This package is available from the Python Package Index. If you have pip you should be able to do:

$ pip install petl

You can also download manually, extract and run python install.

To verify the installation, the test suite can be run with nose, e.g.:

$ pip install nose
$ nosetests -v petl

petl has been tested with Python versions 2.7 and 3.4-3.6 under Linux and Windows operating systems.

Dependencies and extensions

This package is written in pure Python and has no installation requirements other than the Python core modules.

Some domain-specific and/or experimental extensions to petl are available from the petlx package.

Some of the functions in this package require installation of third party packages. These packages are indicated in the relevant parts of the documentation for each file format.

Also is possible to install some of dependencies when installing petl by specifying optional extra features, e.g.:

$ pip install petl['avro', 'interval', 'remote']

The available extra features are:


For using records from Databases with SQLAlchemy.

Note that is also required installing the package for the desired database.

For using Interval transformations with intervaltree
For using Avro files with fastavro
For using DataFrames with pandas
For using Arrays with numpy
For using Excel/LO files with xlrd/xlwt
For using Excel/LO files with openpyxl
For using XPath expressions with lxml
For using Bcolz ctables with bcolz
For using Text indexes with whoosh

For using HDF5 files with PyTables.

Note that also are additional software to be installed.


For reading and writing from Remote Sources with fsspec.

Note that fsspec also depends on other packages for providing support for each protocol as described in

Design goals

This package is designed primarily for convenience and ease of use, especially when working interactively with data that are unfamiliar, heterogeneous and/or of mixed quality.

petl transformation pipelines make minimal use of system memory and can scale to millions of rows if speed is not a priority. However if you are working with very large datasets and/or performance-critical applications then other packages may be more suitable, e.g., see pandas, pytables, bcolz and blaze. See also Related Work.

ETL pipelines

This package makes extensive use of lazy evaluation and iterators. This means, generally, that a pipeline will not actually be executed until data is requested.

E.g., given a file at ‘example.csv’ in the current working directory:

>>> example_data = """foo,bar,baz
... a,1,3.4
... b,2,7.4
... c,6,2.2
... d,9,8.1
... """
>>> with open('example.csv', 'w') as f:
...     f.write(example_data)

…the following code does not actually read the file or load any of its contents into memory:

>>> import petl as etl
>>> table1 = etl.fromcsv('example.csv')

Rather, table1 is a table container (see Conventions - table containers and table iterators below) which can be iterated over, extracting data from the underlying file on demand.

Similarly, if one or more transformation functions are applied, e.g.:

>>> table2 = etl.convert(table1, 'foo', 'upper')
>>> table3 = etl.convert(table2, 'bar', int)
>>> table4 = etl.convert(table3, 'baz', float)
>>> table5 = etl.addfield(table4, 'quux', lambda row: * row.baz)

…no actual transformation work will be done until data are requested from table5 (or any of the other tables returned by the intermediate steps).

So in effect, a 5 step pipeline has been set up, and rows will pass through the pipeline on demand, as they are pulled from the end of the pipeline via iteration.

A call to a function like petl.util.vis.look(), or any of the functions which write data to a file or database (e.g.,,,,, will pull data through the pipeline and cause all of the transformation steps to be executed on the requested rows, e.g.:

>>> etl.look(table5)
| foo | bar | baz | quux               |
| 'A' |   1 | 3.4 |                3.4 |
| 'B' |   2 | 7.4 |               14.8 |
| 'C' |   6 | 2.2 | 13.200000000000001 |
| 'D' |   9 | 8.1 |  72.89999999999999 |

…although note that petl.util.vis.look() will by default only request the first 5 rows, and so the minimum amount of processing will be done to produce 5 rows.

Functional and object-oriented programming styles

The petl package supports both functional and object-oriented programming styles. For example, the example in the section on ETL pipelines above could also be written as:

>>> import petl as etl
>>> table = (
...     etl
...     .fromcsv('example.csv')
...     .convert('foo', 'upper')
...     .convert('bar', int)
...     .convert('baz', float)
...     .addfield('quux', lambda row: * row.baz)
... )
>>> table.look()
| foo | bar | baz | quux               |
| 'A' |   1 | 3.4 |                3.4 |
| 'B' |   2 | 7.4 |               14.8 |
| 'C' |   6 | 2.2 | 13.200000000000001 |
| 'D' |   9 | 8.1 |  72.89999999999999 |

A wrap() function is also provided to use the object-oriented style with any valid table container object, e.g.:

>>> l = [['foo', 'bar'], ['a', 1], ['b', 2], ['c', 2]]
>>> table = etl.wrap(l)
>>> table.look()
| foo | bar |
| 'a' |   1 |
| 'b' |   2 |
| 'c' |   2 |

Interactive use

When using petl from within an interactive Python session, the default representation for table objects uses the petl.util.vis.look() function, so a table object can be returned at the prompt to inspect it, e.g.:

>>> l = [['foo', 'bar'], ['a', 1], ['b', 2], ['c', 2]]
>>> table = etl.wrap(l)
>>> table
| foo | bar |
| 'a' |   1 |
| 'b' |   2 |
| 'c' |   2 |

By default data values are rendered using the built-in repr() function. To see the string (str()) values instead, print() the table, e.g.:

>>> print(table)
| foo | bar |
| a   |   1 |
| b   |   2 |
| c   |   2 |

IPython notebook integration

Table objects also implement _repr_html_() and so will be displayed as an HTML table if returned from a cell in an IPython notebook. The functions petl.util.vis.display() and petl.util.vis.displayall() also provide more control over rendering of tables within an IPython notebook.

For examples of usage see the repr_html notebook.

petl executable

Also included in the petl distribution is a script to execute simple transformation pipelines directly from the operating system shell. E.g.:

$ petl "dummytable().tocsv()" > example.csv
$ cat example.csv | petl "fromcsv().cut('foo', 'baz').convert('baz', float).selectgt('baz', 0.5).head().data().totsv()"

The petl script is extremely simple, it expects a single positional argument, which is evaluated as Python code but with all of the functions in the petl namespace imported.

Conventions - table containers and table iterators

This package defines the following convention for objects acting as containers of tabular data and supporting row-oriented iteration over the data.

A table container (also referred to here as a table) is any object which satisfies the following:

  1. implements the __iter__ method
  2. __iter__ returns a table iterator (see below)
  3. all table iterators returned by __iter__ are independent, i.e., consuming items from one iterator will not affect any other iterators

A table iterator is an iterator which satisfies the following:

  1. each item returned by the iterator is a sequence (e.g., tuple or list)
  2. the first item returned by the iterator is a header row comprising a sequence of header values
  3. each subsequent item returned by the iterator is a data row comprising a sequence of data values
  4. a header value is typically a string (str) but may be an object of any type as long as it implements __str__ and is pickleable
  5. a data value is any pickleable object

So, for example, a list of lists is a valid table container:

>>> table = [['foo', 'bar'], ['a', 1], ['b', 2]]

Note that an object returned by the csv.reader() function from the standard Python csv module is a table iterator and not a table container, because it can only be iterated over once. However, it is straightforward to define functions that support the table container convention and provide access to data from CSV or other types of file or data source, see e.g. the function.

The main reason for requiring that table containers support independent table iterators (point 3) is that data from a table may need to be iterated over several times within the same program or interactive session. E.g., when using petl in an interactive session to build up a sequence of data transformation steps, the user might want to examine outputs from several intermediate steps, before all of the steps are defined and the transformation is executed in full.

Note that this convention does not place any restrictions on the lengths of header and data rows. A table may contain a header row and/or data rows of varying lengths.

Extensions - integrating custom data sources

The module has functions for extracting data from a number of well-known data sources. However, it is also straightforward to write an extension that enables integration with other data sources. For an object to be usable as a petl table it has to implement the table container convention described above. Below is the source code for an ArrayView class which allows integration of petl with numpy arrays. This class is included within the module but also provides an example of how other data sources might be integrated:

>>> import petl as etl
>>> class ArrayView(etl.Table):
...     def __init__(self, a):
...         # assume that a is a numpy array
...         self.a = a
...     def __iter__(self):
...         # yield the header row
...         header = tuple(self.a.dtype.names)
...         yield header
...         # yield the data rows
...         for row in self.a:
...             yield tuple(row)

Now this class enables the use of numpy arrays with petl functions, e.g.:

>>> import numpy as np
>>> a = np.array([('apples', 1, 2.5),
...               ('oranges', 3, 4.4),
...               ('pears', 7, 0.1)],
...              dtype='U8, i4,f4')
>>> t1 = ArrayView(a)
>>> t1
| f0        | f1 | f2        |
| 'apples'  | 1  | 2.5       |
| 'oranges' | 3  | 4.4000001 |
| 'pears'   | 7  | 0.1       |

>>> t2 = t1.cut('f0', 'f2').convert('f0', 'upper').addfield('f3', lambda row: row.f2 * 2)
>>> t2
| f0        | f2        | f3                  |
| 'APPLES'  | 2.5       |                 5.0 |
| 'ORANGES' | 4.4000001 |  8.8000001907348633 |
| 'PEARS'   | 0.1       | 0.20000000298023224 |

If you develop an extension for a data source that you think would also be useful for others, please feel free to submit a PR to the petl GitHub repository, or if it is a domain-specific data source, the petlx GitHub repository.


This package tries to make efficient use of memory by using iterators and lazy evaluation where possible. However, some transformations cannot be done without building data structures, either in memory or on disk.

An example is the petl.transform.sorts.sort() function, which will either sort a table entirely in memory, or will sort the table in memory in chunks, writing chunks to disk and performing a final merge sort on the chunks. Which strategy is used will depend on the arguments passed into the petl.transform.sorts.sort() function when it is called.

In either case, the sorting can take some time, and if the sorted data will be used more than once, it is undesirable to start again from scratch each time. It is better to cache the sorted data, if possible, so it can be re-used.

The petl.transform.sorts.sort() function, and all functions which use it internally, provide a cache keyword argument which can be used to turn on or off the caching of sorted data.

There is also an explicit petl.util.materialise.cache() function, which can be used to cache in memory up to a configurable number of rows from any table.