What's new in h5py 2.9
======================

New features
------------

* A convenient high-level API for creating virtual datasets, HDF5 objects which
  act as a view over one or more real datasets (:issue:`1060`, :issue:`1126`).
  See :ref:`vds` for details.
* :class:`File` can now be constructed with a Python file-like object,
  making it easy to create an HDF5 file in memory using :class:`io.BytesIO`
  (:issue:`1061`, :issue:`1105`, :issue:`1116`).
  See :ref:`file_fileobj` for details.
* :class:`File` now accepts parameters to control the chunk cache
  (:issue:`1008`). See :ref:`file_cache` for details.
* New options to record the order of insertion for attributes and group entries.
  Iterating over these collections now follows insertion order if it was
  recorded, or name order if not (:issue:`1098`).
* A new method :meth:`Group.create_dataset_like` to create a new dataset with
  similar properties to an existing one (:issue:`1085`).
* Datasets can now be created with storage backed by external non-HDF5 files
  (:issue:`1000`).
* Lists or tuples of unicode strings can now be stored as HDF5 attributes
  (:issue:`1032`).
* Inspecting the view returned by ``.keys()`` now shows the key names, for
  convenient interactive use (:issue:`1049`).

Exposing HDF5 functions
-----------------------

* ``H5LTopen_file_image`` as :func:`h5py.h5f.open_file_image` (:issue:`1075`).
* External dataset storage functions ``H5Pset_external``, ``H5Pget_external``
  and ``H5Pget_external_count`` as methods on :class:`h5py.h5p.PropDCID`
  (:issue:`1000`).

Bugfixes
--------

* Fix reading/writing of float128 data (:issue:`1114`).
* Converting data to float16 when creating a dataset (:issue:`1115`).

Support for old Python
----------------------

Support for Python 3.3 has been dropped.

Support for Python 2.6 has been dropped.