What's new in lxml 2.0?
During the development of the lxml 1.x series, a couple of quirks were
discovered in the design that made the API less obvious and its future
extensions harder than necessary. lxml 2.0 is a soft evolution of lxml 1.x
towards a simpler, more consistent and more powerful API - with some major
extensions. Wherever possible, lxml 1.3 comes close to the semantics of lxml
2.0, so that migrating should be easier for code that currently runs with 1.3.
One of the important internal changes was the switch from the Pyrex
compiler to Cython, which provides better optimisation and improved
support for newer Python language features. This allows the code of
lxml to become more Python-like again, while the performance improves
as Cython continues its own development. The code simplification,
which will continue throughout the 2.x series, will hopefully make it
even easier for users to contribute.
A graduation towards a more consistent API cannot go without a certain amount
of incompatible changes. The following is a list of those differences that
applications need to take into account when migrating from lxml 1.x to lxml
- lxml 0.9 introduced a feature called namespace implementation. The
global Namespace factory was added to register custom element classes
and have lxml.etree look them up automatically. However, the later
development of further class lookup mechanisms made it appear less and less
adequate to register this mapping at a global level, so lxml 1.1 first
removed the namespace based lookup from the default setup and lxml 2.0
finally removes the global namespace registry completely. As all other
lookup mechanisms, the namespace lookup is now local to a parser, including
the registry itself. Applications that use a module-level parser can easily
map its get_namespace() method to a global Namespace function to
mimic the old behaviour.
- Some API functions now require passing options as keyword arguments,
as opposed to positional arguments. This restriction was introduced
to make the API usage independent of future extensions such as the
addition of new positional arguments. Users should not rely on the
position of an optional argument in function signatures and instead
pass it explicitly named. This also improves code readability - it
is common good practice to pass options in a consistent way
independent of their position, so many people may not even notice
the change in their code. Another important reason is compatibility
with cElementTree, which also enforces keyword-only arguments in a
couple of places.
- XML tag names are validated when creating an Element. This does not
apply to HTML tags, where only HTML special characters are
forbidden. The distinction is made by the SubElement() factory,
which tests if the tree it works on is an HTML tree, and by the
.makeelement() methods of parsers, which behave differently for
the XMLParser() and the HTMLParser().
- XPath now raises exceptions specific to the part of the execution that
failed: XPathSyntaxError for parser errors and XPathEvalError for
errors that occurred during the evaluation. Note that the distinction only
works for the XPath() class. The other two evaluators only have a
single evaluation call that includes the parsing step, and will therefore
only raise an XPathEvalError. Applications can catch both exceptions
through the common base class XPathError (which also exists in earlier
- Network access in parsers is now disabled by default, i.e. the
no_network option defaults to True. Due to a somewhat 'interesting'
implementation in libxml2, this does not affect the first document (i.e. the
URL that is parsed), but only subsequent documents, such as a DTD when
parsing with validation. This means that you will have to check the URL you
pass, instead of relying on lxml to prevent any access to external
resources. As this can be helpful in some use cases, lxml does not work
- The type annotations in lxml.objectify (the pytype attribute) now use
NoneType for the None value as this is the correct Python type name.
Previously, lxml 1.x used a lower case none.
- Another change in objectify regards the way it deals with ambiguous types.
Previously, setting a value like the string "3" through normal attribute
access would let it come back as an integer when reading the object
attribute. lxml 2.0 prevents this by always setting the pytype
attribute to the type the user passed in, so "3" will come back as a
string, while the number 3 will come back as a number. To remove the
type annotation on serialisation, you can use the deannotate() function.
- The C-API function findOrBuildNodeNs() was replaced by the more generic
findOrBuildNodeNsPrefix() that accepts an additional default prefix.
Most of the enhancements of lxml 2.0 were made under the hood. Most people
won't even notice them, but they make the maintenance of lxml easier and thus
facilitate further enhancements and an improved integration between lxml's
- lxml.objectify now has its own implementation of the E factory. It uses
the built-in type lookup mechanism of lxml.objectify, thus removing the need
for an additional type registry mechanism (as previously available through
the typemap parameter).
- XML entities are supported through the Entity() factory, an Entity
element class and a parser option resolve_entities that allows to keep
entities in the element tree when set to False. Also, the parser will now
report undefined entities as errors if it needs to resolve them (which is
still the default, as in lxml 1.x).
- A major part of the XPath code was rewritten and can now benefit from a
bigger overlap with the XSLT code. The main benefits are improved thread
safety in the XPath evaluators and Python RegExp support in standard XPath.
- The string results of an XPath evaluation have become 'smart' string
subclasses. Formerly, there was no easy way to find out where a
string originated from. In lxml 2.0, you can call its
getparent() method to find the Element that carries it. This
works for attributes (//@attribute) and for text() nodes,
i.e. Element text and tails. Strings that were constructed in the
path expression, e.g. by the string() function or extension
functions, will return None as their parent.
- Setting a QName object as value of the .text property or as
an attribute value will resolve its prefix in the respective context
- Following ElementTree 1.3, the iterfind() method supports
efficient iteration based on XPath-like expressions.
The parsers also received some major enhancements:
- iterparse() can parse HTML when passing the boolean html
- Parse time XML Schema validation by passing an
XMLSchema object to the schema keyword argument of a parser.
- Support for a target object that implements ElementTree's
- The encoding keyword allows overriding the document encoding.
The following functions and methods are now deprecated. They are
still available in lxml 2.0 and will be removed in lxml 2.1:
- The tounicode() function was replaced by the call
- CamelCaseNamed module functions and methods were renamed to their
underscore equivalents to follow PEP 8 in naming.
- etree.setDefaultParser() -> etree.set_default_parser()
- etree.getDefaultParser() -> etree.get_default_parser()
- etree.useGlobalPythonLog() -> etree.use_global_python_log()
- XMLParser.setElementClassLookup() -> .set_element_class_lookup()
- HTMLParser.setElementClassLookup() -> .set_element_class_lookup()
- The .getiterator() method on Elements and ElementTrees was
renamed to .iter() to follow ElementTree 1.3.
The most visible changes in lxml 2.0 regard the new modules that were added.
A very useful module for doctests based on XML or HTML is
lxml.doctestcompare. It provides a relaxed comparison mechanism
for XML and HTML in doctests. Using it for XML comparisons is as
>>> import lxml.usedoctest
and for HTML comparisons:
>>> import lxml.html.usedoctest
The largest new package that was added to lxml 2.0 is lxml.html. It
contains various tools and modules for HTML handling. The major features
include support for cleaning up HTML (removing unwanted content), a readable
HTML diff and various tools for working with links.
The Cascading Stylesheet Language (CSS) has a very short and generic path
language for pointing at elements in XML/HTML trees (CSS selectors). The module
lxml.cssselect provides an implementation based on XPath.