What’s new in h5py 2.3¶
Support for arbitrary vlen data¶
Variable-length data is no longer restricted to strings. You can use this feature to produce “ragged” arrays, whose members are 1D arrays of variable length.
The implementation of special types was changed to use the NumPy dtype
“metadata” field. This change should be transparent, as access to special types
is handled through
Improved exception messages¶
H5py has historically suffered from low-detail exception messages generated automatically by HDF5. While the exception types in 2.3 remain identical to those in 2.2, the messages have been substantially improved to provide more information as to the source of the error.
ValueError: Unable to set extend dataset (Dimension cannot exceed the existing maximal size (new: 100 max: 1)) IOError: Unable to open file (Unable to open file: name = 'x3', errno = 2, error message = 'no such file or directory', flags = 0, o_flags = 0) KeyError: "Unable to open object (Object 'foo' doesn't exist)"
Improved setuptools support¶
setup.py now uses
setup_requires to make installation via pip friendlier.
Multiple low-level additions¶
Improved support for opening datasets via the low-level interface, by
H5Dopen2 and many new property-list functions.
Improved support for MPI features¶
Added support for retrieving the MPI communicator and info objects from an open file. Added boilerplate code to allow compiling cleanly against newer versions of mpi4py.
Readonly files can now be opened in default mode¶
When opening a read-only file with no mode flags, now defaults to opening the file on RO mode rather than raising an exception.
Single-step build for HDF5 on Windows¶
Building h5py on windows has typically been hamstrung by the need to build a compatible version of HDF5 first. A new Paver-based system located in the “windows” distribution directory allows single-step compilation of HDF5 with settings that are known to work with h5py.
For more, see:
Pierre de Buyl
Robert David Grant
Many others who contributed bug reports