Release notes for PyTables 2.2 series
|Author:||Francesc Alted i Abad|
Updated Blosc to 1.0 (final).
Filter ID of Blosc changed from wrong 32010 to reserved 32001. This will prevent PyTables 2.2 (final) to read files created with Blosc and PyTables 2.2 pre-final. ptrepack can be used to retrieve those files, if necessary. More info in ticket #281.
Recent benchmarks suggest a new parametrization is better in most scenarios:
- The default chunksize has been doubled for every dataset size. This works better in most of scenarios, specially with the new Blosc compressor.
- The HDF5 CHUNK_CACHE_SIZE parameter has been raised to 2 MB in order to better adapt to the chunksize increase. This provides better hit ratio (at the cost of consuming more memory).
Some plots have been added to the User's Manual (chapter 5) showing how the new parametrization works.
- A new version of Blosc (0.9.5) is included. This version is now considered to be stable and apt for production. Thanks for all PyTables users that have contributed to find and report bugs.
- Added a new IO_BUFFER_SIZE parameter to tables/parameters.py that allows to set the internal PyTables' buffer for doing I/O. This replaces CHUNKTIMES but it is more general because it affects to all Leaf objects and also the tables.Expr module (and not only tables as before).
- BUFFERTIMES parameter in tables/parameters.py has been renamed to BUFFER_TIMES which is more consistent with other parameter names.
- On Windows platforms, the path to the tables module is now appended to sys.path and the PATH environment variable. That way DLLs and PYDs in the tables directory are to be found now. Thanks to Christoph Gohlke for the hint.
- A replacement for barriers for Mac OSX, or other systems not implementing them, has been carried out. This allows to compile PyTables on such platforms. Fixes #278
- Fixed a couple of warts that raise compatibility warnings with forthcoming Python 2.7.
- HDF5 1.8.5 is used in Windows binaries.
Numexpr is not included anymore in PyTables and has become a requisite instead. This is because Numexpr already has decent enough installers and is available in the PyPI repository also, so it should be easy for users to fulfill this dependency.
When using a Numexpr package that is turbo-loaded with Intel's VML/MKL, the parameter MAX_THREADS will control the number of threads that VML can use during computations. For a finer control, the numexpr.set_vml_num_threads() can always be used.
Cython is used now instead of Pyrex for Pyrex extensions.
Updated to 0.9 version of Blosc compressor. This version can make use of threads so as to accelerate the compression/decompression process. In order to change the maximum number of threads that Blosc can use (2 by default), you can modify the MAX_THREADS variable in tables/parameters.py or make use of the new setBloscMaxThreads() global function.
Reopening already opened files is supported now, provided that there is not incompatibility among intended usages (for example, you cannot reopen in append mode an already opened file in read-only mode).
Option --print-versions for test_all.py script is now preferred over the deprecated --show-versions. This is more consistent with the existing print_versions() function.
Fixed a bug that, under some circumstances, prevented the use of table iterators in itertool.groupby(). Now, you can safely do things like:
sel_rows = table.where('(row_id >= 3)') for group_id, grouped_rows in itertools.groupby(sel_rows, f_group): group_mean = average([row['row_id'] for row in grouped_rows])
Copies of Array objects with multidimensional atoms (coming from native HDF5 files) work correctly now (i.e. the copy holds the atom dimensionality). Fixes #275.
The tables.openFile() function does not try anymore to open/close the file in order to guess whether it is a HDF5 or PyTables one before opening it definitely. This allows the fcntl.flock() and fcntl.lockf() Python functions to work correctly now (that's useful for arbitrating access to the file by different processes). Thanks to Dag Sverre Seljebotn and Ivan Vilata for their suggestions on hunting this one! Fixes #185.
The estimation of the chunksize when using multidimensional atoms in EArray/Carray was wrong because it did not take in account the shape of the atom. Thanks to Ralf Juengling for reporting. Fixes #273.
Non-contiguous arrays can now safely be saved as attributes. Before, if arrays were not contiguous, incorrect data was saved in attr. Fixes #270.
EXTDIM attribute for CArray/EArray now saves the correct extendeable dimension, instead of rubbish. This does not affected functionality, because extendeable dimension was retrieved directly from shape information, but it was providing misleading information to the user. Fixes #268.
- Now, Table.Cols.__len__() returns the number of top level columns instead of the number of rows in table. This is more consistent in that Table.Cols is an accessor for columns. Fixes #276.
- Blosc compressor has been added as an additional filter, in addition to the existing Zlib, LZO and bzip2. This new compressor is meant for fast compression and extremely fast decompression. Fixes #265.
- In File.copyFile() method, copyuserattrs was set to false as default. This was unconsistent with other methods where the default value for copyuserattrs is true. The default for this is true now. Closes #261.
- tables.copyFile and File.copyFile recognize now the parameters present in tables/parameters.py. Fixes #262.
- Backported fix for issue #25 in Numexpr (OP_NEG_LL treats the argument as an int, not a long long). Thanks to David Cooke for this.
- CHUNK_CACHE_NELMTS in tables/paramters.py set to a prime number as Neil Fortner suggested.
- Workaround for a problem in Python 2.6.4 (and probably other versions too) for pickling strings like "0" or "0.". Fixes #253.
- Support for HDF5 hard links, soft links and external links (when PyTables is compiled against HDF5 1.8.x series). A new tutorial about its usage has been added to the 'Tutorials' chapter of User's Manual. Closes #239 and #247.
- Added support for setting HDF5 chunk cache parameters in file opening/creating time. 'CHUNK_CACHE_NELMTS', 'CHUNK_CACHE_PREEMPT' and 'CHUNK_CACHE_SIZE' are the new parameters. See "PyTables' parameter files" appendix in User's Manual for more info. Closes #221.
- New Unknown class added so that objects that HDF5 identifies as H5G_UNKNOWN can be mapped to it and continue operations gracefully.
- Added flag --dont-create-sysattrs to ptrepack so as to not create sys attrs (default is to do it).
- Support for native compound types in attributes. This allows for better compatibility with HDF5 files. Closes #208.
- Support for native NumPy dtype in the description parameter of File.createTable(). Closes #238.
- Added missing _c_classId attribute to the UnImplemented class. ptrepack no longer chokes while copying Unimplemented classes.
- The FIELD_* sys attrs are no longer copied when the PYTABLES_SYS_ATTRS parameter is set to false.
- File.createTable() no longer segfaults if description=None. Closes #248.
- Workaround for avoiding a Python issue causing a segfault when saving and then retrieving a string attribute with values "0" or "0.". Closes #253.
Row.__contains__() disabled because it has little sense to query for a key in Row, and the correct way should be to query for it in Table.colnames or Table.colpathnames better. Closes #241.
[Semantic change] To avoid a common pitfall when asking for the string representation of a Row class, Row.__str__() has been redefined. Now, it prints something like:
>>> for row in table: ... print row ... /newgroup/table.row (Row), pointing to row #0 /newgroup/table.row (Row), pointing to row #1 /newgroup/table.row (Row), pointing to row #2
>>> for row in table: ... print row ... ('Particle: 0', 0, 10, 0.0, 0.0) ('Particle: 1', 1, 9, 1.0, 1.0) ('Particle: 2', 2, 8, 4.0, 4.0)
Use print row[:] idiom if you want to reproduce the old behaviour. Closes #252.
- After some improvements in both HDF5 and PyTables, the limit before emitting a PerformanceWarning on the number of children in a group has been raised from 4096 to 16384.
Added Expr, a class for evaluating expressions containing array-like objects. It can evaluate expressions (like '3*a+4*b') that operate on arbitrary large arrays while optimizing the resources (basically main memory and CPU cache memory) required to perform them. It is similar to the Numexpr package, but in addition to NumPy objects, it also accepts disk-based homogeneous arrays, like the Array, CArray, EArray and Column PyTables objects.
Added support for NumPy's extended slicing in all Leaf objects. With that, you can do the next sort of selections:
array1 = array # simple selection array2 = array[4:1000:2] # slice selection array3 = array[1, ..., ::2, 1:4, 4:] # general slice selection array4 = array[1, [1,5,10], ..., -1] # fancy selection array5 = array[np.where(array[:] > 4)] # point selection array6 = array[array[:] > 4] # boolean selection
Thanks to Andrew Collette for implementing this for h5py, from which it has been backported. Closes #198 and #209.
Numexpr updated to 1.3.1. This can lead to up a 25% improvement of the time for both in-kernel and indexed queries for unaligned tables.
HDF5 1.8.3 supported.
- Fixed problems when modifying multidimensional columns in Table objects. Closes #228.
- Row attribute is no longer stalled after a table move or rename. Fixes #224.
- Array.__getitem__(scalar) returns a NumPy scalar now, instead of a 0-dim NumPy array. This should not be noticed by normal users, unless they check for the type of returned value. Fixes #222.
Added a dtype attribute for all leaves. This is the NumPy dtype that most closely matches the leaf type. This allows for a quick-and-dirty check of leaf types. Closes #230.
Added a shape attribute for Column objects. This is formed by concatenating the length of the column and the shape of its type. Also, the representation of columns has changed an now includes the length of the column as the leading dimension. Closes #231.
Added a new maindim attribute for Column which has the 0 value (the leading dimension). This allows for a better similarity with other *Array objects.
In order to be consistent and allow the extended slicing to happen in VLArray objects too, VLArray.__setitem__() is not able to partially modify rows based on the second dimension passed as key. If this is tried, an IndexError is raised now. Closes #210.
The forceCSI flag has been replaced by checkCSI in the next Table methods: copy(), readSorted() and itersorted(). The change reflects the fact that a re-index operation cannot be triggered from these methods anymore. The rational for the change is that an indexing operation is a potentially very expensive operation that should be carried out explicitly instead of being triggered by methods that should not be in charge of this task. Closes #216.
After the introduction of the shape attribute for Column objects, the shape information for multidimensional columns has been removed from the dtype attribute (it is set to the base type of the column now). Closes #232.
-- The PyTables Team