NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI
PyPI, #2530 most downloaded on PyPI
Hierarchical datasets for Python
Latest stable
3.11.1
Last release 8 months ago
01 Mar 2026
Release timing varies
gaps range from 2 months to 2.2 years
Some releases are documented
notes for 17 of 45 stable releases
1 version withdrawn
withdrawn after publishing
15 years old
49 releases · first in 2011
Drag across the ruler to compare, or tap a release to open it.
Fix blosc2 loading (see #1305).
The tests folders have been moved outside the tables package to make it lighter once installed.
Consequently the tables.test function is no longer available.
Moreover, the print_versions function has been moved to tables.utils.
The setialization-de-serialization of attributes and datasets via pickle now raised a PickleSecurityWarning (see section below).
Fix path traversal in filenode.read_from_filenode. When the target is a directory, the output file name is taken from the filenode's _filename attribute, which is stored in the HDF5 file being read, so a crafted file could use a name such as ../x or an absolute path to write outside that directory. A stored name that is not a plain file name is now rejected with a ValueError.
Fix heap buffer overflow reading a single-column compound datatype with trailing padding. Description.__init__ only rebuilt the offset-aware NumPy dtype (keeping the on-disk itemsize that carries the padding) when the compound had more than one column, so a compound with a single member smaller than its H5Tget_size was repacked and the NumPy container ended up smaller than the on-disk element: H5Dread/H5Aread then wrote past it for a flat table column, a single-member nested column and an attribute alike. The on-disk itemsize is now kept from one column on.
Fix heap buffer overflow reading a nested H5T_ARRAY datatype (an array whose base type is itself an array). hdf5_to_np_ext_type kept the dimensions of the outer array but discarded those of the base, so the NumPy container was built smaller than H5Tget_size while H5Dread and H5Aread wrote the full element, overrunning the buffer for an Array leaf, a nested-array table column and an attribute alike. The base dimensions are now carried through, so the value reads back with its full shape.
Fix crash reading a table whose compound type has an integer column with no matching native size (3, 5, 6 or 7 bytes). get_nested_type builds the native in-memory compound sized from the on-disk type and inserts each member at its on-disk offset, which only holds while every member keeps its on-disk size. An odd-precision integer has no NumPy counterpart, so get_native_type widens it and the inserted member overruns the compound; H5Dread then walks past the malformed datatype on read (crash in H5T__sort_value). Such a column is now rejected and the leaf opens as UnImplemented.
Fix heap buffer overflow and wild pointer dereference reading a leaf whose CLASS attribute disagrees with its datatype. A CLASS attribute selects the reader without re-checking the datatype, so a foreign file could point the VLArray reader at a fixed H5T_ARRAY dataset (read as an array of hvl_t handles, overflowing the row buffer and dereferencing file-controlled bytes as pointers) or the Array/CArray/EArray readers at a variable-length dataset (hvl_t handles written into a buffer sized for the base atom). H5VLARRAYget_info now requires an H5T_VLEN datatype and Array._open_array rejects H5T_VLEN, so the leaf opens as UnImplemented instead.
Build wheels against HDF5 2.2.0; fix blosc2 filter buffer size.
The HDF5 1.14.6 bundled in PyTables v3.11.1 wheels was affected by ~30 published CVEs whose fixes shipped only on the HDF5 2.x line (see HDFGroup/cve_hdf5), including CVE-2025-44905 and CVE-2025-44904 (both rated 8.8 HIGH by NVD) and the CVE-2026-17572/17573/17574 batch fixed in 2.2.0.
Fix heap buffer overflow reading Blosc2 datasets whose filter typesize disagrees with the datatype. get_blosc2_slice took the element size from the Blosc2 filter metadata (cd_values[2]) instead of the dataset's datatype, so a crafted file could make the optimized slice read copy past the receive buffer. The optimized path now falls back to the plain HDF5 read when the two sizes disagree.
Fix heap buffer overflow reading tables with a variable-length column. get_nested_type mapped a variable-length compound member to its base atom and dropped the on-disk hvl_t size, so the NumPy record was smaller than the on-disk row while H5Dread filled the whole hvl_t, writing past the read buffer. Such a column is now rejected and the leaf opens as UnImplemented.
Fix out-of-bounds write opening a Table or VLArray leaf whose dataset has more than one dimension. Table._get_info and H5VLARRAYget_info read the dataset extent into a single hsize_t (both are 1-D entities), so a chunked compound or vlen dataset of higher rank made H5Sget_simple_extent_dims write past that buffer, reachable from a forged CLASS attribute. Both readers now reject a rank other than one and the leaf opens as UnImplemented.
Keep allow_pickle=False when dereferencing an ExternalLink. The linked file was opened with the default policy, so a file opened with unpickling disabled could still have pickles loaded from a second file it links to. The linked file now inherits the disabled policy unless allow_pickle is passed explicitly to the link call.
Do not use plain eval in PyTables utilities (fixes security issues GHSA-54cf-28h8-9p23 and GHSA-6mmx-p77c-4hmr).
Fix NULL-pointer dereference when opening a malformed HDF5 root attribute (Closes GHSA-jc4m-9v2h-9348).
Pickle payloads can execute arbitrary code while they are loaded.
Untrusted files can now be opened with open_file(..., allow_pickle=False).
The parameters.ALLOW_PICKLE parameter controls the default for files opened without an explicit allow_pickle argument; set it to False to disable unpickling globally for new file handles.
The same per-file opt-out can be passed through pandas as pandas.read_hdf(..., allow_pickle=False) or pandas.HDFStore(..., allow_pickle=False).
Writing remains enabled and emits PickleSecurityWarning, since the resulting data requires unpickling to be read. The ptdump and ptrepack utilities keep unpickling disabled unless --allow-pickle is given, and pttree never unpickles.
Automatic unpickling remains enabled by default for compatibility.
The default is expected to change to disabled in PyTables 3.13.
See also GHSA-3r7g-mq96-7hhc.
Point selections preserve coordinate views and accept read-only index arrays.
Atom shapes reject non-integer dimensions instead of silently truncating them.
Array indexing rejects out-of-range and non-integer indices before converting coordinates or accessing data.
Reusing an Expr after selecting an input range or a smaller output no longer truncates later results. Evaluation and iteration preserve the expression's input shape, so ranges can also be reset between calls.
Columns named like a Numexpr function (e.g. where or abs) can be used in Table.where and the other query methods. They were taken for the function, so the query failed (638).
Reading a table whose on-disk compound type has a nested compound column with padding no longer overflows the read buffer. The description now keeps the on-disk itemsize and member offsets of nested columns (as it already did for the top-level type, see #765), so the NumPy container matches H5Tget_size and H5Dread cannot write past it.
indexesextension.keysort now rejects a second array shorter than the first instead of reordering it out of bounds. It reorders the first len(array1) elements of array2 in place, so a shorter array2 was read and written past its end.
Indexed queries now reject an index whose bounds or ranges arrays do not match the dtype (and, for ranges, the shape) of its sorted values. The search buffers are sized after sorted, so a wider bounds array in a crafted file made H5Dread write past them.
Opening a leaf or attribute whose datatype is an enum with a base integer wider than 8 bytes no longer overflows a heap buffer. get_native_type and get_type_enum asked HDF5 for the native type with H5Tget_native_type, which overruns its buffer for such a base; the width is now checked first and the type is treated as unsupported.
Group and node walks, including ptdump, terminate when group hard links form cycles (794). Links to an ancestor are listed without traversing their children again; non-cyclic aliases retain their descendants.
Preserve pending Row.update changes when iteration stops early. Table.flush, File.flush and File.close now write them even if the iterator has been discarded (8, 1098).
Table.where (and hence Table.read_where, Table.get_where_list and Table.append_where) now treats a start with no stop like a Python slice, i.e. the rows from start to the last one are considered. Previously only one row was considered (797).
Preserve field values when Table.append converts structured arrays with a different dtype or byteorder (658), and copy strided arrays before appending their records.
Fix Table.remove_rows with a step greater than 1, which removed the wrong rows or raised OverflowError. An empty range now removes nothing instead of raising HDF5ExtError, and Table.remove_row accepts negative indices.
Indexing a multidimensional Array with a single list or 1-D array of integers, e.g. array[[0, 2]], now selects along the first axis like NumPy instead of raising HDF5ExtError.
Modifying a nested column of an indexed table (with Row.update, Table.modify_column or Table.modify_columns) no longer raises KeyError, and indexes on its subcolumns are marked dirty (699).
Setting a table row from a record scalar, e.g. table[2] = table[0], no longer raises ValueError.
Queries that compare an indexed column with another column, e.g. table.where("a > b") or table.where("a == a"), no longer raise AttributeError: 'Column' object has no attribute 'tolist'. The index of such a column is not used for that comparison.
Table.iterrows and Table.read with a negative step now return the same rows as the equivalent Python slice. Before, they could raise OverflowError, skip the last row or (for Table.read) return wrong or uninitialized values.
Table.append accepts a single row given as a tuple or a record (e.g. table.append(table[0])) instead of raising IndexError, and Table.modify_rows accepts a single record.
The error raised by EArray.append for an object of the wrong rank now shows the rank of the EArray instead of (myrank).
VLArray.get_row_size accepts negative row numbers, counting from the end like indexing does, instead of raising OverflowError.
Fix issue with non-zero direct-chunk filter mask (1325).
Fix infinite busy loop in ObjectCache.updateslot (1254).
Fix VLArray.append after truncate() writing past the new extent and reading back empty values (1102).
Fix a one byte stack buffer overflow when opening a dataset whose HDF5 datatype has a byte order PyTables does not support, i.e. VAX ordered or mixed endian types.
Fix a heap buffer overflow when reading a compound (record) attribute whose HDF5 datatype carries trailing padding; the read buffer now keeps the on-disk itemsize, matching how padded tables are already handled.
Fix an out of bounds stack read when opening a dataset whose filter pipeline holds more client data values than get_filter_names() reads at once; the count HDF5 reports is now clamped to the buffer it filled. Filter client data values above 0x7fffffff are also no longer reported as negative numbers on platforms where long is 32 bit, i.e. Windows.
Fix a heap buffer overflow when reading an H5T_BITFIELD attribute or dataset wider than one byte. PyTables mapped every bitfield to the 1-byte bool dtype, so the read buffer was sized below the on-disk type and H5Aread/H5Dread wrote past it. Only 1-byte bitfields (booleans) are produced by PyTables, so wider ones from foreign files are now reported as an unsupported type instead.
Fix a heap buffer overflow when reading a string system attribute that holds more than one element. H5ATTRget_attribute_string sized its buffer for a single element, but H5Aread fills the whole attribute, so a non-scalar string attribute wrote past it. is_pytables_file reads PYTABLES_FORMAT_VERSION through that path, so merely identifying a foreign file could trigger the overflow; the buffer is now sized for the whole attribute.
Fix typos in docstrings, comments and error messages.
Minimum Python version is now 3.11.
Minimum Cython version is now v3.2.
Minimum numpy version is nor 2.0.
Add git submodule for hdf5-blosc2 v3.0.1.
Fix compatibility with c-blosc2 v3.x
Read scalar HDF5 enum attributes as NumPy scalars with the corresponding integer dtype, including enum attributes stored with a foreign byte order.
Move test data to a dedicated sub-folder to tests.
Add Windows ARM64 wheels builds.
Fix compatibility with numpy >= 2.5.
Update the vendored c-blosc to v1.21.7.
Thanks to:
Teddy Tennant
Jacob Rideout
Adrian Altenhoff
Max (maxtaran2010)
YuuLuo
Raashish Aggarwal
Naveed
CAOShurong
Eric Larson
Nikhil Dabas
Lim Jooyoung
Jeff
Jay (jay23606)
Your coding agent can read these notes before it upgrades. Set up the MCP server →
Set Cython "freethreading_compatible" directive. Please refer to https://www.pytables.org/cookbook/threading.html for details about threading support.
Thanks to:
Wheels for Python v3.13 are now provided (#1217).
numpy.typing.DTypeLike over np.dtype for parameters. Thanks to Joel T. Collins.IsDescription, dict, numpy.dtype as table descriptions. Thanks to Joel T. Collins.chunkshape when creating arrays. Thanks to Joel T. Collins.setup.py finds runtime libs (see #1219 and #1223). Thanks to Jason Bacon.createparents=True (#1208).README.rst (#1218). Thanks to David Poznik.sys.maxint in pttree (#1224).Fix version constraints for the numpy runtime requirements (#1204). For a mistake it didn't allow to use PyTables with numpy 2.x.
numpy runtime requirements
(#1204).
For a mistake it didn't allow to use PyTables with numpy 2.x.1205), Thanks to Michał Górny.numpy >= 2 as build constraint (see discussion in #1200).faulthandler in tables.tests.test_all. Thanks to Eric Larson.New direct chunking API which allows access to raw chunk data skipping the HDF5 filter pipeline (cutting overhead, see "Optimization tips" in User's G
atom.py (#1079). This also narrows some types, only allowing bytes to be stored in VLStringAtom and only str in VLUnicodeAtom. Thanks to Nils Carlson.dparams.schunk on decompression (#1110 and #1109). Thanks to Tom Birch.createparents argument in Leaf.copy (#1127 and #1125). Thanks to Ko Stehner.ExternalLink class (#1095). Thanks to erikdl-zeiss.test_queries (#1169). Thanks to Miro Hrončok.TypeError when computing Blosc2 search paths with missing library (#1188 and #1100). Thanks to martinowitsch, Padraic Calpin and Eric Larson.RuntimeWarning on NumPy expectedrows value (#1010). Thanks to wony-zheng and Ivan Vilata.MACOSX_DEPLOYMENT_TARGET in Docker (#1050 and #1165). Thanks to Clemens Brunner, Antonio Valentino, Maximilian Linhoff and Eric Larson.getLibrary C code. Thanks to Antonio Valentino.In alphabetical order:
Fix the assembly of returned slice data in Blosc2 NDim optimized slice reads by using Blosc2's b2nd_copy_buffer (#1078). The bug only showed up when t
b2nd_copy_buffer (#1078). The bug only showed up when the chunk did not fully cover the innermost dimension. Add unit tests to ckeck for regressions, along with foreign-generated files, and enable and fix Blosc2 NDim tests which were not being run. Thanks to Ivan Vilata.b2nd_copy_buffer function).In alphabetical order:
Avoid NumPy >= 1.25 deprecations, use numpy.all, numpy.any, etc. instead. Thanks to Antonio Valentino.
CArray and EArray, with Blosc2 NDim 2-level partitioning for multidimensional arrays (#1056). See "Multidimensional slicing and chunk/block sizes" in the User's Guide. Thanks to Marta Iborra and Ivan Vilata. This development was funded by a NumFOCUS grant.Col._v_col_attrs (#893 and #821). Thanks to Jonathan Wheeler, Thorben Menne, Ezequiel Cimadevilla Alvarez, odidev, Sander Roet, Antonio Valentino, Munehiro Nishida, Zbigniew Jędrzejewski-Szmek, Laurent Repiton, xmatthias, Logan Kilpatrick.H5Dchunk_iter (when available) to speed up walking over many chunks in a very large table, as well as with random reads (#991, #997, #999). Thanks to Francesc Alted and Mark Kittisopikul.setup.py (now using pyproject.toml as per PEP 518) and blosc2 discovery mechanism. Blosc2 may be used both via python-blosc2 or system c-blosc2 (#987, #1000, #998, #1017, #1045). Thanks to Antonio Valentino, Ben Greiner, Iwo-KX, nega.SECURITY.md with security policy (#1012 and #1011). Thanks to Joyce Brum.numpy.all, numpy.any, etc. instead. Thanks to Antonio Valentino.cibuildwheel, more clear distinctions between build and runtime dependencies.ptrepack failures with external links (#938 and #990). Thanks to Adrian Altenhoff.File.get_node docstring (#1021). Thanks to Steffen Rehberg.extern "C" block (#1026). Thanks to Ivan Vilata.In alphabetical order:
Avoid NumPy >= 1.25 deprecations, use numpy.all, numpy.any, etc. instead. Thanks to Antonio Valentino.
CArray and EArray, with Blosc2 NDim 2-level partitioning for multidimensional arrays (#1056). See "Multidimensional slicing and chunk/block sizes" in the User's Guide. Thanks to Marta Iborra and Ivan Vilata. This development was funded by a NumFOCUS grant.Col._v_col_attrs (#893 and #821). Thanks to Jonathan Wheeler, Thorben Menne, Ezequiel Cimadevilla Alvarez, odidev, Sander Roet, Antonio Valentino, Munehiro Nishida, Zbigniew Jędrzejewski-Szmek, Laurent Repiton, xmatthias, Logan Kilpatrick.H5Dchunk_iter (when available) to speed up walking over many chunks in a very large table, as well as with random reads (#991, #997, #999). Thanks to Francesc Alted and Mark Kittisopikul.setup.py (now using pyproject.toml as per PEP 518) and blosc2 discovery mechanism. Blosc2 may be used both via python-blosc2 or system c-blosc2 (#987, #1000, #998, #1017, #1045). Thanks to Antonio Valentino, Ben Greiner, Iwo-KX, nega.SECURITY.md with security policy (#1012 and #1011). Thanks to Joyce Brum.numpy.all, numpy.any, etc. instead. Thanks to Antonio Valentino.cibuildwheel, more clear distinctions between build and runtime dependencies.ptrepack failures with external links (#938 and #990). Thanks to Adrian Altenhoff.File.get_node docstring (#1021). Thanks to Steffen Rehberg.extern "C" block (#1026). Thanks to Ivan Vilata.In alphabetical order:
Support for Python 3.11 has been added (PR #962).
Stop using appveyor and deprecated ci-helpers (closes #827).
Compatibility with Python 3.10, numpy 1.21 and HDF5 1.12.
Support for Python 3.5 has been dropped (#840 and #850).
Windows: Significantly faster import tables PR #781. Thanks to Christoph Gohlke.
Internal C-Blosc sources updated to 1.21.1 (#931). Note that, starting from C-Blosc 1.19 does not include the Snappy codec sources anymore, so Snappy will be not available if you compile from included sources; other packages (like conda or wheels), may (or may not) include it.
Stop using appveyor and deprecated ci-helpers (closes #827).
Switch to git submodule for the management of vendored c-blosc sources.
CI moved to GitHub Actions (GHA).
Drop Travis-CI.
Improved code formatting and notation consistency (#873, #868, #865 thanks to Miroslav Šedivý).
Improve the use of modern Python including pathlib, f-strings (#859, #855, #839 and #818 thanks to Miroslav Šedivý).
Several improvements to wheels generation in CI (thanks to Andreas Motl @amotl and Matthias @xmatthias).
Simplified management of version information.
Drop dependency on the deprecated distutils.
Modernize the setup script and add support for PEP517 (#907).
Fix pkg-config (setup.py) for Python 3.9 on Debian. Thanks to Marco Sulla PR #792.
Fix ROFileNode fails to return the fileno() (#633).
Do not flush read only files (#915 thanks to @lrepiton).
Drop the deprecated hdf5Version and File.open_count.
the get_tables_version and get_hdf5_version functions are now deprecated please use the coresponding tables.__version__ and tables.hdf5_version instead.
Maintenance release to fix packaging issues.
Maintenance release to fix packaging issues.
No new features or bug fixes.
Maintenance release to fix compatibility with Python 3.8
Maintenance release to fix compatibility with Python 3.8
Maintenance release to fix how PyPi repo is handling wheel versions.
description is a NumPy struct array with padding, this
is honored now. The previous behaviour (i.e. getting rid of paddings) can
be replicated by passing the new allow_padding parameter when opening
a file. For some examples, see the new examples/tables-with-padding.py
and examples/attrs-with-padding.py. For details on the implementation
see :issue:720.--dont-allow-padding in ptrepack utility so as to
replicate the previous behaviour of removing padding during file copies.
The default is to honor the original padding in copies.Environment variable to control the use of embedded libraries. Thanks to avalentino.
690. Thanks to katrinleinweber.685. Thanks to cgohlke.687. Thanks to mingwandroid.688 & :issue:689. Thanks to oleksandr-pavlyk.Fixed deprecation warnings with numpy 1.14. Thanks to oleksandr-pavlyk.
__dir__() method
autocompletes children that are named as valid python identifiers.
:issue:624 & :issue:625 thanks to ankostis.Group.__getitem__() to have groups act as python-containers,
so code like this works: hfile.root['some child'].
:issue:628 thanks to ankostis.os.fspath method.
Thanks to mruffalo._convert_time64() with an empty
nparr array.
:issue:653. Thanks to alobbs.675. Thanks to jwiggins.651. Thanks to jackdbd.setup.py detects conda env and uses installed conda (hdf5, bzip2, lzo and/or blosc) packages when building from source.
616. Thanks to oscar6echo.Fixed critical bug in ptrepack
Fixed critical bug in ptrepack
Support for HDF5 v1.10.x (see 582)
582)590. Thanks to Yaroslav Halchenkolong type to six.integer_types in atom.py.
See also :issue:598. Thanks to Kyle Keppler for reporting.583.514Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
One column per quarter.