General questions

What is PyTables?

PyTables is a package for managing hierarchical datasets designed to efficiently cope with extremely large amounts of data.

It is built on top of the HDF5 [1] library, the Python language [2] and the NumPy [3] package. It features an object-oriented interface that, combined with C extensions for the performance-critical parts of the code, makes it a fast yet extremely easy-to-use tool for interactively storing and retrieving very large amounts of data.

What are PyTables’ licensing terms?

PyTables is free for both commercial and non-commercial use, under the terms of the BSD license.

I’m having problems. How can I get support?

The most common and efficient way is to subscribe (remember you need to subscribe prior to send messages) to the PyTables users mailing list [4], and send there a brief description of your issue and, if possible, a short script that can reproduce it. Hopefully, someone on the list will be able to help you. It is also a good idea to check out the archives of the user’s list [5] (you may want to check the Gmane archives [6] instead) so as to see if the answer to your question has already been dealed with.

Why HDF5?

HDF5 [1] is the underlying C library and file format that enables PyTables to efficiently deal with the data. It has been chosen for the following reasons:

  • Designed to efficiently manage very large datasets.
  • Lets you organize datasets hierarchically.
  • Very flexible and well tested in scientific environments.
  • Good maintenance and improvement rate.
  • Technical excellence (R&D 100 Award [7]).
  • It’s Open Source software

Why Python?

  1. Python is interactive.

    People familiar with data processing understand how powerful command line interfaces are for exploring mathematical relationships and scientific data sets. Python provides an interactive environment with the added benefit of a full featured programming language behind it.

  2. Python is productive for beginners and experts alike.

    PyTables is targeted at engineers, scientists, system analysts, financial analysts, and others who consider programming a necessary evil. Any time spent learning a language or tracking down bugs is time spent not solving their real problem. Python has a short learning curve and most people can do real and useful work with it in a day of learning. Its clean syntax and interactive nature facilitate this.

  3. Python is data-handling friendly.

    Python comes with nice idioms that make the access to data much easier: general slicing (i.e. data[start:stop:step]), list comprehensions, iterators, generators ... are constructs that make the interaction with your data very easy.

Why NumPy?

NumPy [3] is a Python package to efficiently deal with large datasets in-memory, providing containers for homogeneous data, heterogeneous data, and string arrays. PyTables uses these NumPy containers as in-memory buffers to push the I/O bandwith towards the platform limits.

Where can PyTables be applied?

In all the scenarios where one needs to deal with large datasets:

  • Industrial applications
    • Data acquisition in real time
    • Quality control
    • Fast data processing
  • Scientific applications
    • Meteorology, oceanography
    • Numerical simulations
    • Medicine (biological sensors, general data gathering & processing)
  • Information systems
    • System log monitoring & consolidation
    • Tracing of routing data
    • Alert systems in security

Is PyTables safe?

Well, first of all, let me state that PyTables does not support transactional features yet (we don’t even know if we will ever be motivated to implement this!), so there is always the risk that you can lose your data in case of an unexpected event while writing (like a power outage, system shutdowns ...). Having said that, if your typical scenarios are write once, read many, then the use of PyTables is perfectly safe, even for dealing extremely large amounts of data.

Can PyTables be used in concurrent access scenarios?

It depends. Concurrent reads are no problem at all. However, whenever a process (or thread) is trying to write, then problems will start to appear. First, PyTables doesn’t support locking at any level, so several process writing concurrently to the same PyTables file will probably end up corrupting it, so don’t do this! Even having only one process writing and the others reading is a hairy thing, because the reading processes might be reading incomplete data from a concurrent data writing operation.

The solution would be to lock the file while writing and unlock it after a flush over the file has been performed. Also, in order to avoid cache (HDF5 [1], PyTables) problems with read apps, you would need to re-open your files whenever you are going to issue a read operation. If a re-opening operation is unacceptable in terms of speed, you may want to do all your I/O operations in one single process (or thread) and communicate the results via sockets, Queue.Queue objects (in case of using threads), or whatever, with the client process/thread.

The examples directory contains two scripts demonstrating methods of accessing a PyTables file from multiple processes.

The first, multiprocess_access_queues.py, uses a multiprocessing.Queue object to transfer read and write requests from multiple DataProcessor processes to a single process responsible for all access to the PyTables file. The results of read requests are then transferred back to the originating processes using other Queue objects.

The second example script, multiprocess_access_benchmarks.py, demonstrates and benchmarks four methods of transferring PyTables array data between processes. The four methods are:

  • Using multiprocessing.Pipe from the Python standard library.
  • Using a memory mapped file that is shared between two processes. The NumPy array associated with the file is passed as the out argument to the tables.Array.read() method.
  • Using a Unix domain socket. Note that this example uses the ‘abstract namespace’ and will only work under Linux.
  • Using an IPv4 socket.

What kind of containers does PyTables implement?

PyTables does support a series of data containers that address specific needs of the user. Below is a brief description of them:

Lets you deal with heterogeneous datasets. Allows compression. Enlargeable. Supports nested types. Good performance for read/writing data.
Provides quick and dirty array handling. Not compression allowed. Not enlargeable. Can be used only with relatively small datasets (i.e. those that fit in memory). It provides the fastest I/O speed.
Provides compressed array support. Not enlargeable. Good speed when reading/writing.
Most general array support. Compressible and enlargeable. It is pretty fast at extending, and very good at reading.
Supports collections of homogeneous data with a variable number of entries. Compressible and enlargeable. I/O is not very fast.
The structural component. A hierarchically-addressable container for HDF5 nodes (each of these containers, including Group, are nodes), similar to a directory in a UNIX filesystem.

Please refer to the Library Reference for more specific information.

Cool! I’d like to see some examples of use.

Sure. Go to the HowToUse section to find simple examples that will help you getting started.

Can you show me some screenshots?

Well, PyTables is not a graphical library by itself. However, you may want to check out ViTables [8], a GUI tool to browse and edit PyTables & HDF5 [1] files.

Is PyTables a replacement for a relational database?

No, by no means. PyTables lacks many features that are standard in most relational databases. In particular, it does not have support for relationships (beyond the hierarchical one, of course) between datasets and it does not have transactional features. PyTables is more focused on speed and dealing with really large datasets, than implementing the above features. In that sense, PyTables can be best viewed as a teammate of a relational database.

For example, if you have very large tables in your existing relational database, they will take lots of space on disk, potentially reducing the performance of the relational engine. In such a case, you can move those huge tables out of your existing relational database to PyTables, and let your relational engine do what it does best (i.e. manage relatively small or medium datasets with potentially complex relationships), and use PyTables for what it has been designed for (i.e. manage large amounts of data which are loosely related).

How can PyTables be fast if it is written in an interpreted language like Python?

Actually, all of the critical I/O code in PyTables is a thin layer of code on top of HDF5 [1], which is a very efficient C library. Cython [9] is used as the glue language to generate “wrappers” around HDF5 calls so that they can be used in Python. Also, the use of an efficient numerical package such as NumPy [3] makes the most costly operations effectively run at C speed. Finally, time-critical loops are usually implemented in Cython [9] (which, if used properly, allows to generate code that runs at almost pure C speeds).

If it is designed to deal with very large datasets, then PyTables should consume a lot of memory, shouldn’t it?

Well, you already know that PyTables sits on top of HDF5, Python and NumPy [3], and if we add its own logic (~7500 lines of code in Python, ~3000 in Cython and ~4000 in C), then we should conclude that PyTables isn’t effectively a paradigm of lightness.

Having said that, PyTables (as HDF5 [1] itself) tries very hard to optimize the memory consumption by implementing a series of features like dynamic determination of buffer sizes, Least Recently Used cache for keeping unused nodes out of memory, and extensive use of compact NumPy [3] data containers. Moreover, PyTables is in a relatively mature state and most memory leaks have been already addressed and fixed.

Just to give you an idea of what you can expect, a PyTables program can deal with a table with around 30 columns and 1 million entries using as low as 13 MB of memory (on a 32-bit platform). All in all, it is not that much, is it?.

Why was PyTables born?

Because, back in August 2002, one of its authors (Francesc Alted [10]) had a need to save lots of hierarchical data in an efficient way for later post-processing it. After trying out several approaches, he found that they presented distinct inconveniences. For example, working with file sizes larger than, say, 100 MB, was rather painful with ZODB (it took lots of memory with the version available by that time).

The netCDF3 [11] interface provided by Scientific Python [12] was great, but it did not allow to structure the hierarchically; besides, netCDF3 [11] only supports homogeneous datasets, not heterogeneous ones (i.e. tables). (As an aside, netCDF4 [11] overcomes many of the limitations of netCDF3 [11], although curiously enough, it is based on top of HDF5 [1], the library chosen as the base for PyTables from the very beginning.)

So, he decided to give HDF5 [1] a try, start doing his own wrappings to it and voilà, this is how the first public release of PyTables (0.1) saw the light in October 2002, three months after his itch started to eat him ;-).

Does PyTables have a client-server interface?

Not by itself, but you may be interested in using PyTables through pydap [13], a Python implementation of the OPeNDAP [14] protocol. Have a look at the PyTables plugin of pydap [13].

How does PyTables compare with the h5py project?

Well, they are similar in that both packages are Python interfaces to the HDF5 [1] library, but there are some important differences to be noted. h5py [16] is an attempt to map the HDF5 [1] feature set to NumPy [3] as closely as possible. In addition, it also provides access to nearly all of the HDF5 [1] C API.

Instead, PyTables builds up an additional abstraction layer on top of HDF5 [1] and NumPy [3] where it implements things like an enhanced type system, an engine for enabling complex queries, an efficient computational kernel [17], advanced indexing capabilities [18] or an undo/redo feature, to name just a few. This additional layer also allows PyTables to be relatively independent of its underlying libraries (and their possible limitations). For example, PyTables can support HDF5 [1] data types like enumerated or time that are available in the HDF5 [1] library but not in the NumPy [3] package; or even perform powerful complex queries that are not implemented directly in neither HDF5 [1] nor NumPy [3].

Furthermore, PyTables also tries hard to be a high performance interface to HDF5/NumPy, implementing niceties like internal LRU caches for nodes and other data and metadata, automatic computation of optimal chunk sizes for the datasets, a variety of compressors, ranging from slow but efficient (bzip2 [19]) to extremely fast ones (Blosc [20]) in addition to the standard zlib [21]. Another difference is that PyTables makes use of numexpr [22] so as to accelerate internal computations (for example, in evaluating complex queries) to a maximum.

For contrasting with other opinions, you may want to check the PyTables/h5py comparison in a similar entry of the FAQ of h5py [23].

I’ve found a bug. What do I do?

The PyTables development team works hard to make this eventuality as rare as possible, but, as in any software made by human beings, bugs do occur. If you find any bug, please tell us by file a bug report in the issue tracker [24] on GitHub [25].

Is it possible to get involved in PyTables development?

Indeed. We are keen for more people to help out contributing code, unit tests, documentation, and helping out maintaining this wiki. Drop us a mail on the users mailing list and tell us in which area do you want to work.

How can I cite PyTables?

The recommended way to cite PyTables in a paper or a presentation is as following:

  • Author: Francesc Alted, Ivan Vilata and others
  • Title: PyTables: Hierarchical Datasets in Python
  • Year: 2002 -
  • URL: http://www.pytables.org

Here’s an example of a BibTeX entry:

  author =    {Francesc Alted and Ivan Vilata and others},
  title =     {{PyTables}: Hierarchical Datasets in {Python}},
  year =      {2002--},
  url = "http://www.pytables.org/"

PyTables 2.x issues

I’m having problems migrating my apps from PyTables 1.x into PyTables 2.x. Please, help!

Sure. However, you should first check out the Migrating from PyTables 1.x to 2.x document. It should provide hints to the most frequently asked questions on this regard.

For combined searches like table.where(‘(x<5) & (x>3)’), why was a & operator chosen instead of an and?

Search expressions are in fact Python expressions written as strings, and they are evaluated as such. This has the advantage of not having to learn a new syntax, but it also implies some limitations with logical and and or operators, namely that they can not be overloaded in Python. Thus, it is impossible right now to get an element-wise operation out of an expression like ‘array1 and array2’. That’s why one has to choose some other operator, being & and | the most similar to their C counterparts && and ||, which aren’t available in Python either.

You should be careful about expressions like ‘x<5 & x>3’ and others like ‘3 < x < 5’ which ‘’won’t work as expected’‘, because of the different operator precedence and the absence of an overloaded logical and operator. More on this in the appendix about condition syntax in the HDF5 manual [26].

There are quite a few packages affected by those limitations including NumPy [3] themselves and SQLObject [27], and there have been quite longish discussions about adding the possibility of overloading logical operators to Python (see PEP 335 [28] and this thread [30] for more details).

I can not select rows using in-kernel queries with a condition that involves an UInt64Col. Why?

This turns out to be a limitation of the numexpr [22] package. Internally, numexpr [22] uses a limited set of types for doing calculations, and unsigned integers are always upcasted to the immediate signed integer that can fit the information. The problem here is that there is not a (standard) signed integer that can be used to keep the information of a 64-bit unsigned integer.

So, your best bet right now is to avoid uint64 types if you can. If you absolutely need uint64, the only way for doing selections with this is through regular Python selections. For example, if your table has a colM column which is declared as an UInt64Col, then you can still filter its values with:

[row['colN'] for row in table if row['colM'] < X]

However, this approach will generally lead to slow speed (specially on Win32 platforms, where the values will be converted to Python long values).

I’m already using PyTables 2.x but I’m still getting numarray objects instead of NumPy ones!

This is most probably due to the fact that you are using a file created with PyTables 1.x series. By default, PyTables 1.x was setting an HDF5 attribute FLAVOR with the value ‘numarray’ to all leaves. Now, PyTables 2.x sees this attribute and obediently converts the internal object (truly a NumPy object) into a numarray one. For PyTables 2.x files the FLAVOR attribute will only be saved when explicitly set via the leaf.flavor property (or when passing data to an Array or Table at creation time), so you will be able to distinguish default flavors from user-set ones by checking the existence of the FLAVOR attribute.

Meanwhile, if you don’t want to receive numarray objects when reading old files, you have several possibilities:

  • Remove the flavor for your datasets by hand:

    for leaf in h5file.walkNodes(classname='Leaf'):
        del leaf.flavor
  • Use the :program:’ptrepack` utility with the flag --upgrade-flavors so as to convert all flavors in old files to the default (effectively by removing the FLAVOR attribute).

  • Remove the numarray (and/or Numeric) package from your system. Then PyTables 2.x will return you pure NumPy objects (it can’t be otherwise!).

Installation issues


Error when importing tables

You have installed the binary installer for Windows and, when importing the tables package you are getting an error like:

The command in "0x6714a822" refers to memory in "0x012011a0". The
procedure "written" could not be executed.
Click to ok to terminate.
Click to abort to debug the program.

This problem can be due to a series of reasons, but the most probable one is that you have a version of a DLL library that is needed by PyTables and it is not at the correct version. Please, double-check the versions of the required libraries for PyTables and install newer versions, if needed. In most cases, this solves the issue.

In case you continue getting problems, there are situations where other programs do install libraries in the PATH that are optional to PyTables (for example BZIP2 or LZO), but that they will be used if they are found in your system (i.e. anywhere in your PATH). So, if you find any of these libraries in your PATH, upgrade it to the latest version available (you don’t need to re-install PyTables).

Can’t find LZO binaries for Windows

Unfortunately, the LZO binaries for Windows seems to be unavailable from its usual place at http://gnuwin32.sourceforge.net/packages/lzo.htm. So, in order to allow people to be able to install this excellent compressor easily, we have packaged the LZO binaries in a zip file available at: http://www.pytables.org/download/lzo-win. This zip file follows the same structure that a typical GnuWin32 [29] package, so it is just a matter of unpacking it in your GNUWIN32 directory and following the instructions in the PyTables Manual [15].

Hopefully somebody else will take care again of maintaining LZO for Windows again.

Testing issues

Tests fail when running from IPython

You may be getting errors related with Doctest when running the test suite from IPython. This is a known limitation in IPython (see http://lists.ipython.scipy.org/pipermail/ipython-dev/2007-April/002859.html). Try running the test suite from the vanilla Python interpreter instead.

Tests fail when running from Python 2.5 and Numeric is installed

Numeric doesn’t get well with Python 2.5, even on 32-bit platforms. This is a consequence of Numeric not being maintained anymore and you should consider migrating to NumPy as soon as possible. To get rid of these errors, just uninstall Numeric.

[1](1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15) http://www.hdfgroup.org/HDF5
[3](1, 2, 3, 4, 5, 6, 7, 8, 9, 10) http://www.numpy.org
[9](1, 2) http://www.cython.org
[11](1, 2, 3, 4) http://www.unidata.ucar.edu/software/netcdf
[13](1, 2) http://www.pydap.org
[22](1, 2, 3) https://github.com/pydata/numexpr