What’s new in h5py 3.0

New features

  • The interface for storing & reading strings has changed - see Strings in HDF5. The new rules are hopefully more consistent, but may well require some changes in coding using h5py.

  • Reading & writing data now releases the GIL, so another Python thread can continue while HDF5 accesses data. Where HDF5 can call back into Python, such as for data conversion, h5py re-acquires the GIL. However, HDF5 has its own global lock, so this won’t speed up parallel data access using multithreading.

  • Numpy datetime and timedelta arrays can now be stored and read as HDF5 opaque data (GH1339), though other tools will not understand them. See Storing other types as opaque data for more information.

  • New Dataset.iter_chunks() method, to iterate over chunks within the given selection.

  • Compatibility with HDF5 1.12.

  • Methods which accept a shape tuple, e.g. to create a dataset, now also allow an integer for a 1D shape (PR 1340).

  • Casting data to a specified type on reading (Dataset.astype()) can now be done without a with statement, like this:

    data = dset.astype(np.int32)[:]
  • A new Dataset.fields() method lets you read only selected fields from a dataset with a compound datatype.

  • Reading data has less overhead, as selection has been implemented in Cython. Making many small reads from the same dataset can be as much as 10 times faster, but there are many factors that can affect performance.

  • A new NumPy-style Dataset.nbytes attribute to get the size of the dataset’s data in bytes. This differs from the size attribute, which gives the number of elements.

  • The external argument of Group.create_dataset(), which specifies any external storage for the dataset, accepts more types (GH1260), as follows:

    • The top-level container may be any iterable, not only a list.
    • The names of external files may be not only str but also bytes or os.PathLike objects.
    • The offsets and sizes may be NumPy integers as well as Python integers.

    See also the deprecation related to the external argument.

  • Support for setting file space strategy at file creation. Includes option to persist empty space tracking between sessions. See File for details.

  • More efficient writing when assiging a scalar to a chunked dataset, when the number of elements to write is no more than the size of one chunk.

  • Introduced support for the split file driver (PR 1468).

  • Allow making virtual datasets which can grow as the source data is resized - see Virtual Datasets (VDS).

  • New allow_unknown_filter option to Group.create_dataset(). This should only be used if you will compress the data before writing it with the low-level write_direct_chunk() method.

  • The low-level chunk query API provides information about dataset chunks in an HDF5 file: get_num_chunks(), get_chunk_info() and get_chunk_info_by_coord().

  • The low-level h5py.h5f.FileID.get_vfd_handle() method now works for any file driver that supports it, not only the sec2 driver.

Breaking changes & deprecations

  • h5py now requires Python 3.6 or above; it is no longer compatible with Python 2.7.

  • The default mode for opening files is now ‘r’ (read-only). See Opening & creating files for other possible modes if you need to write to a file.

  • In previous versions, creating a dataset from a list of bytes objects would choose a fixed length string datatype to fit the biggest item. It will now use a variable length string datatype. To store fixed length strings, use a suitable dtype from h5py.string_dtype().

  • Variable-length UTF-8 strings in datasets are now read as bytes objects instead of str by default, for consistency with other kinds of strings. See Strings in HDF5 for more details.

  • When making a virtual dataset, a dtype must be specified in VirtualLayout. There is no longer a default dtype, as this was surprising in some cases.

  • The external argument of Group.create_dataset() no longer accepts the following forms (GH1260):

    • a list containing name, [offset, [size]];
    • a list containing name1, name2, …; and
    • a list containing tuples such as (name,) and (name, offset) that lack the offset or size.

    Furthermore, each nameoffsetsize triplet now must be a tuple rather than an arbitrary iterable. See also the new feature related to the external argument.

  • The MPI mode no longer supports mpi4py 1.x.

  • The deprecated h5py.h5t.available_ftypes dictionary was removed.

  • The deprecated Dataset.value property was removed. Use ds[()] to read all data from any dataset.

  • The deprecated functions new_vlen, new_enum, get_vlen and get_enum have been removed. See Special types for the newer APIs.

  • Removed deprecated File.fid attribute. Use File.id instead.

  • Remove the deprecated h5py.highlevel module. The high-level API is available directly in the h5py module.

  • The third argument of h5py._hl.selections.select() is now an optional high-level Dataset object, rather than a DatasetID. This is not really a public API - it has to be imported through the private _hl module - but probably some people are using it anyway.

Exposing HDF5 functions

  • H5Dget_num_chunks
  • H5Dget_chunk_info
  • H5Dget_chunk_info_by_coord
  • H5Oget_info1
  • H5Oget_info_by_name1
  • H5Oget_info_by_idx1
  • H5Ovisit1
  • H5Ovisit_by_name1
  • H5Pset_attr_phase_change
  • H5Pset_fapl_split
  • H5Pget_file_space_strategy
  • H5Pset_file_space_strategy
  • H5Sencode1
  • H5Tget_create_plist

Bug fixes

  • Fix segmentation fault when accessing vlen of strings (GH1336).
  • Fix the storage of non-contiguous arrays, such as numpy slices, as HDF5 vlen data (GH1649).
  • Fix pathologically slow reading/writing in certain conditions with integer indexing (GH492).
  • Fix bug when Group.copy() source is a high-level object and destination is a Group (GH1005).
  • Fix reading data for region references pointing to an empty selection.
  • Unregister converter functions at exit, preventing segfaults on exit in some situations with threads (PR 1440).
  • As HDF5 1.10.6 and later support UTF-8 paths on Windows, h5py built against HDF5 1.10.6 will use UTF-8 for file names, allowing all filenames.
  • Fixed h5py.h5d.DatasetID.get_storage_size() to report storage size of zero bytes without raising an exception (GH1475).
  • Attribute Managers (obj.attrs) can now work on HDF5 stored datatypes (GH1476).
  • Remove broken inherited ds.dims.values() and ds.dims.items() methods. The dimensions interface behaves as a sequence, not a mapping (GH744).
  • Fix creating attribute with Empty by converting its dtype to a numpy dtype object.
  • Fix getting maxshape on empty/null datasets.
  • The File.swmr_mode property is always available (GH1580).
  • The File.mode property handles SWMR access modes in addition to plain RDONLY/RDWR modes
  • Importing an MPI build of h5py no longer initialises MPI immediately, which will hopefully avoid various strange behaviours.
  • Avoid launching a subprocess by using platform.machine() at import time. This could trigger a warning in MPI.
  • Removed an equality comparison with an empty array, which will cause problems with future versions of numpy.
  • Better error message if you try to use the mpio driver and h5py was not built with MPI support.
  • Improved error messages when requesting chunked storage for an empty dataset.
  • Data conversion functions should fail more gracefully if no memory is available.
  • Fix some errors for internal functions that were raising “TypeError: expected bytes, str found” instead of the correct error.
  • Use relative path for virtual data sources if the source dataset is in the same file as the virtual dataset.
  • Generic exception types used in tests’ assertRaise (exception types changed in new HDF5 version)
  • Use dtype=object in tests with ragged arrays

Building h5py

  • The setup.py configure command was removed. Configuration for the build can be specified with environment variables instead. See Custom installation for details.

  • It is now possible to specify separate include and library directories for HDF5 via environment variables. See Custom installation for more details.

  • The pkg-config name to use when looking up the HDF5 library can now be configured, this can assist with selecting the correct HDF5 library when using MPI. See Custom installation for more details.

  • Using bare char* instead of array.array in h5d.read_direct_chunk since array.array is a private CPython C-API interface

  • Define NPY_NO_DEPRECATED_API to silence a warning.

  • Make the lzf filter build with HDF5 1.10 (GH1219).

  • If HDF5 is not loaded, an additional message is displayed to check HDF5 installation

  • Rely much more on the C-interface provided by Cython to call Python and NumPy.

  • Removed an old workaround which tried to run Cython in a subprocess if cythonize() didn’t work. This shouldn’t be necessary for any recent version of setuptools.

  • Migrate all Cython code base to Cython3 syntax

    • The only noticeable change is in exception raising from cython which use bytes
    • Massively use local imports everywhere as expected from Python3
    • Explicitly mark several Cython functions as non-binding


  • Unregistering converter functions on exit (PR 1440) should allow profiling and code coverage tools to work on Cython code.