Changelog

 since version 1.7.2 release
 ===========================
 * raise more informative error when trying to iterate or
   perform a membership operation on a Dataset (issue #1383)
 * fix type hint for createEnumType (issue #1378)
 * add python 3.13 to windows wheel builds (PR #1377)

 version 1.7.2 (tag v1.7.2rel)
 =============================
 * add static type hints (PRs #1302, #1349)
 * Expose nc_rc_set, nc_rc_get (via rc_set, rc_get module functions). (PR #1348)
 * Add Variable.get_fill_value and allow `fill_value='default'` to 
   set `_FillValue` attribute using default fill values.  (issue #1374, PR #1375).
 * Fix NETCDF3 endian error (issue #1373, PR #1355).

 version 1.7.1 (tag v1.7.1rel)
===============================
 * include nc_complex source code from v0.2.0 tag (instead of using submodule).
 * add aarch64 wheels.

 version 1.7.0 (tag v1.7.0rel)
===============================
 * add support for complex numbers via `auto_complex` keyword to `Dataset` (PR #1295)
 * fix for deprecated Cython `DEF` and `IF` statements using compatibility header
   with shims for unavailable functionality (PR #1277)
 * use `szip` as the library name on Windows (PR #1304)
 * add support for MS-MPI `MPI_Message` detection (PR #1305)
 * fix for issue #1306 - surprising result when indexing vlen str with non-contiguous
   indices.
 * Fix bug in set_collective introduced in PR #1277 (collective mode was
   always set).

 version 1.6.5 (tag v1.6.5rel)
===============================
 * fix for issue #1271 (mask ignored if bool MA assigned to uint8 var)
 * include information on specific object when reporting errors from netcdf-c
 * python 3.12 wheels added, support for python 3.7 removed.

 version 1.6.4 (tag v1.6.4rel)
===============================
 * set path to SSL certificates internally, so https DAP URLs work with wheels
   (issue #1246, requires nc_rc_set function available starting with netcdf-c
   4.9.1, plus bugfix in netcdf-c PR #2690).
   Added certifi as a dependency.
 * Added `isopen` method to `MFDataset` object to check if underlying files are open.

 version 1.6.3 (tag v1.6.3rel)
==============================
 * Use ``nc_put_vars`` for strided writes for netcdf-c >= 4.6.2 (issue #1222).
 * _Unsigned="false" should be same as not having _Unsigned set (issue #1232).
   _Unsigned now must be set to "true" or "True" for variable to be interpreted
   as unsigned, instead of just having _Unsigned be set (to anything).
 * pypi wheels built with netcdf-c 4.9.1.

 version 1.6.2 (tag v1.6.2rel)
==============================
 * Added ``netCDF4.__has_set_alignment__`` property to help identify if the
   underlying netcdf4 supports setting the HDF5 alignment.
 * Slicing multi-dimensional variables with an all False boolean index array
   now returns an empty numpy array (instead of raising an exception - issue #1197).
   Behavior now consistent with numpy slicing.
 * fix problem with compiling using netcdf-c < 4.9.0 (issue #1209)
 * pypi wheels build with netcdf-c 4.9.0.

 version 1.6.1 (tag v1.6.1rel)
==============================
 * add Dataset methods has_<name>_filter (where <name>=zstd,blosc,bzip2,szip)
   to check for availability of extra compression filters.
 * release GIL for all C-lib calls (issue #1180).
 * Add support for nc_set_alignment and nc_get_alignment to control alignment
   of data within HDF5 files.

 version 1.6.0 (tag v1.6.0rel)
==============================
 * add support for new quantization functionality in netcdf-c 4.9.0 via "signficant_digits"
   and "quantize_mode" kwargs in Dataset.createVariable. Default quantization_mode is "BitGroom",
   but alternate methods "BitRound" and GranularBitRound" also supported.
 * opening a Dataset in append mode (mode = 'a' or 'r+') creates a Dataset
   if one does not already exist (similar to python open builtin).  Issue #1144.
   Added a mode='x' option (as in python open) which is the same as mode='w' with
   clobber=False.
 * allow createVariable to accept either Dimension instances or Dimension
   names in "dimensions" tuple kwarg (issue #1145).
 * remove all vestiges of python 2 in _netCDF4.pyx and set cython language_level
   directive to 3 in setup.py.
 * add 'compression' kwarg to createVariable to enable new compression
   functionality in netcdf-c 4.9.0.  'None','zlib','szip','zstd','bzip2'
   'blosc_lz','blosc_lz4','blosc_lz4hc','blosc_zlib' and 'blosc_zstd'
   are currently supported. 'blosc_shuffle',
   'szip_mask' and 'szip_pixels_per_block' kwargs also added.
   compression='zlib' is equivalent to (the now deprecated) zlib=True.
   If the environment variable NETCDF_PLUGIN_DIR is set to point to the
   directory with the compression plugin lib__nc* files, then the compression plugins will
   be installed within the package and be automatically available (the binary
   wheels have this).  Otherwise, the environment variable HDF5_PLUGIN_PATH
   needs to be set at runtime  to point to plugins in order to use the new compression
   options.
 * MFDataset did not aggregate 'name' variable attribute (issue #1153).
 * issue warning instead of raising an exception if missing_value or
   _FillValue can't be cast to the variable type when creating a
   masked array (issue #1152).
 * Define MPI_Session for compatibility with current mpi4py (PR #1156).

 version 1.5.8 (tag v1.5.8rel)
==============================
 * Fix Enum bug (issue #1128): the enum_dict member of an EnumType read from a file
   contains invalid values when the enum is large enough (more than 127 or 255
   members).
 * Binary wheels for aarch64 and python 3.10.

 version 1.5.7 (tag v1.5.7rel)
==============================
 * don't try to mask vlens with default  _FillValue, since vlens don't have a default _FillValue.
   This gets rid of numpy DeprecationWarning (issue #1099).
 * update docs to reflect the fact that a variable must be in collective mode before writing
   compressed data to it in parallel. Added a test for this (examples/mpi_example_compressed.py).
   Issue #1108.
 * Fix OverflowError when dimension sizes become greater than 2**32-1 elements on Windows (Issue #1112).
 * Don't return masked arrays for vlens (only for primitive and enum types - issue #1115).

 version 1.5.6 (tag v1.5.6rel)
==============================
 * move CI/CD tests from travis/appveyor to Github Actions (PR #1061).
 * move netCDF4 dir under src so module can be imported in source directory (PR #1062).
 * change numpy.bool to numpy.bool_ and numpy.float to numpy.float_ (float and
   bool are deprecated in numpy 1.20, issue #1065)
 * clean up docstrings so that they work with latest pdoc.
 * update cython numpy API to remove deprecation warnings.
 * Add "fromcdl" and "tocdl" Dataset methods for import/export of CDL
   via ncdump/ncgen called externally via the subprocess module (issue #1078).
 * remove python 2.7 support.
 * broadcast data (if possible)to conform to variable shape when writing to a slice
   (issue #1083).

 version 1.5.5.1 (tag v1.5.5.1rel)
==================================
 * rebuild binary wheels for linux and OSX to link netcdf-c 4.7.4 and hdf5 1.12.0.

 version 1.5.5 (tag v1.5.5rel)
==============================
 * have setup.py always try use nc-config first to find paths to netcdf and
 hdf5 libraries and headers. Don't use pkg-config to find HDF5 if HDF5 env
 vars are set (or read from setup.cfg).
 * Change MIT license text to standard OSI wording (PR #1046).

 version 1.5.4 (tag v1.5.4rel)
==============================
 * fix printing of variable objects for variables that end with the letter 'u'
   (issue #983).
 * make sure root group has 'name' attribute (issue #988).
 * add the ability to pack vlen floats to integers using
   scale_factor/add_offset (issue #1003)
 * use len instead of deprecated numpy.alen (issue #1008)
 * check size on valid_range instead of using len (issue #1013).
 * add `set_chunk_cache/get_chunk_cache` module functions to reset the
   default chunk cache sizes before opening a Dataset (issue #1018).
 * replace use of numpy's deprecated tostring() method with tobytes()
   (issue #1023).
 * bump minimal numpy version to 1.9 (first version to have tobytes()).

 version 1.5.3 (tag v1.5.3rel)
==============================
 * make sure arrays are masked that are not filled when auto_fill is off
   (issue #972).
 * python 3.8 binary wheels.

 version 1.5.2 (tag v1.5.2rel)
==============================
 * fix for scaling bug when _Unsigned attribute is set and byteorder of data
   does not match native byteorder (issue #930).
 * revise documentation for Python 3 (issue #946).
 * establish support for Python 2.7, 3.5, 3.6 and 3.7 (issue #948).
 * use dict built-in instead of OrderedDict for Python 3.7+
   (pull request #955).
 * remove underline ANSI in Dataset string representation (pull request #956).
 * remove newlines from string representation (pull request #960).
 * fix for issue #957 (size of scalar var is a float since numpy.prod(())=1.0).
 * make sure Variable.setncattr fails to set _FillValue (issue #959).
 * fix detection of parallel HDF5 support with netcdf-c 4.6.1 (issue #964).

 version 1.5.1.2 (tag v1.5.1.2rel)
==================================
 * fix another slicing bug introduced by the fix to issue #906 (issue #922).

 version 1.5.1.1 (tag v1.5.1.1rel)
==================================
 * fixed __version__ attribute (was set incorrectly in 1.5.1 release).
 * fix for issue #919  (assigning 2d array to 3d variable with singleton
   first dimension with v[:] = a).
 * minimum numpy changed from 1.9.0 to 1.10.0.

 version 1.5.1 (tag v1.5.1rel)
==============================
 * fix issue #908 by adding workaround for incorrect value returned
   by nc_inq_var_fill for netcdf-c < 4.5.1.
 * fix bug writing slice to unlimited dimension that is not the first
   (leftmost).  Issue  #906.
 * make sure data gets converted to type of scale_factor when add_offset=0
   and scale_factor=1 (issue #913).
 * fix for reading empty (NIL) string attributes (issue #915).

 version 1.5.0.1 (tag v1.5.0.1rel)
==================================
 * binary wheels for linux and macosx rebuilt against netcdf-c 4.6.3 (instead
   of 4.4.1.1).
 * add read-shared mode (mode='rs'). Significantly speeds up reads of NETCDF3
   files (pull request #902).

 version 1.5.0 (tag v1.5.0rel)
===============================
 * added support for parallel IO in the classic netcdf-3 formats through the
   pnetcdf library (pull request #897).

 version 1.4.3.2 (tag v1.4.3.2)
===============================
 * include missing membuf.pyx file in release source tarball.

 version 1.4.3.1 (tag v1.4.3.1)
===============================
 * fix bug in implementation of NETCDF4_CLASSIC support for parallel IO
   in v1.4.3 release.

 version 1.4.3 (tag v1.4.3rel)
=============================
 * make set_always_mask work in MFDataset.
 * fix saving diskless files to disk with netcdf-c >= 4.6.2.
 * write to an in-memory Dataset, memoryview buffer returned by Dataset.close()
   (issue #865, requires netcdf-c >= 4.6.2)
 * fix performance regression when using large sequences of consecutive
   integers for indexing with netcdf-c >= 4.6.2 (issue #870).
 * improved error messages for ncinfo and other utilities (issue #873).
 * fix for int64 attributes not being created for NETCDF3_64BIT_DATA (CDF5)
   files (issue #878).
 * fix for MPI parallel error ("NetCDF: Attempt to use feature that was not
   turned on when netCDF was built") using netcdf-c 4.6.2 (issue #883).
 * Added methods `set_ncstring_attrs()` to Dataset, Group and Variable that
   forces all text attributes to be written as variable length strings (netCDF
   type NC_STRING - issue #882).
 * Allow parallel mode with NETCDF4_CLASSIC files (issue #890).

 version 1.4.2 (tag v1.4.2rel)
=============================
 * add get_dims Variable method (issue #824)
 * make sure format keyword not ignored when mode is 'ws' (issue #827)
 * fix numpy FutureWarning (non-tuple sequence for
   multidimensional indexing is deprecated), issue #833.
 * add 'master_file' kwarg to MFDataset.__init__ (issue #835).
 * always use nc_get_vars for strided access over OpenDAP (issue #838).
 * raise FutureWarning when trying to set multi-dimensional array attribute
   while still silently flattening the array (issue #841). Will change
   to ValueError in next release (1.4.3).
 * fix parallel writes when both nc4 parallel and pnetcdf parallel options
   enabled in the netcdf-c library (issue #820).
 * fix for writing masked scalar character variable (issue #850).

 version 1.4.1 (tag v1.4.1rel)
=============================
 * disable workaround for slow nc_get_vars for __netcdflibversion__ >= 4.6.2,
   since a fix was added to speed up nc_get_vars in the C library.  Issue 680.
 * new Dataset and Variable methods (set_always_mask) to optionally
   re-enable old behaviour (return masked arrays
   only if selected slice contains missing values) (issue #809).

 version 1.4.0 (tag v1.4.0rel)
=============================
 * fixed bug in detection of CDF5 library support in setup.py (pull request
   #736, issue #713).
 * fixed reading of variables with zero-length dimensions in NETCDF3_CLASSIC
   files (issue #743).
 * allow integer-like objects in VLEN slices (not just python ints, issue
   #526, pull request #757).
 * treating _FillValue as a valid_min/valid_max was too surprising, despite
   the fact the thet netcdf docs 'attribute best practices' suggests that
   clients should to this.  Revert this change from issue #576 (issue #761).
 * remove netcdftime, since it is now a separate package. date2num, num2date
   and date2index still importable from netCDF4.
 * fix 'Unreachable code' cython warning (issue #767).
 * Change behavior of string attributes so that nc.stringatt = ['foo','bar']
   produces an vlen string array attribute in NETCDF4, instead of concatenating
   into a single string ('foobar').  In NETCDF3/NETCDF4_CLASSIC, an IOError
   is now raised, instead of writing 'foobar'. Issue #770.
 * fix loading of enum type names (issue #775).
 * make sure missing_value applies only to scaled short integers if
   auto-scaling is on (issue #777).
 * automatically create views of compound types with character arrays as
   numpy strings (issue #773).  Can be disabled using
   'set_auto_chartostring(False)'. Numpy structured
   array dtypes with 'SN' string subtypes can now be used to
   define netcdf compound types (they get converted to ('S1',N)
   character array types automatically).
 * always return masked array by default, even if there are no
   masked values (too surprising to get ndarray or MaskedArray depending
   on slice, issue #785).
 * treat valid_min/valid_max/_FillValue/missing_value as unsigned
   integers if _Unsigned is set (to mimic behaviour of netcdf-java).
   Conversion to unsigned type now occurs before masking and scale/offset
   operation. Issue #794.


 version 1.3.1 (tag v1.3.1rel)
=============================
 * add parallel IO capabilities.  netcdf-c and hdf5 must be compiled with MPI
   support, and mpi4py must be installed.  To open a file for parallel access,
   use `parallel=True` in `Dataset.__init__` and optionally pass the mpi4py Comm instance
   using the `comm` kwarg and the mpi4py Info instance using the `info` kwarg.
   IO can be toggled between collective and independent using `Variable.set_collective`.
   See `examples/mpi_example.py`. Issue #717, pull request #716.
   Minimum cython dependency bumped from 0.19 to 0.21.
 * Add optional `MFTime` calendar overload to use across all files, for example,
   `'standard'` or `'gregorian'`. If `None` (the default), check that the calendar
   attribute is present on each variable and values are unique across files raising
   a `ValueError` otherwise.
 * Allow _FillValue to be set for vlen string variables (issue #730).

 version 1.3.0 (tag v1.3.0rel)
==============================
 * always search for HDF5 headers when building, even when nc-config is used
   (since nc-config does not always include the path to the HDF5 headers).
   Also use H5get_libversion to obtain HDF5 version info instead of
   H5public.h. Fixes issue #677.
 * encoding kwarg added to Dataset.__init__ and Dataset.filepath (default
   is to use sys.getfilesystemencoding()) so that oddball
   encodings (such as cp1252 on windows) can be handled in Dataset
   filepaths (issue #686).
 * Calls to nc_get_vars are avoided, since nc_get_vars is very slow (issue
   #680).  Strided slices are now converted to multiple calls to
   nc_get_vara.  This speeds up strided slice reads by a factor of 10-100
   (especially for NETCDF4/HDF5 files) in most cases. In some cases, strided reads
   using nc_get_vars are faster (e.g. strided reads over many dimensions
   such as var[:,::2,::2,::2])), so a variable method use_nc_get_vars was added.
   var.use_nc_get_vars(True) will tell the library to use nc_get_vars instead
   of multiple calls to nc_get_vara, which was the default behaviour previous
   to this change.
 * fix utc offset time zone conversion in netcdftime - it was being done
   exactly backwards (issue #685 - thanks to @pgamez and @mdecker).
 * Fix error message for illegal ellipsis slicing, add test (issue #701).
 * Improve timezone format parsing in netcdftime
   (https://github.com/Unidata/netcdftime/issues/17).
 * make sure numpy datatypes used to define CompoundTypes have
   isalignedstruct flag set to True (issue #705), otherwise.
   segfaults can occur. Fix required raising them minimum numpy requirement
   from 1.7.0 to 1.9.0.
 * ignore missing_value, _FillValue, valid_range, valid_min and valid_max
   when creating masked arrays if attribute cannot be safely
   cast to variable data type (and issue a warning).  When setting
   these attributes don't cast to variable dtype unless it can
   be done safely and issue a warning. Issue #707.

 version 1.2.9 (tag v1.2.9rel)
==============================
 * Fix for auto scaling and masking when _Unsigned attribute set (create
   view as unsigned type after scaling and masking). Issue #671.
 * Always mask values outside valid_min, valid_max (not just when
   missing_value attribute present).  Issue #672.
 * Fix setup.py so pip install doesn't fail if cython not installed.
   setuptools >= 18.0 now required for installation (Issue #666).

 version 1.2.8 (tag v1.2.8rel)
==============================
 * recognize _Unsigned attribute used by netcdf-java to designate unsigned
   integer data stored with a signed integer type in netcdf-3 (issue #656).
 * add Dataset init memory parameter to allow loading a file from memory
   (pull request #652, issues #406 and #295).
 * fix for negative times in num2date (issue #659).
 * fix for failing tests in numpy 1.13 due to changes in numpy.ma (issue #662).
 * Checking for _Encoding attribute for NC_STRING variables, otherwise use
   'utf-8'. 'utf-8' is used everywhere else, 'default_encoding' global module
   variable is no longer used.  getncattr method now takes optional kwarg
   'encoding' (default 'utf-8') so encoding of attributes can be specified
   if desired. If _Encoding is specified for an NC_CHAR ('S1') variable,
   the chartostring utility function is used to convert the array of
   characters to an array of strings with one less dimension (the last
   dimension is interpreted as the length of each string) when reading the
   data. When writing the data, stringtochar is used to convert a numpy
   array of fixed length strings to an array of characters with one more
   dimension. chartostring and stringtochar now also have an 'encoding' kwarg.
   Automatic conversion to/from character to string arrays can be turned off
   via a new set_auto_chartostring Dataset and Variable method (default
   is True). Addresses issue #654.
 * Cython >= 0.19 now required, _netCDF4.c and _netcdftime.c removed from
   repository.

 version 1.2.7 (tag v1.2.7rel)
==============================
 * fix for issue #624 (error in conversion to masked array when variable slice
   returns a scalar). This is a regression introduced in 1.2.5 associated
   with support for vector missing_values. Test (tst_masked5.py) added for
   vector missing_values.
 * fix for python 3.6 compatibility (error retrieving character _FillValue attribute,
   issue #626). Test with python 3.6 using travis CI.

 version 1.2.6 (tag v1.2.6rel)
==============================
 * fix some test failures on big endian PPC64 that were due to
   errors in byte-swapping logic. Also fixed bug in enum
   code exposed on PPC64 (issue #608).
 * remove support for python 2.6 (it probably still will work for a while
   though).
 * Sometimes checking that data being assigned to a variable has
   an 'ndim' attribute is not sufficient, instead check to see that
   the object supports the buffer interface (issue #613).
 * make get_variables_by_attributes work in MFDataset (issue #610)
   The hack is also applied for set_auto_maskandscale, set_auto_scale,
   set_automask, so these don't have to be duplicated in MFDataset (pull
   request #571).

 version 1.2.5 (tag v1.2.5rel)
==============================
 * Add MFDataset.set_auto_maskandscale (plus set_auto_scale, set_auto_mask).
   Fixes issue #570.
 * Use valid_min/valid_max/valid_range attributes when defining mask
   (issue #576).  Values outside the valid range are considered to
   be missing when defining the mask.
 * Fix for issue #584 (add support for dates before -4712-1-1 in 360_day
   and 365_day calendars to netcdftime.utime).
 * Fix for issue #593: add support for datetime.timedelta operations
   (adding and subtracting timedelta, subtracting two datetime
   instances to compute time duration between them), implement
   datetime.replace() and datetime.__str__(). datetime.__repr__()
   includes the full state of an instance. Add datetime.calendar.
   datetime comparison operators have full accuracy now.
 * Fix for issue #585 by increasing the size of the buffer used to store the
   filepath.
 * Fix for issue #592: Add support for string array attributes. (When
   reading, a vlen string array attribute is returned as a list of
   strings. To write, use var.setncattr_string("name", ["two", "strings"]).)
 * Fix for issue #596 - julian day calculations wrong for negative years,
   caused incorrect roundtrip num2date(date2num(date)) roundtrip for dates with year
   < 0.
 * Make sure negative years work in utime.num2date (issue #596).
 * raise NotImplementedError when trying to pickle Dataset, Variable,
   CompoundType, VLType, EnumType and MFDataset (issue #602).
 * Fix for issue #527: initialize vldata[i].p in Variable._get(...).

 version 1.2.4 (tag v1.2.4rel)
==============================
 * Fix for issue #554.  It is now ensured that data is in native endian
   byte order before passing to netcdf-c library.  Data read from variable
   with non-native byte order is also byte-swapped, so that dtype remains
   consistent with netcdf variable.  Behavior now consistent with h5py.
 * raise warning for HDF5 1.10.x (issue #549), since backwards
   incompatible files may be created.
 * raise AttributeError instead of RuntimeError when attribute operation
   fails.  raise IOError instead of RuntimeError when nc_create or
   nc_open fails (issue #546).
 * Use NamedTemporaryFile instead of deprecated mktemp in tests
   (pull request #543).
 * add AppVeyor automated windows tests (pull request #540).

 version 1.2.3.1 (tag v1.2.3.1rel)
==================================
 * fix bug in setup.py (pull request #539, introduced in issue #518).

 version 1.2.3 (tag v1.2.3rel)
==============================
 * try to avoid writing NC_STRING attributes if possible, by
   trying to convert unicode strings to ascii and write as NC_CHAR (issue
   #529).  This preserves compatibility with clients (like Matlab) that
   can't deal with NC_STRING attributes. A 'setncattr_string' method
   was added for Dataset and Variable to that users can force attributes
   to be written as NC_STRING if necessary.
 * fix failing tests with numpy 1.11 (issues #521 and #522).
 * fix indentation bug in nc4tonc3 utility (issue #519).
 * add the capability in setup.py to use pkg-config instead of
   nc-config (pull request #518).
 * make sure slices which return scalar masked arrays
   are consistent with numpy.ma (issue #515).
 * add test/tst_cdf5.py and test/tst_filepath.py (to test new
   NETCDF3_64BIT_DATA format and filepath Dataset method).
 * expose netcdftime.__version__ (issue #504).
 * fix potential memory leak in Dataset.filepath in attempt to fix
   mysterious segfaults on CentOS6 (issue #506). Segfaults
   can apparently still occur on systems like CentOS6 with old versions of glibc.

 version 1.2.2 (tag v1.2.2rel)
 =============================
 * fix failing tests on python 2.6 (issue #497). Change minimum required
   python from 2.5 to 2.6.
 * Potential memory leaks fixed by freeing string pointers internally allocated
   in netcdf-c using nc_free_string. Also use nc_free_vlens to free space allocated
   for vlens inside netcdf-c (issue #495).
 * invoke str on filename argument to Dataset constructor, so pathlib
   instances can be used (issue #489).
 * don't use hardwired NC_MAX_DIMS or NC_MAX_VARS internally to allocate space
   for dimension or variable ids.  Instead, find out the number of dims
   and vars and use malloc.  NC_MAX_NAME is still used to allocate space
   for attribute and variable names, since there is no obvious way to
   determine the length of these names.
 * if trying to write a unicode attribute, check to see if it exists
   first and is NC_CHAR, and if so, delete it and recreate it.  Workaround for C
   lib bug discovered in issue #485.
 * support for NETCDF3_64BIT_DATA format supported in netcdf-c 4.4.0.
   Similar to NETCDF3_64BIT (now NETCDF3_64BIT_OFFSET), but includes
   64 bit dimensions and sizes, plus unsigned and 64 bit integer
   data types.
 * make sure chunksize does not exceed dimension size
   (for non-unlimited dimensions) on variable creation (issue #480).
 * add 'size' attribute to Dimension (same as len(d), where
   d is a Dimension instance, issue #477).
 * fix bug in nc3tonc4 with --unpackshort=1 (issue #474).
 * dates do not have to be contiguous, i.e. can be before and after the
   missing dates in Gregorian calendar (pull request #476).

 version 1.2.1 (tag v1.2.1rel)
 =============================
 * add the capability to slice variables with unsorted integer sequences,
   or integer sequences with duplicates (issue #467). This was done
   by converting boolean array slices to integer array slices internally,
   instead of the other way around.
 * raise TypeError if masked array assigned to a VLEN str variable slice
   (issue #464).
 * Ellipsis now can be used with scalar VLEN str variables (issue #458).
   Slicing of scalar VLEN (non-str) variables now works.
 * Allow non-positive reference years in non-real-world calendars
   (issue #442).

 version 1.2.0 (tag v1.2.0rel)
 =============================
 * Fixes to setup.py for building on windows (issue #460).
 * warnings now issued if file being read contains unsupported
   variables or data types (they were previously being silently
   skipped).
 * added 'get_variables_by_attributes' method (issue #454).
 * check for 'units' attribute in date2index (issue #453).
 * added support for enum types (issue #452).
 * added 'isopen' Dataset method (issue #450).
 * raise ValueError if year 0 or negative year used in time units string.
   The year 0 does not exist in the Julian and Gregorian
   calendars (issue #442).

 version 1.1.9 (tag v1.1.9rel)
 =============================
 * fix for issue #391 (data is already byte-swapped to native
   endian format by the HDF4 library).
 * fix for issue #415 (copy.deepcopy does not work on
   netcdftime datetime object).
 * fix for issue #420 - len(v) where v is a scalar variable returned
   unexpected IndexError, now returns "TypeError: len() on unsized object"
   (same as numpy does for len() on a scalar array).
 * translate docstrings from epydoc markup to markdown, so
   pdoc can be used (epydoc is dead).
 * add small offset in conversion to Julian date for numerical
   stability (more accurate round trip calculations). This offset is removed
   in back conversion only from microseconds. Pull request #433.
 * add detection of unsigned integers to handling of automatic
   packing (set_auto_scale and set_auto_maskandscale) when writing.
   Pull request #435.
 * use USE_SETUPCFG env var to over-ride use of setup.cfg. If USE_SETUPCFG
   evaluates to false, setup.cfg will not be used and all configuration
   variables can be set from environment variables.  Useful when using 'pip
   install' and nc-config is broken (issue #438).
 * fix for integer overflow in date2index (issue #444).

 version 1.1.8 (tag v1.1.8rel)
 =============================
 * v[...] now returns a numpy scalar array (not just a scalar) when
   v is a scalar netcdf variable (issue #413).
 * unix-like paths can now be used in createVariable and createGroup.
   v = nc.createVariable('/path/to/var1',('xdim','ydim'),float)
   will create a Variable named 'var1', while also creating the Groups
   'path' and 'path/to' if they do not already exist.
   Similarly, g = nc.createGroup('/path/to') acts like 'mkdir -p' in unix, creating
   the Groups 'path' and '/path/to', if they don't already exist.
   Users who relied on nc.createGroup(groupname) failing when the
   group already exists will have to modify their code, since nc.createGroup
   will now return the existing group instance.
   Dataset.__getitem__ also added.  nc['/path/to'] returns
   a Group instance, and nc['/path/to/var1'] returns a Variable
   instance.
 * change minimum required numpy to 1.7.0, fix so all tests pass with 1.7.0.
   Added travis tests for minimum required cython, numpy (issue #404).
 * enable abbreviations to time units specification, as allowed in CF (issue
   #402). Now, instead of just 'seconds' and 'seconds', 'secs', 'sec' and 's'
   are also allowed (similar to minutes, days and hours).
 * install utility scripts in utils directory with setuptools entry points
   (pull request #392 from @mindw).  Code for utilities moved
   to netCDF4_utils.py - makes utilities more windows-friendly.
 * make sure booleans are treated correctly in setup.cfg. Add
   use_cython (default True) to setup.cfg.  If set to False, then
   cython will not be used to compile netCDF4.pyx (existing netCDF4.c
   will be used instead).
 * use "from Cython.Build import cythonize" instead of
   "from Cython.Distutils import build_ext" in setup.py (issue #393)
   to conform to new cython build mechanism (CEP 201, described at
   https://github.com/cython/cython/wiki/enhancements-distutils_preprocessing).
 * unicode attributes now written as strings, not bytes (using
   nc_put_att_string instead of nc_put_att_text, issue #388).
 * add __orthogonal_indexing__ attribute to Variable, Dataset and Group (issue #385) to
   denote that Variable objects do not follow numpy indexing semantics for integer and
   boolean array indices.
 * make sure application of scale_factor and add_offset works correctly when
   scale_factor not given (issue #381).
 * add man pages for nc3tonc4, nc4tonc3, ncinfo in man directory.
   Not installed by setup.py (contributed by Ross Gammon, issue #383).
 * replace tabs with spaces by running reindent.py on all *.py and *.pyx files
   (issue #378).
 * refactor netCDF4_utils and netCDF4 module into netCDF4 package.
   Refactoring effectively removes netCDF4 utils private attributes from
   netCDF4 namespace, so has the potential to break code using private
   attributes (issue #409).

 version 1.1.7 (tag v1.1.7rel)
 =============================
 * check to make sure cython >= 0.19 is available before trying
   to use it (otherwise compilation with fail).  Issue 367.
 * add ipython notebooks from Unidata workshop in examples directory.
 * fix ellipsis variable slicing regression (issue 371).
 * release the Global Interpreter Lock (GIL) when calling the C
   library for read operations.  Speeds up multi-threaded reads
   (issue 369).  Caution - the HDF5 library may need to be compiled
   with the threadsafe option to ensure that global data structures
   are not corrupted by simultaneous manipulation by different threads.
 * Make sure USE_NCCONFIG environment variable takes precedence over value
   of use_ncconfig in setup.cfg.  With this change, 'pip install netCDF4'
   with USE_NCCONFIG=1 will use environment variables to find paths to
   libraries and include files, instead of relying on nc-config (issue #341).

 version 1.1.6 (tag v1.1.6rel)
 =============================

 * fix for issue 353 (num2date can no longer handle units like 'hours since
   2000-01-01 0').
 * fix for issue 354 (num2date no longer supports multi-dimensional arrays).
 * fix for spurious UserWarning about endian-ness mismatch (issue 364).
 * make calendar name keyword for num2date/date2num case insensitive
   (issue 362).
 * make sure units parser returns time-zone naive datetime instance that
   includes UTC offset (issue 357).  UTC offset was applied incorrectly in
   netcdftime.date2num and num2date.  No longer need to depend on
   python-dateutil.

 version 1.1.5 (tag v1.1.5rel)
 =============================

 * add dependency on python-dateutil in setup.py and install docs.
 * use python datetime in num2date and date2num whenever possible. Remove
   duplicate num2date and date2num functions from netcdftime. Addresses issue
   #344. Add microsecond capability to netcdftime.datetime. Roundtrip
   accuracy of num2date/date2num now down to less than a millisecond.
 * use nc-config by default to find dependencies.  setup.py modified to handle failure
   to find nc-config more gracefully (issue #340). If you wish to use env vars to point
   to the libs, you must first move the setup.cfg file out of the way
   (rename it to setup.cfg.save), or set USE_NCCONFIG to 0.
 * if endian-ness of variable is specified, adjust datatype to reflect this when opening
   a file (issue 346).
 * fix for issue #349 (seconds outside the range 0-59 in netcdftime.num2date).

 version 1.1.4 (tag v1.1.4rel)
 =============================

 * speedup conversion of array indices to slices (issue #325).
 * fix for issue #330 (incorrect values for seconds returned by netcdftime).
 * fix reading of scalar vlen variables (issue #333).
 * setting fill_value=False in createVariable for vlen and compound variables
   now does nothing, instead of causing an error when the Dataset is closed
   (issue #331).
 * cython will regenerate netCDF4.c when install is run, not just build.
   Makes 'pip install' do the right thing when cython is installed (issue #263).

 version 1.1.3 (tag v1.1.3rel)
 =============================

 * checked in _datetime.c to git (resolves issue #315).  Note - _datetime.c
   was *not* included in the 1.1.2 release.
 * Changed __str__ to __repr__ in MFDataset, to be consistent with Dataset
   (issue #317). IPython uses __repr__ to make use-friendly human-readable summaries
   of objects in the terminal.

 version 1.1.2 (tag v1.1.2rel)
 =============================

 * fix for issue 312 (allow slicing with objects that can be cast to ints).
 * indexing netCDF variables with integer sequences and boolean arrays now
   behave the same way (integer sequences are converted to boolean arrays
   internally). Addresses issue #300.  Since indexing using integer sequences
   does not behave exactly as before, some client code may break.  For example,
   previously when integer index arrays had the same length, and that length
   was equal to the number of dimensions of the array being indexed,
   netcdf4-python mirrored the numpy indexing behavior and treated the elements
   of the index arrays as individual sets of integer indices.  This special
   case has been removed. An IndexError is now raised when the new behavior
   would produce a different result than the old, i.e. when the
   indices in an integer sequence are not sorted, or there are duplicate
   indices in the sequence.
 * fix for issue #310 (masked arrays not returned correctly when variable
   has non native endian-ness).
 * fix for issue #306 (slicing variable with "-1" when there is only
   one element along that dimension).
 * Improved speed of num2date and date2num for standard, julian, gregorian
   and proleptic gregorian calendars by vectorizing the functions. See Issue #296
 * Fix for issue #301 ("Datestring parser chokes on years with extra space").
 * Add name property for Dimension, Variable and Group instances (to access string
   name associated with instance).
 * Allow for null byte attributes (so _FillValue='\x00' can be set manually).
   Issue 273.
 * Added __repr__ (matching __str__) for all types (pull request #291).
   IPython uses __repr__ to make use-friendly human-readable summaries
   of objects in the terminal.


 version 1.1.1 (tag v1.1.1rel)
 ==============================

 * make sure _FillValue is a byte for character arrays in Python 3 (issue
   271).
 * add numpy to install_requires in setup.py (issue #282, fixes issue #211).
   'pip install netcdf4-python' will no longer fail if numpy not installed.
 * Fix for issue 278 (UnicodeDecodeError reading netcdf.h from setup.py with
   Python 3.4).
 * Make netcdftime.datetime immutable and hashable (issue 255).
 * Fix issue with slicing of scalar VLEN arrays (issue 270).
 * Add set_auto_mask and set_auto_scale methods to control auto scaling and
   auto masking separately. (issue 269).  Also added set_auto_maskandscale,
   set_auto_scale, set_auto_mask Dataset/Group methods that recursively walk
   through all variables in the Dataset/Group.
 * Make sure file_format attribute propagated to Group instances (issue 265).
 * Fix for issue #259 ("Cannot use broadcasting to set all elements of a
   Variable to a given value").

 version 1.1.0 (tag v1.1.0rel)
 =============================

 * revert weakref change, so that previous behaviour (Dimensions and Variables
   keep strong references to parent Dataset) is the default.  New keyword
   argument 'keepweakref' for Dataset.__init__ can be set to true to get
   weak references.

 version 1.0.9 (tag v1.0.9rel)
 =============================

 * speed up the creation of new Group instances (issue 239).

 * fix logic errors in setup.py (issue 236).

 * it is now possible to create and set variable length string variables with
   numpy string datatypes (pull request 224).

 * add .travis.yml (for travis-ci testing on github), silence warnings from
   test output (issue 225).

 * modify __unicode__ for Variable and Dimension to return more useful error
   message when Dataset object has been garbage collected.

 * use weak references to group instances when creating Dimension and Variable
   objects. This prevents cyclic references messing up garbage collection (issue
   218, pull request 219).

 * accessing values from a 0-dimensional Variable now returns a 0-dimensional
   numpy array, not a 1-dimensional array (issue 220). To write code
   compatible with both the old and new (fixed) behavior, wrap values accessed
   from a 0-dimensional Variable with numpy.asscalar.

 * add an __array__ method to Variable to make numpy ufuncs faster (issue 216).

 * change download_url in setup.py to point to pypi instead of googlecode.

 * fix for date2index error when time variable has only one entry (issue 215).

 * silence warnings ("Non-trivial type declarators in shared declaration (e.g.
   mix of pointers and values). Each pointer declaration should be on its own
   line") with Cython 0.2.

 * reduced memory usage for Variable.__getitem__ under Python 2.

 version 1.0.8 (tag v1.0.8rel)
 =============================

 * change file_format Dataset attribute to data_model (keeping file_format
   for backward compatibility).  Add disk_format attribute (underlying
   disk format, one of NETCDF3, HDF4, HDF5, DAP2, DAP4, PNETCDF or UNDEFINED).
   Uses nc_inq_format_extended, added in version 4.3.1 of library.  If using
   earlier version of lib, disk_format will be set to UNDEFINED.

 * default _FillValue now ignored for byte data types (int8 and uint8) as per
   http://www.unidata.ucar.edu/software/netcdf/docs/netcdf-c/Fill-Values.html#Fill-Values
   "If you need a fill value for a byte variable, it is recommended that you
   explicitly define an appropriate _FillValue attribute, as generic utilities
   such as ncdump will not assume a default fill value for byte variables".
   ncinfo now returns fill mode information (issue 209).

 * check to see if filling was disabled before masking data equal to
   default fill value (issue 209).

 * add variable type information to Dataset.__repr__ (output of ncinfo).

 version 1.0.7 (tag v1.0.7rel)
 =============================

 * add the ability to specify the locations of hdf4,jpeg and curl libs, in
   case netCDF4 was built statically with HDF4 and/or OpenDAP support (issue
   207).

 * add 'ncinfo' utility (like 'ncdump -h' but less verbose).

 * more information displayed when Dataset or Group instance is printed.

 * fix for issue 194 (versions after 1.0.5 fail for netcdf 4.1.1, due to
   call to nc_inq_path, which was added in netcdf 4.1.2).  Fixed by adding
   compile time API check similar to what was done for nc_rename_grp.  If
   filepath Dataset method is called an exception will be raised at runtime
   if the module was built with netcdf < 4.1.2, or cython was not installed
   at build time.

 * fix for issues 202 and 206 (exception raised by numpy.isnan for character data types).

 * if dateutils not installed and time unit accuracy < 1 second requested, have
   netcdftime raise an ImportError.

 version 1.0.6 (svn revision 1312)
 ================================

 * issue warning of endian-ness of dtype argument does not match endian kwarg in
   createVariable.

 * make sure netcdf type NC_CHAR always returned in numpy array dtype 'S1'
   (sometimes arrays of type 'U1' were being returned). Fixes intermittently failing
   test tst_compoundatt.py on python 3.3.

 * fix for issue 201 (if data associated with numpy array not the
   same endian-ness as dtype, data was written incorrectly).  Now
   bytes are swapped if necessary.  Variable.endian() now returns
   'native' instead of None for NETCDF3 formatted files.  createVariable
   now enforces endian='native' for NETCDF3 files. Added tst_endian.py
   test case.

 * fix for issue 200 (library version detection failed on cygwin).

 * fix for issue 199 (nc4tonc3 utility not copying global attributes).

 * fix for issue 198 (setup.py chokes when no arguments given).

 * fix for issue 197 (slicing of netCDF variables using lists of integers).

 * create 'path' attribute for group instance using posixpath, instead
   of os.path (to ensure the unix path is used on all platforms). Issue 196.

 * fix for issue 196 (test failures on win32 due to files being deleted
   before they are closed).

 version 1.0.5 (svn revision 1278)
 ================================

 * change setup.py to compile the Cython sources directly, if
   cython is available.
   This allows for "ifdef" like capability to modify source at compile time to
   account for changes in netcdf API (e.g. the forthcoming addition of the
   nc_rename_grp in version 4.3.1).

 * added a "renameGroup" method, which raises an exception if
   the netcdf lib version linked does not support it. Requires
   netcdf >= 4.3.1.

 * support for more than one missing value (missing_value attribute
   is a vector) when converting to masked array.

 * add 'renameAttribute' method to Dataset, Group and Variable.

 * fix so that var[:] = x works if x is a scalar, and var is a netcdf variable
   with an unlimited dimension that has shape () - i.e. no data has been
   written to it yet.  Before this change, var[:] = x did not write any data.
   Now the scalar x will be written as the first entry in var along the
   unlimited dimension.

 * remove dos line feeds from nc3tonc4 (issue 181).

 * add datatype property for Variable that returns numpy dtype for primitive
   datatypes (same as dtype attribute) but returns CompoundType or VLType
   instance for compound or vlen variables (issue 178).

 * fix logic for deciding where to look for nc-config in setup.py (issue 177).

 * issue a warning and don't try to apply scale_factor or add_offset if these
   attributes are not convertible to floats (issue 176).

 * add filepath method to Dataset instance to return file path (or opendap
   URL) used to create Dataset (issue 172).

 * fix for issue 170 (opening a remote DAP dataset fails after creating
   a NETCDF4 formatted file).

 * fix for issue 169 (error in chartostring function on 64-bit windows).

 * add support for missing_value or _FillValue == NaN (issue 168).

 * added a Dimension.group() method (issue 165).

 version 1.0.4 (svn revision 1229)
 =================================

 * fixed alignment bug that could cause memory corruption
   when reading compound type variables. All users of compound
   types should upgrade.

 version 1.0.3 (svn revision 1219)
 =================================

 * don't try to write empty data array to netcdf file (fixed failing
   test with netcdf 4.3.0rc2).

 * date2num, num2date and date2index now can handle units of microseconds and milliseconds
   (for proleptic_gregorian calendar, or gregorian and standard calendars
    as long as the time origin is after 1582-10-15). Issue 159.

 * Added a _grp attribute to Dimension (issue 165).

 * don't bundle ordereddict (issue 164).

 * support reading of vlen string attributes (issue 156).

 * add --vars option to nc3tonc4 (issue 154).

 * Don't try to set fletcher32 checksum on scalar variables (it causes HDF5
   to crash).  Fixes issue 150.

 * Add --istart/--istop options to nc3tonc4 (issue 148, courtesy of
   Rich Signell).

 * fix for proleptic_gregorian in netcdftime.py (courtesy of Matthias Cuntz).

 version 1.0.2 (svn revision 1196)
 =================================

 * disable version check for HDF5, which is broken by hdf5 1.8.10.

 * make sure all files have a calendar attribute in MFTime (issue 144).

 * more robust fix to issue 90 (array shape modified by assignment to
   a netCDF variable with one more dimension), including test case.

 version 1.0.1 (svn revision 1190)
 =================================

 * fix error that occurred when retrieving data from a variable that has a
   missing_value attribute specified as a string (issue 142).

 * automatically close netcdf files when there are no references
   left to Dataset object (using __dealloc__ method).  Fixes issue 137.

 * fix for slicing of scalar vlen string variables (issue 140).

 * fix to allow writing of unicode data to a NC_CHAR variable.

 * allow for writing of large variables (> 2**32 elements). Fixes issue 130.

 version 1.0fix1
 ===============

 * fix python 3 incompatibility in setup.py (issue 125).

 version 1.0 (svn revision 1164)
 ===============================
 * add 'aggdim' keyword to MFDataset, so the name of the dimension
   to aggregate over can be specified (instead of using the unlimited
   dimension).  aggdim=None by default, which results in the previous
   behavior.  aggdim must be the leftmost dimension of all the variables
   to be aggregated.

 * raise IndexError when indexing a netcdf variable out of range
   so iterating over a variable in a for loop behaves as expected
   (as described in http://effbot.org/zone/python-for-statement.htm).
   Fixes issue 121.

 * added MacPorts portfile (so it can be installed via MacPorts
   on macosx using a "local Portfile repository"). Installs
   from svn HEAD using 'port install netcdf4-python'.

 * added experimental 'diskless' file capability (only added to the C
   lib after the 4.2 release).  Controlled by kwarg 'diskless' to
   netCDF4.Dataset (default False). diskless=True when creating a file
   results in a file that exists only in memory, closing the file
   makes the data disappear, except if persist=True keyword given in
   which case it is persisted to a disk file on close.  diskless=True
   when opening a file creates an in-memory copy of the file for faster access.

 * add the ability to specify the location of the required libs
   (and whether to use nc-config) with setup.cfg,
   instead of using environment variables.

 * fix ISO9601 date parser so it recognizes time zone offsets in time unit
   strings (contributed by David Hassel, issue 114, r1117).

 * add setncatts Dataset,Group and Variable method to add a bunch of
   attributes (given in a python dictionary) at once.  Speeds things
   up for NETCDF3 and NETCDF4_CLASSIC files a lot, since nc_redef/nc_enddef
   not need to be called for each attribute (issue 85, r1113). Adding 1000
   attributes is about 35 times faster using setncatts to add them all at once.
   Makes no difference for NETCDF4 formatted files, since nc_redef/nc_enddef
   is not called.

 * only round after apply scale_factor and add_offset if variable type is
   integer (issue 111, r1109).

 * Fixed bug with all False Boolean index (r1107).

 * added support for after, before and nearest selection method to date2index
   fast "first guess" indexing (r1106).

 * Remove white space in time units string (netcdftime._parse_date).
   An extra space in the time units of one CMIP3 model caused an error
   (r1105).

 * based on results with examples/bench_compress2.py, change
   default complevel for zlib compression from 6 to 4. If complevel=0,
   turn compression off entirely (set zlib=False) (r1102).

 version 0.9.9 (svn revision 1099)
 ================================
 * changed default unicode encoding from "latin-1" to "utf-8", since
   this is the python 3 default, and the only encoding that appears to
   work for dimension and variable names.

 * added test case for unicode attributes, variable and dimension names.

 * fixes for unicode variable, dimension and group names.

 * fix for unicode attributes in python3 (ncdump did not intrepret them as
   text strings). Issue 107.

 * add --format option to nc4tonc3 utility (can be either NETCDF3_CLASSIC or
   NETCDF3_64BIT). Fixes issue 104.

 version 0.9.8 (svn revision 1080)
 ================================
 * use numpy.ma.isMA to check for masked array (instead of checking
   for presence of 'mask' attribute).

 * fixes for AIX with ibm xlc compiler.

 * make sure unicode attributes don't get converted to ascii strings (issue
   98).

 version 0.9.7 (svn revision 1073)
 ================================
 * Added __str__ methods to Dataset, Variable, Dimension, CompoundType,
   VLType and MFDataset, so useful human-readable information is provided when these
   objects are printed in an interactive session.

 * don't try to apply scale_factor and offset if scale_factor=1 and
   add_offset=0 (to avoid making copies of large arrays).

 * changed netCDF4._default_fillvals to netCDF4.default_fillvals (to make part
   of public API).  Added to docs (issue 94).

 version 0.9.6 (svn revision 1043)
 =================================
 * changed default unicode encoding from "ascii" to "latin-1" (iso-8859-1).

 * add "unicode_error" module variable to control what happens when characters
   cannot be decoded by the encoding specified by the "default_encoding" module
   variable (which is "ascii" by default).  unicode_error = "replace" by default
   which means bad characters are replace by "?".  Previously an error was
   raised, the old behavior can be obtained by setting unicode_error = 'strict'.
   Fixes issue 92.

 * add __enter__ and __exit__ methods so you can do "with Dataset(url) as f:"
   (issue 89).

 * don't add extra singleton dimensions to rhs numpy arrays when assigning to
   a netcdf variable. Fixes issue 90.

 * coerce missing_value attribute to same type as variable (for primitive
   types). Fixes issue 91.

 version 0.9.5 (svn revision 1031)
 ================================
 * fix for compound variables on python 3.2.

 * fix slicing of masked MFDataset variables (issue 83).

 * round to nearest integer after packing with scale_factor and add_offset
   (instead of truncation) (issue 84).

 * if add_offset missing, but scale_factor present, assume add_offset zero.
   if scale_factor missing, but add_offset present, assume scale_factor one.
   (this is consistent with unidata recommendations - issue 86).

 * only try to convert strings to bytes for python 3 so Dataset can
   be subclassed (issue 87).

 version 0.9.4 (svn revision 1018)
 ================================
 * tested with python 2.7.1/3.1.3 using netcdf 4.1.2 and hdf5 1.8.6.

 * Added a 'default_encoding' module variable that controls how unicode
 strings are encoded into bytes. Default is 'ascii'.

 * now works on Python 3.

 * netCDF3 module removed. If you still need it, get it from netCDF4 0.9.3.

 * regenerated C source with Cython 0.14.1.

 * Added a MFTime class. Provide a unified interface to MFDataset time
 variable using different time units.

 * Fixed bug in netcdftime (issue 75) that occurs when time specified is within
 one second of the end of the month.

 * on unix-like systems, the environment variable USE_NCCONFIG can be set to
 tell setup.py to use the nc-config script installed by netcdf to figure out
 where all the libs and headers are (without having to specify NETCDF_DIR,
 HDF5_DIR, etc).  Only works with netcdf 4.1.2.

 version 0.9.3 (svn revision 930)
 ================================
 * fix chunk sizes bug (chunk sizes pointer should be size_t, not int).  Fixes
 issue 66. Added test in tst_compression.py

 * fixed writing of data with missing values with scale/offset packing. Added
 test (tst_masked2.py).

 * fix iso8601 regex in netcdftime date parser so it can parse 'hours since 1-1-1 ...'
 (year had to be 4 digits previously)

 version 0.9.2 (svn revision 907)
 ================================
 * fix netcdftime bug with '360_day' calendar. Fixes issue 59.

 * make sure scalar slice of 1d variable returns array scalar (not array of
   shape (1,)). Fixes issue 57.

 * updated date parser in netcdftime.  Can now handle units like
   "seconds since 1970-01-01T00:00:00Z".

 * added support in setup.py for specifying the locations of the HDF5/netcdf-4 headers
   and libs separately with environment variables (HDF5_INCDIR, HDF5_LIBDIR).i
   Patch contributed by Patrice Dumas.

 * add masked array support to num2date (dates for missing times set to None).

 * add chunk_cache keyword to createVariable.  HDF5 default is 1mb, which can
   cause problems when creating 1000's of variables.  In such cases,
   chunk_cache can be reduced, or set to zero.

 * add set_var_chunk_cache and get_var_chunk_cache Variable methods.

 * raise AttributeError when trying to set _FillValue attribute (it can only
   be reliably set on variable creation, using the fill_value keyword to
   createVariable).

version 0.9.1 (svn revision 879)
================================
 * raise ImportError if netcdf-4 < 4.1.1 or hdf5 <= 1.8.4.

 * add __netcdf4libversion__ and __hdf5libversion__ module variables.

 * make sure data is not truncated to integers before scale_factor and
   add_offset is applied (issue 46).

 * fix bug in date2num with noleap calendar in netcdftime (issue 45).

 * fix bug in 360day calendar in netcdftime (issue 44).

 * python 2.4 compatibility restored (by modifying OrderedDict).  Fixes issue
   37.

 * make sure int64 attributes cast to int32 when format=NETCDF4_CLASSIC. This
   was causing tst_multifile.py to fail on 64-bit platforms.

 * fix tutorial.py to cast 64 bit integers to 32 bit when writing to 32-bit
   integer vlen (was causing tutorial.py to fail on 64-bit platforms).

 * remove nose dependency from tst_netcdftime.py.

version 0.9 (svn revision 846)
==============================
 * fixed bug (issue 30) with date2index occurring with dates outside the support.

 * make sure that auto masking works with MFDataset.

 * fix bug (issue 34) when slicing MFDataset variables with dimensions of
   length 1.

 * used ordered dictionaries for variables, dimensions, groups etc to
   preserve creation order (makes it easier to copy files, fixes issue 28).

 * change auto_maskandscale default to True.  This means data will
   automatically be converted to and from masked arrays.  Data scaled as short
   integers using the scale_factor and add_offset attributes will also be
   automatically converted to/from float arrays.

 * add setncattr, getncattr, delncattr methods (for setting/getting/deleting netcdf
   attributes with names that clash with the reserved python attributes).

version 0.8.2 (svn revision 769)
================================
 * compound type tests re-enabled. Compound and vlen types now fully
   supported in netcdf-4.1-beta2.

 * make sure data retrieved from a netCDF variable is not coerced to a python
   scalar (it should remain a numpy scalar array).

 * fix docs to point out that an unlimited dimension can be created by setting
   size to *either* None or 0 in createDimension.

 * fix another slicing corner case.

 * remove auto pickling/unpickling into vlen strings (too cute, sometimes
   produced surprising results).

version 0.8.1 (svn revision 744)
================================
 * added 'cmptypes' and 'vltypes' Group/Dataset attributes, which contain
   dictionaries that map the names of compound and vlen types to
   CompoundType and VLType instances.

 * Experimental variable-length (vlen) data type support added.

 * changes to accommodate compound types in netcdf-4.1-beta snapshots.
   Compound types now work correctly for snapshots >= 20090603.

 * Added __len__ method and 'size' property to Variable class.

 * In date2index, replaced the brute force method by the bisection method and
   added a 'select' keyword to find the index of the date before, after or
   nearest the given date.

 * Fixed bug occurring when indexing with a numpy array of length 1.

 * Fixed bug that occurred when -1 was used as a variable index.

 * enabled 'shared access' mode for NETCDF3 formatted files (mode='ws',
   'r+s' or 'as'). Writes in shared mode are unbuffered, which can
   improve performance for non-sequential access.

 * fixed bug in renameVariable that caused failure when new name is longer
   than old name, and file format is NETCDF3_64BIT or NETCDF3_CLASSIC.

version 0.8 (svn revision 685)
==============================
 * added 'stringtoarr' utility function for converting python strings
   to numpy character arrays of a specified size.

 * initial support for compound data types (which are mapped to structured
   numpy arrays). Compound data types are created with the createCompoundTYpe
   Dataset or Group method. Both attributes and variables can be compound
   types.

 * make sure 64-bit integer attributes converted to 32 bits when writing to
   a NETCDF3 formatted file.

 * added nc4tonc3 utility for converted NETCDF4_CLASSIC files to NETCDF3_64BIT
   files (useful for sharing data with colleagues that don't have netcdf-4
   capable clients).

version 0.7.7 (svn revision 626)
================================

 * David Huard reworked fancy indexing - it is now much more efficient and
   less of a memory hog. Now works differently than numpy fancy indexing - 1d
   arrays of boolean or integer indices work independently on each dimension.
   This enables things like:
   >>> tempdat = temp[[0,1,3],lats>0,lons>0]
   (retrieves 1st, 2nd and 4th levels, all Northern Hem. and Eastern
    Hem. grid points - note that this would raise an IndexError in numpy)

* added opendap test (tst_dap.py).

* bugfix for nc3tonc4 utility.

* fix MFDataset.Variable. __getattr__ to raise AttributeError instead of
  KeyError when attribute not found.

* netcdftime version number upped to 0.7.

version 0.7.6 (svn revision 574)
================================

* added date2index function, courtesy of David Huard, which finds the indices
  in a netCDF time variable corresponding to a sequence of datetime instances.

* make _get_att/_set_att raise AttributeError instead of RuntimeError, so that
  getattr(object, 'nonexistantattribute', None) works.  (thanks David Huard)

* v[:] = data now works along unlim dim, i.e. you can do this:

  file = Dataset('test.nc', "w")
  file.createDimension("time", None)     # unlimited dimension
  var = file.createVariable("var", 'd', ("time",))
  # you used to have to do this
  #var[0:10] = numpy.arange(10)
  # but now you can simply do this
  var[:] = numpy.arange(10)

version 0.7.5 (svn revision 549)
================================
* return a scalar array, not a python scalar, when a slice returns a single
  number. This is more consistent with numpy behavior, and fixes a bug
  in MFDataset slicing.
* added 'exclude' parameter to MFDataset.__init__
* added set_auto_maskandscale method to MFDataset variables.

version 0.7.4 (svn revision 540)
================================
* ensure all arithmetic is done with float64 in netcdftime (Rob Hetland).
* fixes for netcdf-4.0-beta2 ('chunking' keyword to createVariable
  replaced by 'contiguous').  Now works with netcdf-4.0-beta2 and hdf5-1.8.0
  final, but is incompatible with netcdf-4.0-beta1.

version 0.7.3.1 (svn revision 507)
==================================
* netCDF3 docs were missing from 0.7.3.
* make sure quantization function preserves fill_value of masked arrays.

version 0.7.3 (svn revision 501)
================================
* MFnetCDF4 module merged into netCDF4 and netCDF3 (now
  called MFDataset).
* added netCDF3 module for those who can't install the netCDF 4 lib.
* added set_auto_maskandscale Variable method to enable automatic
  packing and unpacking of short integers (using scale_factor
  and add_offset attributes) and automatic conversion to/from
  masked arrays (using missing_value or _FillValue attribute)
  on a per-variable basis.
  var.set_auto_maskandscale(True) turns automatic
  conversion on (it is off by default).
* automatically pack/unpack short integer variables
  if scale_factor and add_offset variable attributes are set.
* added support for masked arrays.  If you try to write a masked
  array to a variable with the missing_value or _FillValue attributes
  set, the masked array is filled with that value before being written
  to the file.  If you read data from a variable with the missing_value
  or _FillValue attribute set, a masked array is returned with the
  appropriate values masked.
* added date2num and num2date functions.
* added capability to use 'fancy indexing' with variable objects
  (i.e. using sequences of integers or booleans in slices). WARNING:
  if a sequence of integers or booleans is used to slice a netCDF4
  variable, all of the data in that dimension is read into a numpy
  array, and then the sequence is used to slice the numpy array,
  returning just the requested elements to the user.  This can
  potentially gobble a lot of memory and degrade performance
  (especially if 'fancy indexing' is done on the left-most dimension).
* added convenience functions stringtochar and chartostring for
  converting character arrays to arrays of fixed-length strings and
  vice-versa.   Example usage in examples/test_stringarr.py.

20070826 - version 0.7.1 (svn revision 400)
===========================================
* added 'endian()' and 'chunking()' Variable methods (to inquire about
  endian and chunking variable settings).

* 'ndim' attribute was not public (so it couldn't be accessed from python).
  Fixed.

* added 'endian' kwarg to createVariable (to set the endian-ness
  used in the HDF5 file).

* can now manually set HDF5 chunksizes for each dimension at
  variable creation, using 'chunksizes' kwarg to createVariable.

* added "getlibversion()" function to get info about version
  of netcdf-4 library used to build module.

* if a variable has an unsupported datatype (such as 'compound', or
  'vlen'), then instead of raising an exception, just skip it.
  Print a useful error message when an attribute with an unsupported
  datatype is accessed.

* if variable dimension is specified as 'dimname' or ('dimname')
  in createVariable, it is automatically converted to a tuple ('dimname',).
  Better error messages when specified dimension can't be found.

* createVariable accepts numpy dtype object as datatype.  dtype variable
  attribute is now a numpy dtype object.

20070723 - version 0.7 (svn revision 361)
=========================================
* renamed MFnetCDF4_classic --> MFnetCDF4.

* eliminated netCDF4_classic module (all file formats handled by
  netCDF4 module now).

* removed all user-defined data type stuff (it was hacky and made
  the code too complex - wait till there is a real use case to
  refactor and put back in).

* added 'ndim' variable attribute (number of variable dimensions).

20070424 - version 0.6.3 (svn revision 302)
===========================================
* passes all tests with netcdf-4.0-beta1/hdf5-1.8.0-beta1.

* if slice index is not a slice object, assume it's an integer (and
  try to convert to one if it is not).  This allows numpy scalar arrays
  to work as slice indices.

* (netCDF4_classic only) try to make sure file is not left in 'define mode'
  when exception is raised.

* if slicing a variable results in a array with shape (1,), just return
  a scalar (except for compound types).

* added instructions for using the netCDF4_classic module to serve
  data over http with the DAP using pydap (http://pydap.org).

* added --quiet and --chunk options to nc3tonc4.

* Turned off zlib compression by default so as not to violate the
  'principle of least surprise'.  Shuffle filter still activated
  by default when zlib compression turned on.

* Fixed bug in fletcher32 checksum activation call.  Renamed compression()
  variable method to filters(), include fletcher32 checksum flag in output.

* added utility for converting GRIB1 files to compressed
  NETCDF4_CLASSIC files (requires PyNIO).

* added 'compression()' variable method that returns a dict with
  compression filter parameter settings for that variable. (rev 237)

* reimplemented 'shape' and 'dimensions' variable attributes as
  properties.

* fixed bug when 'chunking' keyword in createVariable was set to 'sub'
  (caused Bus Error on MacOS X).

* Setting 'shuffle=0' keyword in createVariable was turning off
  zlib compression filter instead of shuffle filter.  Fixed.

20070213 - version 0.6.2
========================
* updated for compatibility with netcdf-4.0-alpha18 and hdf5 1.8.0alpha5
  (shared dimensions actually work now).

* netCDF4.createVariable can now use old single character Numeric typecodes
  for datatype specification.

* Improvements to MFDataset (now called MFnetCDF4_classic) by Rob Hetland.

20061121 - version 0.6.1
========================
* bugfixes for negative strides.

* bugfix for empty string attributes.

* support for shared dimensions (variables can use dimensions defined
  only in a parent group).  This doesn't actually work yet, because of
  a bug in netcdf-4.0-alpha17.

* now requires Pyrex (C source files generated on the fly when setup.py
  is run).

20061003 - version 0.6
======================
* if fill_value keyword to createVariable is set to the Boolean
  False (not an integer that evaluates to False), no pre-filling
  is done for that variable.

* updated to be compatible with netcdf-4.0-alpha17.
  Can now install pure-python netcdftime separately with setup-netcdftime.py.
  netcdftime will try to use numpy, but fall back to Numeric if numpy
  not installed.

* generated source files with a version of pyrex
  (from http://codespeak.net/svn/lxml/pyrex/) that produces
  extensions compatible with python 2.5.

* added new module for multi-file access of NETCDF3 and NETCDF4_CLASSIC
  files (MFDataset). Based on CDFMF from pycdf.

* implement negative strides in variable slicing
  (feature missing from Scientific.IO.NetCDF). Now variables support
  full python extended slicing syntax.

20060925 - version 0.5.1
========================
* on 64-bit systems integer attributes in netCDF4_classic failed, since there
  is no 64-bit integer data type. Fixed by downcasting to 32-bit integer.

20060920 - version 0.5
======================
* Compound type support! (members must be fixed data primitive types -
  no user-defined types or NC_STRING variables allowed).  Attributes
  are still restricted to primitive data types (no vlen or compound
  type attributes).

* Assigning single values to a slice now does the Right Thing, i.e.
  >>> data[:] = 1
  fills all the elements with 1 (instead of raising an IndexError).

* Tested with numpy 1.0b5, netcdf-4.0-alpha16, HDF5 1.7.52 alpha.

* Added renameDimension and renameVariable methods to Dataset and Group classes.

* netCDF attributes can be deleted using python del (i.e. 'del dset.foo').

* Moved examples from test and test_classic to examples and
  examples_classic directories.

* Added proper unit tests (in test and test_classic directories).

* NULL characters are removed from text attributes.

* Variable _FillValue can be set using new keyword argument 'fill_value'
  to createVariable Dataset and Group method.

* docstrings now formatted with epydoc (http://epydoc.sf.net).

* improved Scientific.IO.NetCDF compatibility for netCDF4_classic
  (typecode method, ability to use old Numeric typecodes).

* zlib=False or complevel=0 disables shuffle filter in createVariable.

* subversion repository hosted on Google projects
  (http://code.google.com/p/netcdf4-python/).

* examples_classic/bench2.py is a performance comparison with
  Scientific.IO.NetCDF (the numpy version provided by pynetcdf).

* __dict__ attribute of Dataset, Group or Variable provides a python
  dictionary with all netCDF attribute name/value pairs (just like
  Scientific.IO.NetCDF).

20060710 - version 0.4.5
========================
* fixed to work with recent svn versions of numpy

* Now requires at least numpy 0.9.8.

* Raise a AttributeError if user tries to rebind a private attribute
  (like 'variables', 'dimensions' or 'dtype').

20060629 - version 0.4.4
========================
* fixed to work with netcdf-4.0-alpha14.

* automatically cast _FillValue attribute to variable type, to
  avoid surprising error message.

20060320 - version 0.4.3
========================
updated netcdftime module yet again
added 'all_leap'/'366_day' and '360_day' calendars.
netCDFTime class renamed utime, fwd and inv methods
renamed date2num and num2date. These methods can now handle
numpy arrays as well as scalars.
a 'real' python datetime instance is returned if calendar
is gregorian, otherwise a 'datetime-like' instance is returned
(python datetime can't handle funky dates in 'all_leap' and '360_day'
calendars).

20060316 - version 0.4.2
========================
udunits module replaced by pure python version, renamed 'netcdftime'
No longer requires udunits library. Includes 4 calendars
('julian','standard'/'gregorian','proleptic_gregorian','noleap'/'365_day').
Calendar names and their interpretations follow the CF metadata convention.

20060310 - version 0.4.1
========================
udunits module included for doing time conversions.

20060306 - version 0.4
======================

netCDF4_classic module can now write NETCDF3_CLASSIC, NETCDF4_64BIT
as well as NETCDF4_CLASSIC files.  The file format is given as
an optional keyword to the Dataset constructor ('NETCDF4_CLASSIC'
is the default).  Preliminary work on compound types done - but
awaiting the next alpha of the netCDF 4 library to complete (bugs
in alpha12 prevent it from working properly if the compound type
has fields which are arrays).

20060217 - version 0.3.1
========================
refactored user-defined data type support - user-defined
data types are now described by an instance of the class
UserType.  usertype and usertype_name keyword args
eliminated from createVariable.

20060214 - version 0.3
======================
support for variable length strengths (typecode = 'S') and
variable-length, or 'ragged' arrays (vlen user-defined datatype).
Arrays of python objects can be saved as pickled strings with
datatype = 'S'.

20050128 - version 0.2.5
========================
added support for scalar variables (and assignValue, getValue
Variable methods for Scientific.IO.NetCDF compatibility).

20051123 - version 0.2.4
========================
numpy 0.9.4 compatibility
Changed data type codes from ('d', 'f', 'i', 'h', ...) to
('f8', 'f4', 'i4', 'i2', ...).

20050110 - version 0.2.3
========================
added ellipsis slicing capability

20050106 - version 0.2.2
========================
changed scipy_core to numpy.

20051228 - version 0.2.1
========================
bugfixes, added 'nc3tonc4' utility to convert netCDF version 3 files
to NETCDF4_CLASSIC files (with compression).  The converted files
can be read from netCDF 3 clients that have been re-linked to the netCDF 4
library. 'chunking' keyword added to createVariable in netCDF4 module.

20051224 - version 0.2
======================
Added netCDF4_classic module - which creates files in NETCDF4_CLASSIC
format.  These files are compatible with netCDF 3 clients which have
been linked against the netCDF 4 lib.  This module does not use any
new features of the netCDF 4 API except zlib compression.  Unlike
any other netCDF 3 python client, it can transparently compress data
with zlib compression and the HDF5 shuffle filter.

20051222 - version 0.1
======================
First release.  Supports groups, multiple unlimited dimensions, zlib
compression (plus shuffle filter and fletcher32 checksum) and all new
primitive data types.  No support for user-defined data types yet.