Migrating from PyTables 1.x to 2.x


Francesc Alted i Abad




Ivan Vilata i Balaguer



Next are described a series of issues that you must have in mind when migrating from PyTables 1.x to PyTables 2.x series.

New type system

In PyTables 2.x all the data types for leaves are described through a couple of classes:

  • Atom: Describes homogeneous types of the atomic components in *Array

    objects (Array, CArray, EArray and VLArray).

  • Description: Describes (possibly nested) heterogeneous types in Table objects.

So, in order to upgrade to the new type system, you must perform the next replacements:

  • *Array.stype –> *Array.atom.type (PyTables type)

  • *Array.type –> *Array.atom.dtype (NumPy type)

  • *Array.itemsize –> *Array.atom.itemsize (the size of the item)

Furthermore, the PyTables types (previously called “string types”) have changed to better adapt to NumPy conventions. The next changes have been applied:

  • PyTables types are now written in lower case, so ‘Type’ becomes ‘type’. For example, ‘Int64’ becomes now ‘int64’.

  • ‘CharType’ –> ‘string’

  • ‘Complex32’, ‘Complex64’ –> ‘complex64’, ‘complex128’. Note that the numeric part of a ‘complex’ type refers now to the size in bits of the type and not to the precision, as before.

See Appendix I of the Users’ Manual on supported data types for more information on the new PyTables types.

Important changes in Atom specification

  • The dtype argument of EnumAtom and EnumCol constructors has been replaced by the base argument, which can take a full-blown atom, although it accepts bare PyTables types as well. This is a mandatory argument now.

  • vlstring pseudo-atoms used in VLArray nodes do no longer imply UTF-8 (nor any other) encoding, they only store and load raw strings of bytes. All encoding and decoding is left to the user. Be warned that reading old files may yield raw UTF-8 encoded strings, which may be converted back to Unicode in this way:

    unistr = vlarray[index].decode('utf-8')

    If you need to work with variable-length Unicode strings, you may want to use the new vlunicode pseudo-atom, which fully supports Unicode strings with no encoding hassles.

  • Finally, Atom and Col are now abstract classes, so you can’t use them to create atoms or column definitions of an arbitrary type. If you know the particular type you need, use the proper subclass; otherwise, use the Atom.from_*() or Col.from_*() factory methods. See the section on declarative classes in the reference.

    You are also advised to avoid using the inheritance of atoms to check for their kind or type; for that purpose, use their kind and type attributes.

New query system

  • In-kernel conditions, since they are based now in Numexpr, must be written as strings. For example, a condition that in 1.x was stated as:

    result = [row['col2'] for row in table.where(table.cols.col1 == 1)]

    now should read:

    result = [row['col2'] for row in table.where('col1 == 1')]

    That means that complex selections are possible now:

    result = [ row['col2'] for row in
               table.where('(col1 == 1) & (col3**4 > 1)') ]
  • For the same reason, conditions for indexed columns must be written as strings as well.

New indexing system

The indexing system has been totally rewritten from scratch for PyTables 2.0 Pro Edition. The new indexing system has been included into PyTables with release 2.3. Due to this, your existing indexes created with PyTables 1.x will be useless, and although you will be able to continue using the actual data in files, you won’t be able to take advantage of any improvement in speed.

You will be offered the possibility to automatically re-create the indexes in PyTables 1.x format to the new 2.0 format by using the ptrepack utility.

New meanings for atom shape and *Array shape argument

With PyTables 1.x, the atom shape was used for different goals depending on the context it was used. For example, in createEArray(), the shape of the atom was used to specify the dataset shape of the object on disk, while in CArray the same atom shape was used to specify the chunk shape of the dataset on disk. Moreover, for VLArray objects, the very same atom shape specified the type shape of the data type. As you see, all of these was quite a mess.

Starting with PyTables 2.x, an Atom only specifies properties of the data type (à la VLArray in 1.x). This lets the door open for specifying multidimensional data types (that can be part of another layer of multidimensional datasets) in a consistent way along all the *Array objects in PyTables.

As a consequence of this, File.createCArray() and File.createVLArray() methods have received new parameters in order to make possible to specify the shapes of the datasets as well as chunk sizes (in fact, it is possible now to specify the latter for all the chunked leaves, see below). Please have this in mind during the migration process.

Another consequence is that, now that the meaning of the atom shape is clearly defined, it has been chosen as the main object to describe homogeneous data types in PyTables. See the Users’ Manual for more info on this.

New argument chunkshape of chunked leaves

It is possible now to specify the chunk shape for all the chunked leaves in PyTables (all except Array). With PyTables 1.x this value was automatically calculated so as to achieve decent results in most of the situations. However, the user may be interested in specifying its own chunk shape based on her own needs (although this should be done only by advanced users).

Of course, if this parameter is not specified, a sensible default is calculated for the size of the leave (which is recommended).

A new attribute called chunkshape has been added to all leaves. It is read-only (you can’t change the size of chunks once you have created a leaf), but it can be useful for inspection by advanced users.

New flavor specification

As of 2.x, flavors can only be set through the flavor attribute of leaves, and they are persistent, so changing a flavor requires that the file be writable.

Flavors can no longer be set through File.create*() methods, nor the flavor argument previously found in some Table methods, nor through Atom constructors or the _v_flavor attribute of descriptions.

System attributes can be deleted now

The protection against removing system attributes (like FILTERS, FLAVOR or CLASS, to name only a few) has been completely removed. It is now the responsibility of the user to make a proper use of this freedom. With this, users can get rid of all proprietary PyTables attributes if they want to (for example, for making a file to look more like an HDF5 native one).

Byteorder issues

Now, all the data coming from reads and internal buffers is always converted on-the-fly, if needed, to the native byteorder. This represents a big advantage in terms of speed when operating with objects coming from files that have been created in machines with a byte ordering different from native.

Besides, all leaf constructors have received a new byteorder parameter that allows specifying the byteorder of data on disk. In particular, a _v_byteorder entry in a Table description is no longer honored and you should use the aforementioned byteorder parameter.

Tunable internal buffer sizes

You can change the size of the internal buffers for I/O purposes of PyTables by changing the value of the new public attribute nrowsinbuf that is present in all leaves. By default, this contains a sensible value so as to achieve a good balance between speed and memory consumption. Be careful when changing it, if you don’t want to get unwanted results (very slow I/O, huge memory consumption…).

Changes to module names

If your application is directly accessing modules under the tables package, you need to know that the names of all modules are now all in lowercase. This allows one to tell apart the tables.Array class from the tables.array module (which was also called tables.Array before). This includes subpackages like tables.nodes.FileNode.

On top of that, more-or-less independent modules have also been renamed and some of them grouped into subpackages. The most important are:

  • The tables.netcdf3 subpackage replaces the old tables.NetCDF module.

  • The tables.nra subpackage replaces the old nestedrecords.py with the implementation of the NestedRecArray class.

Also, the tables.misc package includes utility modules which do not depend on PyTables.

Other changes

  • Filters.complib is None for filter properties created with complevel=0 (i.e. disabled compression, which is the default).

  • ‘non-relevant’ –> ‘irrelevant’ (applied to byteorders)

  • Table.colstypes –> Table.coltypes

  • Table.coltypes –> Table.coldtypes

  • Added Table.coldescr, dictionary of the Col descriptions.

  • Table.colshapes has disappeared. You can get it this way:

    colshapes = dict( (name, col.shape)
                      for (name, col) in table.coldescr.iteritems() )
  • Table.colitemsizes has disappeared. You can get it this way:

    colitemsizes = dict( (name, col.itemsize)
                         for (name, col) in table.coldescr.iteritems() )
  • Description._v_totalsize –> Description._v_itemsize

  • Description._v_itemsizes and Description._v_totalsizes have disappeared.

  • Leaf._v_chunksize –> Leaf.chunkshape

Enjoy data!

—The PyTables Team