Unicode

Since early Python 2 days unicode was part of all default Python builds. It allows developers to write applications that deal with non-ASCII characters in a straightforward way. But working with unicode requires a basic knowledge about that matter, especially when working with libraries that do not support it.

Werkzeug uses unicode internally everywhere text data is assumed, even if the HTTP standard is not unicode aware as it. Basically all incoming data is decoded from the charset specified (per default utf-8) so that you don’t operate on bytestrings any more. Outgoing unicode data is then encoded into the target charset again.

Unicode in Python

In Python 2 there are two basic string types: str and unicode. str may carry encoded unicode data but it’s always represented in bytes whereas the unicode type does not contain bytes but charpoints. What does this mean? Imagine you have the German Umlaut ö. In ASCII you cannot represent that character, but in the latin-1 and utf-8 character sets you can represent it, but they look differently when encoded:

>>> u'ö'.encode('latin1')
'\xf6'
>>> u'ö'.encode('utf-8')
'\xc3\xb6'

So an ö might look totally different depending on the encoding which makes it hard to work with it. The solution is using the unicode type (as we did above, note the u prefix before the string). The unicode type does not store the bytes for ö but the information, that this is a LATIN SMALL LETTER O WITH DIAERESIS.

Doing len(u'ö') will always give us the expected “1” but len('ö') might give different results depending on the encoding of 'ö'.

Unicode in HTTP

The problem with unicode is that HTTP does not know what unicode is. HTTP is limited to bytes but this is not a big problem as Werkzeug decodes and encodes for us automatically all incoming and outgoing data. Basically what this means is that data sent from the browser to the web application is per default decoded from an utf-8 bytestring into a unicode string. Data sent from the application back to the browser that is not yet a bytestring is then encoded back to utf-8.

Usually this “just works” and we don’t have to worry about it, but there are situations where this behavior is problematic. For example the Python 2 IO layer is not unicode aware. This means that whenever you work with data from the file system you have to properly decode it. The correct way to load a text file from the file system looks like this:

f = file('/path/to/the_file.txt', 'r')
try:
    text = f.decode('utf-8')    # assuming the file is utf-8 encoded
finally:
    f.close()

There is also the codecs module which provides an open function that decodes automatically from the given encoding.

Error Handling

With Werkzeug 0.3 onwards you can further control the way Werkzeug works with unicode. In the past Werkzeug ignored encoding errors silently on incoming data. This decision was made to avoid internal server errors if the user tampered with the submitted data. However there are situations where you want to abort with a 400 BAD REQUEST instead of silently ignoring the error.

All the functions that do internal decoding now accept an errors keyword argument that behaves like the errors parameter of the builtin string method decode. The following values are possible:

ignore

This is the default behavior and tells the codec to ignore characters that it doesn’t understand silently.

replace

The codec will replace unknown characters with a replacement character (U+FFFD REPLACEMENT CHARACTER)

strict

Raise an exception if decoding fails.

Unlike the regular python decoding Werkzeug does not raise an UnicodeDecodeError if the decoding failed but an HTTPUnicodeError which is a direct subclass of UnicodeError and the BadRequest HTTP exception. The reason is that if this exception is not caught by the application but a catch-all for HTTP exceptions exists a default 400 BAD REQUEST error page is displayed.

There is additional error handling available which is a Werkzeug extension to the regular codec error handling which is called fallback. Often you want to use utf-8 but support latin1 as legacy encoding too if decoding failed. For this case you can use the fallback error handling. For example you can specify 'fallback:iso-8859-15' to tell Werkzeug it should try with iso-8859-15 if utf-8 failed. If this decoding fails too (which should not happen for most legacy charsets such as iso-8859-15) the error is silently ignored as if the error handling was ignore.

Further details are available as part of the API documentation of the concrete implementations of the functions or classes working with unicode.

Request and Response Objects

As request and response objects usually are the central entities of Werkzeug powered applications you can change the default encoding Werkzeug operates on by subclassing these two classes. For example you can easily set the application to utf-7 and strict error handling:

from werkzeug.wrappers import BaseRequest, BaseResponse

class Request(BaseRequest):
    charset = 'utf-7'
    encoding_errors = 'strict'

class Response(BaseResponse):
    charset = 'utf-7'

Keep in mind that the error handling is only customizable for all decoding but not encoding. If Werkzeug encounters an encoding error it will raise a UnicodeEncodeError. It’s your responsibility to not create data that is not present in the target charset (a non issue with all unicode encodings such as utf-8).