accept a scalar sample argument in fill by @henryiii in #1124
• Rebin with groups: MultiCell support, flow handling, and a group_mapping protocol by @henryiii in #1148

• Reject arithmetic with mismatched storages by @divyanshu-iitian in #1166
• Return a zero-filled vector from MultiCell sum() over zero bins instead of an empty list by @henryiii in #1132
• Raise TypeError instead of ValueError for a non-integer locator offset and a string sample= by @henryiii in #1163
• Clear error for unsupported histogram in-place operations by @henryiii in #1130
• Support common transformed axes in the numpy shortcuts by @henryiii in #1122
• Correct Mean storage string rejection and __setitem__ error message by @henryiii in #1117
• Several fixes from automated review by @henryiii in #1116 and #1118
• Promote integer storages to double on division by @henryiii in #1138
• Do not mutate input edges and validate histogramdd dimensions in the numpy module by @henryiii in #1144
• Fix out= handling, scalar division, and broadcasting in structured views by @henryiii in #1146
• Fix non-contiguous array handling and memory safety in the C++ bindings by @henryiii in #1147
• Fix indexing, __setitem__ flow transfer, and fill edge cases by @henryiii in #1149
• Avoid Boost indexed-range UB in rank-0 sum()/empty() by @henryiii in #1153
• Normalize negative integer slice bounds in indexing by @henryiii in #1156
• Remove ODR-violating operator<< for accumulators::sum by @henryiii in #1161

• Avoid deep copies in comparisons and hot paths by @henryiii in #1150

• Cleanup, modernization, and gap-filling in the user guide by @henryiii in #1141
• Note that fancy indexing on views returns a copy by @henryiii in #1136
• Correct the wheel table and add a version-support checklist by @henryiii in #1168

• Use pytest.approx instead of NumPy assertion helpers by @gohil-jay in #780
• Validate UHI serialization against the uhi JSON schema by @henryiii in #1140
• Broaden MultiCell storage coverage by @henryiii in #1127
• Add regression test for scalar-broadcast fill on non-inclusive axes by @henryiii in #1125
• Add ND projection test for flow=False by @Rishabh-git10 in #1111
• Disable cloudpickle tests on PyPy by @henryiii in #1120
• Skip the pytest plugin end-to-end test on Emscripten by @henryiii in #1152

• Bump to Boost 1.91 by @henryiii in #1112
• Bump setup-uv to the maintained tag scheme by @henryiii in #1110
• Update cibuildwheel to v4.0 by @henryiii in #1119
• Update AGENTS instructions and ignore the CLAUDE symlink by @henryiii in #1121
• Make the nox hist session fail on a missing dir and mypy override errors by @henryiii in #1135
• Set up uv cache keys by @henryiii in #1142
• pybind11 v3 pin, MSVC /WX, config and stub cleanups by @henryiii in #1145
• Remove dead code and minor C++ cleanups by @henryiii in #1151
• Remove the needs-changelog labeler by @henryiii in #1164

• No storage option/support for serialization, by @henryiii in #1098

• Correct repr for MultiCell by @henryiii in #1087
• Fix serialization round-trip for 3D histograms with empty axes by @Copilot in #1094
• Fix __setitem__ axis index mismatch when slices are not leading indices by @Copilot in #1097
• Support 0D MeanView/WeightedMeanView by @Copilot in #1100
• Broadcast scalar axis args to match array sample length in Mean/WeightedMean fill by @Copilot in #1101

• Slightly better getitem type by @henryiii in #1086

• Use pattern matching in another place by @henryiii in #1092
• Move to using prek by @henryiii in #1105
• Add AGENTS.md by @henryiii in #1106

• Make the Histogram type generic, by @henryiii in #1081, #1082, #1083
• The MultiCell storage takes an argument (missed in typing), by @henryiii in #1080

• MultiCell storage, supporting a variable number of values, by @Superharz in #1008
• Record growth=True in writer-info (serialization), and respect it if present by @henryiii in #1074
• Drop Python 3.9 by @henryiii in #1057
• Drop Python 3.13t (use 3.14t instead) by @henryiii in #1079

• Don't add empty metadata dicts by @henryiii in #1048
• Avoid compiler warning: declaration of 'i' hides previous local declaration by @Superharz in #1068

• Refactor __getitem__ using pattern matching by @henryiii in #1067
• pytest log_level is better than log_cli_level by @henryiii in #1059
• Setting __file__ no longer required (pybind11 3.0) by @henryiii in #1066
• This should work on iOS now in cibuildwheel by @henryiii in #1049

This release fixes several issues with serialization as it's being prepared in Hist/uhi.

• Support __dict__ on histogram like axes #1041
• Allow library to be specified in writer_info removal #1042

• Avoid mutating lists given in indexing #1047
• Histogram metadata not being serialized correctly #1038
• Unset classic metadata shouldn't show up in serialization #1039
• Set dict correctly when unserializing #1040
• Serialization was broken for N>1D complex storage histograms #1043

This release adds support for UHI serialization for use with the uhi library, supported in the next release of uhi. A new diagnostic test has been added in the wheel, so you can quickly check your compile of boost-histogram without the full test suite. Quite a few fixes are present, as well, including a fix for *=, setting a range with a scalar, setting with a callable in a slice. Backend changes include many new clang-tidy checks that reduce the number of copies and refcount changes. We now build with pybind11 3.0.

This release drops support for Python 3.8, and adds Python 3.14(t), iOS, Windows ARM, and GraalPy.

• Initial support for uhi's serialization #997, #1030, #1033, #1034, #1035
• Drop Python 3.8 #976
• Support Windows ARM #1001
• Support Python 3.14 and 3.14t #1018
• Support iOS #1020, #1027
• Added diagnostics test #1022
• Recommend/use np.s_ for slicing in dicts #1002

• Support setting ranges with histograms (fixes *= too) #1036
• Support a callable in range setting #1036
• Support setting a range with a scalar #1000
• Account for fuzzy edge values in rebinning with edges #999
• Rename C++ metadata to raw_metadata #979
• Address lots of clang-tidy suggestions, should be faster with less copies/refcount bumps #1009, #1010, #1011, #1012, #1013, #1014, #1015

• Add CMake presets for testing and clang-tidy #1009

Fix for axis metadata not passing though non-uniform rebinning correctly. Flow bins are now preserved when doing a non-uniform rebinning. Also adds the ability to rebin by edges or an existing axis.

This is expected to be the last release with Python 3.8 support. Manylinux support will likely be bumped from 2024 to 2_28 soon, as well.

• Support edges= and axis= in bh.rebin #977
• Allow axis to be passed through bh.rebin #980, #982

• Axis metadata was broken when rebinning #978
• Flow bins were lost when using variable rebinning #977
• Scikit-build-core 0.11, with PEP 639 license metadata #986, #988
• Prepare for an potential upcoming pybind11 3 change #994

• Remove extras like [test] in favor of dependency-groups #973
• Multithreaded test running #987

• Make non-uniform rebinning work for Weight() and friends #972

• Support variable rebinning #913
• Removed support for Python 3.7 #952
• Added support for Python 3.13 and free-threaded Python 3.13t #950
• Add GitHub artifact attestations to package distribution #933

• Move build to scikit-build-core #887
• Clean up sdist files #928
• Minor touchup to build backend, use scikit-build-core 0.10+ #948

• Try cibuildwheel pyodide #935
• Try to fix pyodide build #934
• Update emscripten build #937
• Bump to Pyodide 3.12 #938
• Speed up pyodide job by a few seconds #941
• Update jobs and noxfile #929
• Use pylint 3.2 github formatter instead #932
• Add some CI speedups #939
• Faster cibuildwheel #942
• Adapt to final release of NumPy 2.1.0 #951

• NumPy 2 support. #918
• 32-bit Windows Python 3.12 wheel added (matching NumPy). #920

• Support filling Integer axes with unsigned integers #917
• Avoid triggering NumPy 2 dev release install on Python 3.12. #914

• Add missing API docs #909
• Use boost 1.84 #920

• overflow=False is now supported for IntCategory and StrCategory. #883

• Using _storage_type now produces a DeprecationWarning instead of PendingDeprecationWarning. #801
• Updated Boost to 1.82. The upper limit on Regular axes without overflow is now inclusive like NumPy. #802
• Produce more detailed error messages on C++ errors #848

• Make filling an integer axis with a float array (also) an error. #876
• Include -latomic on armv7l #823

• Add Python 3.12 support and binary wheels, also latest PyPy. manylinux2014+ required. #880, #878
• Drop Python 3.6 support. #798
• Drop pre-built wheels for 32-bit Linux (NumPy also dropped). #849
• Add testing for WebAssembly (Pyodide). #850
• Use Ruff #829

• Added storage_type operator and storage_type() function #781, with pending deprecation for _storage_type. #786 #790
• Better errors generated for missing or incorrect sample to mean storage. #782
• Better error message when views are set with an incompatible array. #794

• Patch broken sum with fully empty (0 bin) axis. #718
• Fix zero range bh.numpy.histogram to match numpy.histogram behavior. #721
• Avoid triggering __init__ when copying (better support for subclasses with custom init's). #759
• IntCategory now supports numbers larger than 2^24 (now 2^53). #792
• Pick a subset now supported inside a larger expression. #793

• Minor optimizations for UFuncs. #771
• Added Python 3.11 wheels. #789
• Include PyPy 3.9 binary wheels. #730
• Using pybind11 2.10 #767
• Explicit reset() documentation. #783
• Minor cleanup and further removal of a little Python 2 back-compat code.
• Warnings have better stacklevel settings.

• Fixed regression with invalid .project input causing segfaults. #708
• Minor skips for specific tests on ppc64le, primarily for a NumPy bug. #707
• Avoid using EH for program control, better on Pyodide. #709
• Fix regression with exact float not being accepted for .index for IntCategory in 1.3.0. Add hist nox session to check downstream (manually for the moment). #710

• PyPy 3.8 now supported with binary wheels. #677
• The GIL is released a little more often now. #662
• AxesTuple does not allow construction of non-axes. #680
• KeyError is now thrown when accessing a non-existent item in a Category Axis #689
• WeightedViews now support np.cumsum #699

• Fixed WeightedMean storages producing NaN for .variances() #695
• Modify local include slightly to enable WebAssembly compilation in Pyodide #702

• Use PyLint in CI to check for some style issues #690
• Developer (CMake) installs no longer require toml #698

• musllinux wheels now provided along with manylinux #656

• Fixed single-element negative growth fill #654

• No longer require Docker for clang-format, runs online too #610
• Using pybind11 2.8.0 #658

• Python 3.10 officially supported, with wheels.
• Support subtraction on histograms #636
• Integer histograms are now signed #636

• Support custom setters on AxesTuple subclasses. #627
• Faster picking if slices are not also used #645 or if they are #648 (1000x or more in some cases)
• Throw an error when an AxesTuple setter is the wrong length (inspired by zip strict in Python 3.10) #627
• Fix error thrown on comparison with axis and non-axis object #631
• Static typing no longer thinks storage= is required #604

• Support NumPy 1.21 for static type checking #625
• Use newer Boost 1.77 and Boost.Histogram 1.77+1 #594
• Provide nox support #647

• Experimentally support list selection on categorical axes #577
• Support Python 3.8 on Apple Silicon #600
• Scaling and addition with a scalar affect flow bins too #580
• Change sum_of_deltas_squared to _sum_of_deltas_squared (was an implementation detail) #602

• Fix "picking" on a flow bin #576
• Better error message on getattr #596

• Test on Python 3.10 beta releases #600
• Provide a CMakeLists for quick standalone Boost.Histogram C++ experiments #591
• Adding logging with pytest failure output #575

• Fix scaling a weighted storage #559
• Fix partial summation over a Categorical axis #564
• Support running type checking from Python < 3.8 #542

• A family= is no longer required if you only subclass Histogram. #533

• Fix summing of Mean/WeightedMean accumulators #537
• Added missing dependency on typing_extensions for Python 3.6 & 3.7 #529

• Added Ellipsis support to typing. #525
• Better typing for Views. #530
• Fixed issue with Histogram copy constructor requiring metadata #532

Dropped support for Python 2 and 3.5; removed large numbers of workarounds. Fully statically typed. API compatible with the final 0.x release for most uses, except for subclassing; subclassing histogram components now uses Python 3 class keyword syntax to set families.

• Dropped Python 2.7 and 3.5 support #512
• Removed deprecated .options from axes. Use .traits instead. #503
• Full static typing available, UHI 0.1.2+ supported. #516, #517, #519, #520, #521, #523

• Use keyword class family setting when subclassing histogram components instead of custom decorator. #513
• Structure of internal repr creation changed and made slightly more public. #518

• Consistently show metadata= in repr if present; refactored internal repr handling #518
• Minor typing related fixes for rare bugs (especially in numpy.py, #521)

• Backport fix scaling a weighted storage
• Backport fix partial summation over a Categorical axis

• Backport fix for Mean/WeightedMean summing.
• Backport fix for boost_histogram.numpy density.
• Backport missing metadata from the repr's.
• Ignore family= on Histogram subclassing to make subclassing Histogram only possible in 1.x + 0.x code.

PlottableProtocol provides a way to plot in different libraries, and easy access to common quantities. This is expected to be the final release for Python 2, and mostly equivalent in API to 1.0.

• Support for PlottableProtocol. You can now access .values(), .counts(), and .variances() on all storages; used by plotting libraries. .kind describes the Kind of the histogram (bh.Kind.COUNT or bh.Kind.MEAN). .options has been renamed to .traits, and a few more useful traits were added, like .discrete. Most other portions of the Protocol were already present. #476
• Removed deprecated .rank on histograms (since 0.8). Use .ndim instead. #505
• Supports converting user histogram objects that provide a _to_boost_histogram_ method. #483
• A view=True parameter must now be passed to get a View instead of a standard NumPy values array from to_numpy(). #498

• Added additional support for typing, fixing a couple of rare Python 2 bugs in the process #493.
• The resulting histogram from bh.numpy.* functions is now reducible #508

• Use GitHub Actions for ARM compiling #474
• Apple Silicon support (since 0.12) #495
• Support compiling with C++17 #502
• Rename NPY_NUM_BUILD_JOBS to CMAKE_BUILD_PARALLEL_LEVEL for consistency with other Scikit-HEP projects. #502

Pressing forward to 1.0.

• You can now set all complex storages, either on a Histogram or a View with an (N+1)D array #475
• Axes are now normal __dict__ classes, you can manipulate the __dict__ as normal. Axes construction now lets you either use the old metadata shortcut or the __dict__ inline. #477

• Fixed slicing projection with one-sided slices #479
• Fixed issue if final bin of Variable histogram was infinite by updating to Boost 1.75 #470
• NumPy arrays can be used for weights in bh.numpy #472
• Vectorization for WeightedMean accumulators was broken #475

• Bumped to pybind11 version #470
• Black formatting used in notebooks too #470

If you are using Axis.options, please transition to Axis.traits. traits includes all the old options, along with some new traits, and matches the PlottableProtocol requirements.

Updating pybind11 to 2.6.0. #443 Features:

• Python 3.9 support
• PyPy2 / PyPy3.6 / PyPy3.7 support
• Warnings on latest AppleClang fixed
• 40% faster accumulator fills, simpler implementation
• Segfaults when passing an object with a throwing repr fixed
• kwargs replaced older workarounds (partially at the moment)
• Using new py::type instead of pybind11::detail usage
• Enhanced CMake support, finds conda and venv now, uses pybind11_find_import
• Using setuptools support from pybind11 (previously vendored, so benefits have been available since 0.11.0)

Also cleans up SDists a bit. #467

A release focused on preparing for the upcoming Hist 2.0 release.

• Arbitrary items can be set on an axis or histogram. #450, #456
• Subclasses can customize the conversion procedure. #456

• Fixed reading pickles from boost-histogram 0.6-0.8 #445
• Minor correctness fix #446
• Accidental install of typing on Python 3.5+ fixed
• Scalar ND fill fixed #453

• Updated to Boost 1.74 #442
• CMake installs version.py now too #449
• Updated setuptools infrastructure no longer requires NumPy #451
• Some basic clang-tidy checks are now being run #455

Quick fix for extra print statement in fill.

• Fixed debugging print statement in fill. #438

• Added CI/pre-commit check for print statements #438
• Formatting CMakeLists now too #439

Several fixes were made, mostly related to Weight storage histograms from Uproot 4.

• Reduction on h.axes.widths supported again #428
• WeightedSumView supports standard array operations #432
• Operations shallow copy (non-copyable metadata supported) #433
• Pandas Series as samples/weights supported #434
• Support NumPy scalars in operations #436

This version was released during PyHEP 2020. Several improvements were made to usability when plotting and indexing.

• AxesTuple array now support operations via ArrayTuple #414
• Support sum and bh.rebin without slice #424
• Nicer error messages in some cases #415
• Made a few properties hidden for accumulators that were not public #418
• Boolean now supports reduction, faster compile #422
• AxesTuple now available publicly for subprojects #419

• Histograms support operations with arrays, no longer take the first element only #417

This version was released just before PyHEP 2020. Several important fixes were made, along with a few new features to better support downstream projects.

• metadata supported and propagated on Histograms (slots added) #403
• Added dd=True option in to_numpy #406
• Deprecated cpp module removed #402

• Subclasses can override axes generation #401
• [dev] extra now installs pytest #401

• Fix numpy.histogramdd return structure #406
• Travis deploy multi-arch fixes #399
• Selecting on a bool axes supports 2D+ histograms #398
• Warnings fixed on NumPy 1.19+ #404

This version was released just before SciPy 2020 and Boost 1.74. Highlights include better accumulator views, simpler summing, better NumPy and Pandas compatibility, and sums on growing axes. Lots of backend work, including a new wheel building system, internal changes and better reliance on Boost.Histogram's C++ tools for actions like cropping.

• Weighted histogram cells can now be assigned directly from iterables #375
• Weighted views can be summed and added #368
• Sum is now identical to the built-in sum function #365
• Adding growing axis is better supported #358
• Slicing an AxesTuple now keeps the type #384
• ndim replaces rank for NumPy compatibility #385
• Any array-like supported in fill #391, any iterable can be used for Categories #392
• Added Boolean axes, from Boost.Histogram 1.74 #390
• Division between histograms is supported #393
• More deprecated functionality removed

• Support older versions of CloudPickle (issue also fixed upstream) #343
• Drop extra printout #338
• Throw an error instead of returning an incorrect result in more places #386

• Update Boost to 1.73 #359, pybind11 to 2.5.0 #351, Boost.Histogram to pre-1.74 #388
• Cropping no longer uses workaround #373
• Many more checks added to pre-commit #366
• Deprecating cpp interface #391
• Wheelbuilding migrated to cibuildwheel and GHA #361

This version removes deprecated functionality, and has several backend improvements. The most noticeable user-facing change is the multithreaded fill feature, which can enable significant speedups when you have a dataset that is much larger than the number of bins in your histogram and have free cores to use. Several small bugs have been fixed.

• Added threads= keyword to .fill and NumPy functions; 0 for automatic, default is 1 #325
• .metadata is now settable directly from the AxesTuple #303
• Deprecated items from 0.5.x now dropped #301
• cpp mode updates and fixes #317

• Dict indexing is now identical to positional indexing, fixes "picking" axes in dict #320
• Passing samples=None is now always allowed in .fill #325

• Build system update, higher requirements for developers (only) #314
  • Version is now obtained from setuptools_scm, no longer stored in repo
• Removed futures requirement for Python 2 tests
• Updated Boost.Histogram, cleaner code with fewer workarounds

Common analysis tasks are now better supported. Much more complete documentation. Now using development branch of Boost.Histogram again.

• Fix sum over category axes in indexing #298
• Allow single category item selection #298
• Allow slicing on axes without flow bins #288, #300
• Sum repr no longer throws error #293

• Now using scikit-hep/azure-wheel-helpers via subtree #292

Examples and notebooks are now up to date with the current state of the library. Using Boost 1.72 release.

• Slices and single values can be mixed in indexing #279
• UHI locators supported on axes #280

• Properties on accumulator views now resolve correctly #273
• Division of a histogram by a number is supported again #278
• Setting a histogram with length one slice fixed #279
• NumPy functions now work with NumPy ints in bins= #282
• In-place addition avoids a copy #284

This version fills out most of the remaining features missing from the 0.5.x series. You can now use all the storages without the original caveats; even the accumulators can be accessed array-at-a-time without copy, pickled quickly, and set array-at-a-time, as well.

The API has changed considerably, providing a more consistent experience in Python. Most of the classic API still works in this release, but will issue a warning and will be removed from the next release. Please use this release to transition existing 0.5.x code to the new API.

• Histogram and Axis classes now follow PEP 8 naming scheme (histogram->Histogram, regular->Regular, int->Int64 etc.) #192, #255
• You can now view a histogram with accumulators, with property access such as h.view().value #194
• Circular variable and integer axes added #231
• Split Category into StrCategory and IntCategory, now allows empty categories when growth=True #221
• StrCategory fills are safer and faster #239, #244
• Added axes transforms #192
• Function(forward, inverse) transform added, allowing ultra-fast C function pointer transforms #231
• You can now set histogram contents directly #250
• You can now sum over a range with endpoints #185
• h.axes now has the functions from axis as well. #183
• bh.project has become bh.sum #185
• .reduce(...) and the reducers in bh.algorithm have been removed in favor of dictionary based UHI slicing #259
• bh.numpy module interface updates, histogram=bh.Histogram replaces cryptic bh=True, and density=True is now supported in NumPy mode #256
• Added hist.copy() #218 and hist.shape #264
• Signatures are much nicer in Python 3 #188
• Reprs are better, various properties like __module__ are now set correctly #200

• Unlimited and AtomicInt storages now allow single item access #194
• .view() now no longer makes a copy #194
• Fixes related to string category axis fills #233, #230
• Axes are no longer copies, support setting metadata #238, #246
• Pickling accumulator storages is now comparable in performance simple storages #258

• The linux wheels are now 10-20x smaller #229
• The hist/axis classes are now pure Python, with a C++ object inside #183
• Most internal names changed, core->_core, etc. #183
• The uhi module is now tag. #183
• boost_histogram.cpp as bh provides C++ high-compatibility mode. #183
• Indexing tags now use full UHI instead of workarounds #185
• Removed log and sqrt special axes types#231
• Family and registration added, new casting system #200

• bh.loc supports an offset #164
• Nicer reprs in several places #167
• Deprecate .at and .axis #170

• Use relative paths in setup.py to avoid resolving WSL paths on Windows #162, #163
• Better pybind11 support for Python 3.8 #168

• Serialization code shared with Boost.Histogram #166
• Avoid unused PEP 517 isolation for now #171 (may return with proper PEP 518 support eventually)

• Removed the bh.indexed/h.indexed iterator #150
• Added .axes AxisTuple, with direct access to properties #150
• Cleaned up tab completion in IPython #150

• Fixed a bug in the sdist missing Boost.Variant2 #154
• Fixed filling on strided inputs #158

First beta release and beginning of the changelog.

• Unlimited storage does not support pickling or classic multiprocessing
• Some non-simple storages do not support some forms of access, like .view
• Indexing and the array versions (such as centers) are incomplete and subject to change
• The numpy module is provisional and subject to change
• Docstrings and signatures will improve in later versions (especially on Python 3)