Designing Unicode Aware APIs

APIs that deal with byte str and unicode strings are difficult to get right. Here are a few strategies with pros and cons of each.

Contents

• Designing Unicode Aware APIs
  • Take either bytes or unicode, output only unicode
  • Take either bytes or unicode, output the same type
  • Separate functions
  • Deciding whether to take str or unicode when no value is returned
    • Writing to external data
    • Updating data structures
  • APIs to Avoid
    • Returning unicode unless a conversion fails
    • Ignoring values with no chance of recovery
    • Raising a UnicodeException with no chance of recovery
  • Knowing your data
    • Do you need to operate on both bytes and unicode?
    • Can you restrict the encodings?
      • Single byte encodings
      • Multibyte encodings
        • Fixed width
        • Variable Width
          • ASCII compatible
          • Escaped
          • Other

Take either bytes or unicode, output only unicode

In this strategy, you allow the user to enter either unicode strings or byte str but what you give back is always unicode. This strategy is easy for novice endusers to start using immediately as they will be able to feed either type of string into the function and get back a string that they can use in other places.

However, it does lead to the novice writing code that functions correctly when testing it with ASCII-only data but fails when given data that contains non-ASCII characters. Worse, if your API is not designed to be flexible, the consumer of your code won’t be able to easily correct those problems once they find them.

Here’s a good API that uses this strategy:

from kitchen.text.converters import to_unicode

def truncate(msg, max_length, encoding='utf8', errors='replace'):
    msg = to_unicode(msg, encoding, errors)
    return msg[:max_length]

The call to truncate() starts with the essential parameters for performing the task. It ends with two optional keyword arguments that define the encoding to use to transform from a byte str to unicode and the strategy to use if undecodable bytes are encountered. The defaults may vary depending on the use cases you have in mind. When the output is generally going to be printed for the user to see, errors='replace' is a good default. If you are constructing keys to a database, raisng an exception (with errors='strict') may be a better default. In either case, having both parameters allows the person using your API to choose how they want to handle any problems. Having the values is also a clue to them that a conversion from byte str to unicode string is going to occur.

Note

If you’re targeting python-3.1 and above, errors='surrogateescape' may be a better default than errors='strict'. You need to be mindful of a few things when using surrogateescape though:

• surrogateescape will cause issues if a non-ASCII compatible encoding is used (for instance, UTF-16 and UTF-32.) That makes it unhelpful in situations where a true general purpose method of encoding must be found. PEP 383 mentions that surrogateescape was specifically designed with the limitations of translating using system locales (where ASCII compatibility is generally seen as inescapable) so you should keep that in mind.
• If you use surrogateescape to decode from bytes to unicode you will need to use an error handler other than strict to encode as the lone surrogate that this error handler creates makes for invalid unicode that must be handled when encoding. In Python-3.1.2 or less, a bug in the encoder error handlers mean that you can only use surrogateescape to encode; anything else will throw an error.

Evaluate your usages of the variables in question to see what makes sense.

Here’s a bad example of using this strategy:

from kitchen.text.converters import to_unicode

def truncate(msg, max_length):
    msg = to_unicode(msg)
    return msg[:max_length]

In this example, we don’t have the optional keyword arguments for encoding and errors. A user who uses this function is more likely to miss the fact that a conversion from byte str to unicode is going to occur. And once an error is reported, they will have to look through their backtrace and think harder about where they want to transform their data into unicode strings instead of having the opportunity to control how the conversion takes place in the function itself. Note that the user does have the ability to make this work by making the transformation to unicode themselves:

from kitchen.text.converters import to_unicode

msg = to_unicode(msg, encoding='euc_jp', errors='ignore')
new_msg = truncate(msg, 5)

Take either bytes or unicode, output the same type

This strategy is sometimes called polymorphic because the type of data that is returned is dependent on the type of data that is received. The concept is that when you are given a byte str to process, you return a byte str in your output. When you are given unicode strings to process, you return unicode strings in your output.

This can work well for end users as the ones that know about the difference between the two string types will already have transformed the strings to their desired type before giving it to this function. The ones that don’t can remain blissfully ignorant (at least, as far as your function is concerned) as the function does not change the type.

In cases where the encoding of the byte str is known or can be discovered based on the input data this works well. If you can’t figure out the input encoding, however, this strategy can fail in any of the following cases:

1. It needs to do an internal conversion between byte str and unicode string.
2. It cannot return the same data as either a unicode string or byte str.
3. You may need to deal with byte strings that are not byte-compatible with ASCII

First, a couple examples of using this strategy in a good way:

def translate(msg, table):
    replacements = table.keys()
    new_msg = []
    for index, char in enumerate(msg):
        if char in replacements:
            new_msg.append(table[char])
        else:
            new_msg.append(char)

    return ''.join(new_msg)

In this example, all of the strings that we use (except the empty string which is okay because it doesn’t have any characters to encode) come from outside of the function. Due to that, the user is responsible for making sure that the msg, and the keys and values in table all match in terms of type (unicode vs str) and encoding (You can do some error checking to make sure the user gave all the same type but you can’t do the same for the user giving different encodings). You do not need to make changes to the string that require you to know the encoding or type of the string; everything is a simple replacement of one element in the array of characters in message with the character in table.

import json
from kitchen.text.converters import to_unicode, to_bytes

def first_field_from_json_data(json_string):
    '''Return the first field in a json data structure.

    The format of the json data is a simple list of strings.
    '["one", "two", "three"]'
    '''
    if isinstance(json_string, unicode):
        # On all python versions, json.loads() returns unicode if given
        # a unicode string
        return json.loads(json_string)[0]

    # Byte str: figure out which encoding we're dealing with
    if '\x00' not in json_data[:2]
        encoding = 'utf8'
    elif '\x00\x00\x00' == json_data[:3]:
        encoding = 'utf-32-be'
    elif '\x00\x00\x00' == json_data[1:4]:
        encoding = 'utf-32-le'
    elif '\x00' == json_data[0] and '\x00' == json_data[2]:
        encoding = 'utf-16-be'
    else:
        encoding = 'utf-16-le'

    data = json.loads(unicode(json_string, encoding))
    return data[0].encode(encoding)

In this example the function takes either a byte str type or a unicode string that has a list in json format and returns the first field from it as the type of the input string. The first section of code is very straightforward; we receive a unicode string, parse it with a function, and then return the first field from our parsed data (which our function returned to us as json data).

The second portion that deals with byte str is not so straightforward. Before we can parse the string we have to determine what characters the bytes in the string map to. If we didn’t do that, we wouldn’t be able to properly find which characters are present in the string. In order to do that we have to figure out the encoding of the byte str. Luckily, the json specification states that all strings are unicode and encoded with one of UTF32be, UTF32le, UTF16be, UTF16le, or UTF-8. It further defines the format such that the first two characters are always ASCII. Each of these has a different sequence of NULLs when they encode an ASCII character. We can use that to detect which encoding was used to create the byte str.

Finally, we return the byte str by encoding the unicode back to a byte str.

As you can see, in this example we have to convert from byte str to unicode and back. But we know from the json specification that byte str has to be one of a limited number of encodings that we are able to detect. That ability makes this strategy work.

Now for some examples of using this strategy in ways that fail:

import unicodedata
def first_char(msg):
    '''Return the first character in a string'''
    if not isinstance(msg, unicode):
        try:
            msg = unicode(msg, 'utf8')
        except UnicodeError:
            msg = unicode(msg, 'latin1')
    msg = unicodedata.normalize('NFC', msg)
    return msg[0]

If you look at that code and think that there’s something fragile and prone to breaking in the try: except: block you are correct in being suspicious. This code will fail on multi-byte character sets that aren’t UTF-8. It can also fail on data where the sequence of bytes is valid UTF-8 but the bytes are actually of a different encoding. The reasons this code fails is that we don’t know what encoding the bytes are in and the code must convert from a byte str to a unicode string in order to function.

In order to make this code robust we must know the encoding of msg. The only way to know that is to ask the user so the API must do that:

import unicodedata
def number_of_chars(msg, encoding='utf8', errors='strict'):
    if not isinstance(msg, unicode):
        msg = unicode(msg, encoding, errors)
    msg = unicodedata.normalize('NFC', msg)
    return len(msg)

Another example of failure:

import os
def listdir(directory):
    files = os.listdir(directory)
    if isinstance(directory, str):
        return files
    # files could contain both bytes and unicode
    new_files = []
    for filename in files:
        if not isinstance(filename, unicode):
            # What to do here?
            continue
        new_files.appen(filename)
    return new_files

This function illustrates the second failure mode. Here, not all of the possible values can be represented as unicode without knowing more about the encoding of each of the filenames involved. Since each filename could have a different encoding there’s a few different options to pursue. We could make this function always return byte str since that can accurately represent anything that could be returned. If we want to return unicode we need to at least allow the user to specify what to do in case of an error decoding the bytes to unicode. We can also let the user specify the encoding to use for doing the decoding but that won’t help in all cases since not all files will be in the same encoding (or even necessarily in any encoding):

import locale
import os
def listdir(directory, encoding=locale.getpreferredencoding(), errors='strict'):
    # Note: In python-3.1+, surrogateescape may be a better default
    files = os.listdir(directory)
    if isinstance(directory, str):
        return files
    new_files = []
    for filename in files:
        if not isinstance(filename, unicode):
            filename = unicode(filename, encoding=encoding, errors=errors)
        new_files.append(filename)
    return new_files

Note that although we use errors in this example as what to pass to the codec that decodes to unicode we could also have an errors argument that decides other things to do like skip a filename entirely, return a placeholder (Nondisplayable filename), or raise an exception.

This leaves us with one last failure to describe:

def first_field(csv_string):
    '''Return the first field in a comma separated values string.'''
    try:
        return csv_string[:csv_string.index(',')]
    except ValueError:
        return csv_string

This code looks simple enough. The hidden error here is that we are searching for a comma character in a byte str but not all encodings will use the same sequence of bytes to represent the comma. If you use an encoding that’s not ASCII compatible on the byte level, then the literal comma ',' in the above code will match inappropriate bytes. Some examples of how it can fail:

• Will find the byte representing an ASCII comma in another character
• Will find the comma but leave trailing garbage bytes on the end of the string
• Will not match the character that represents the comma in this encoding

There are two ways to solve this. You can either take the encoding value from the user or you can take the separator value from the user. Of the two, taking the encoding is the better option for two reasons:

1. Taking a separator argument doesn’t clearly document for the API user that the reason they must give it is to properly match the encoding of the csv_string. They’re just as likely to think that it’s simply a way to specify an alternate character (like ”:” or “|”) for the separator.

2. It’s possible for a variable width encoding to reuse the same byte sequence for different characters in multiple sequences.

   Note

   UTF-8 is resistant to this as any character’s sequence of bytes will never be a subset of another character’s sequence of bytes.

With that in mind, here’s how to improve the API:

def first_field(csv_string, encoding='utf-8', errors='replace'):
    if not isinstance(csv_string, unicode):
        u_string = unicode(csv_string, encoding, errors)
        is_unicode = False
    else:
        u_string = csv_string

    try:
        field = u_string[:U_string.index(u',')]
    except ValueError:
        return csv_string

    if not is_unicode:
        field = field.encode(encoding, errors)
    return field

Note

If you decide you’ll never encounter a variable width encoding that reuses byte sequences you can use this code instead:

def first_field(csv_string, encoding='utf-8'):
    try:
        return csv_string[:csv_string.index(','.encode(encoding))]
    except ValueError:
        return csv_string

Separate functions

Sometimes you want to be able to take either byte str or unicode strings, perform similar operations on either one and then return data in the same format as was given. Probably the easiest way to do that is to have separate functions for each and adopt a naming convention to show that one is for working with byte str and the other is for working with unicode strings:

def translate_b(msg, table):
    '''Replace values in str with other byte values like unicode.translate'''
    if not isinstance(msg, str):
        raise TypeError('msg must be of type str')
    str_table = [chr(s) for s in xrange(0,256)]
    delete_chars = []
    for chr_val in (k for k in table.keys() if isinstance(k, int)):
        if chr_val > 255:
            raise ValueError('Keys in table must not exceed 255)')
        if table[chr_val] == None:
            delete_chars.append(chr(chr_val))
        elif isinstance(table[chr_val], int):
            if table[chr_val] > 255:
                raise TypeError('table values cannot be more than 255 or less than 0')
            str_table[chr_val] = chr(table[chr_val])
        else:
            if not isinstance(table[chr_val], str):
                raise TypeError('character mapping must return integer, None or str')
            str_table[chr_val] = table[chr_val]
    str_table = ''.join(str_table)
    delete_chars = ''.join(delete_chars)
    return msg.translate(str_table, delete_chars)

def translate(msg, table):
    '''Replace values in a unicode string with other values'''
    if not isinstance(msg, unicode):
        raise TypeError('msg must be of type unicode')
    return msg.translate(table)

There’s several things that we have to do in this API:

• Because the function names might not be enough of a clue to the user of the functions of the value types that are expected, we have to check that the types are correct.
• We keep the behaviour of the two functions as close to the same as possible, just with byte str and unicode strings substituted for each other.

Deciding whether to take str or unicode when no value is returned

Not all functions have a return value. Sometimes a function is there to interact with something external to python, for instance, writing a file out to disk or a method exists to update the internal state of a data structure. One of the main questions with these APIs is whether to take byte str, unicode string, or both. The answer depends on your use case but I’ll give some examples here.

Writing to external data

When your information is going to an external data source like writing to a file you need to decide whether to take in unicode strings or byte str. Remember that most external data sources are not going to be dealing with unicode directly. Instead, they’re going to be dealing with a sequence of bytes that may be interpreted as unicode. With that in mind, you either need to have the user give you a byte str or convert to a byte str inside the function.

Next you need to think about the type of data that you’re receiving. If it’s textual data, (for instance, this is a chat client and the user is typing messages that they expect to be read by another person) it probably makes sense to take in unicode strings and do the conversion inside your function. On the other hand, if this is a lower level function that’s passing data into a network socket, it probably should be taking byte str instead.

Just as noted in the API notes above, you should specify an encoding and errors argument if you need to transform from unicode string to byte str and you are unable to guess the encoding from the data itself.

Updating data structures

Sometimes your API is just going to update a data structure and not immediately output that data anywhere. Just as when writing external data, you should think about both what your function is going to do with the data eventually and what the caller of your function is thinking that they’re giving you. Most of the time, you’ll want to take unicode strings and enter them into the data structure as unicode when the data is textual in nature. You’ll want to take byte str and enter them into the data structure as byte str when the data is not text. Use a naming convention so the user knows what’s expected.

APIs to Avoid

There are a few APIs that are just wrong. If you catch yourself making an API that does one of these things, change it before anyone sees your code.

Returning unicode unless a conversion fails

This type of API usually deals with byte str at some point and converts it to unicode because it’s usually thought to be text. However, there are times when the bytes fail to convert to a unicode string. When that happens, this API returns the raw byte str instead of a unicode string. One example of this is present in the python standard library: python2’s os.listdir():

>>> import os
>>> import locale
>>> locale.getpreferredencoding()
'UTF-8'
>>> os.mkdir('/tmp/mine')
>>> os.chdir('/tmp/mine')
>>> open('nonsense_char_\xff', 'w').close()
>>> open('all_ascii', 'w').close()
>>> os.listdir(u'.')
[u'all_ascii', 'nonsense_char_\xff']

The problem with APIs like this is that they cause failures that are hard to debug because they don’t happen where the variables are set. For instance, let’s say you take the filenames from os.listdir() and give it to this function:

def normalize_filename(filename):
    '''Change spaces and dashes into underscores'''
    return filename.translate({ord(u' '):u'_', ord(u' '):u'_'})

When you test this, you use filenames that all are decodable in your preferred encoding and everything seems to work. But when this code is run on a machine that has filenames in multiple encodings the filenames returned by os.listdir() suddenly include byte str. And byte str has a different string.translate() function that takes different values. So the code raises an exception where it’s not immediately obvious that os.listdir() is at fault.

Ignoring values with no chance of recovery

An early version of python3 attempted to fix the os.listdir() problem pointed out in the last section by returning all values that were decodable to unicode and omitting the filenames that were not. This lead to the following output:

>>> import os
>>> import locale
>>> locale.getpreferredencoding()
'UTF-8'
>>> os.mkdir('/tmp/mine')
>>> os.chdir('/tmp/mine')
>>> open(b'nonsense_char_\xff', 'w').close()
>>> open('all_ascii', 'w').close()
>>> os.listdir('.')
['all_ascii']

The issue with this type of code is that it is silently doing something surprising. The caller expects to get a full list of files back from os.listdir(). Instead, it silently ignores some of the files, returning only a subset. This leads to code that doesn’t do what is expected that may go unnoticed until the code is in production and someone notices that something important is being missed.

Raising a UnicodeException with no chance of recovery

Believe it or not, a few libraries exist that make it impossible to deal with unicode text without raising a UnicodeError. What seems to occur in these libraries is that the library has functions that expect to receive a unicode string. However, internally, those functions call other functions that expect to receive a byte str. The programmer of the API was smart enough to convert from a unicode string to a byte str but they did not give the user the chance to specify the encodings to use or how to deal with errors. This results in exceptions when the user passes in a byte str because the initial function wants a unicode string and exceptions when the user passes in a unicode string because the function can’t convert the string to bytes in the encoding that it’s selected.

Do not put the user in the position of not being able to use your API without raising a UnicodeError with certain values. If you can only safely take unicode strings, document that byte str is not allowed and vice versa. If you have to convert internally, make sure to give the caller of your function parameters to control the encoding and how to treat errors that may occur during the encoding/decoding process. If your code will raise a UnicodeError with non-ASCII values no matter what, you should probably rethink your API.

Knowing your data

If you’ve read all the way down to this section without skipping you’ve seen several admonitions about the type of data you are processing affecting the viability of the various API choices.

Here’s a few things to consider in your data:

Do you need to operate on both bytes and unicode?

Much of the data in libraries, programs, and the general environment outside of python is written where strings are sequences of bytes. So when we interact with data that comes from outside of python or data that is about to leave python it may make sense to only operate on the data as a byte str. There’s two times when this may make sense:

1. The user is intended to hand the data to the function and then the function takes care of sending the data outside of python (to the filesystem, over the network, etc).
2. The data is not representable as text. For instance, writing a binary file format.

Even when your code is operating in this area you still need to think a little more about your data. For instance, it might make sense for the person using your API to pass in unicode strings and let the function convert that into the byte str that it then sends over the wire.

There are also times when it might make sense to operate only on unicode strings. unicode represents text so anytime that you are working on textual data that isn’t going to leave python it has the potential to be a unicode-only API. However, there’s two things that you should consider when designing a unicode-only API:

1. As your API gains popularity, people are going to use your API in places that you may not have thought of. Corner cases in these other places may mean that processing bytes is desirable.
2. In python2, byte str and unicode are often used interchangably with each other. That means that people programming against your API may have received str from some other API and it would be most convenient for their code if your API accepted it.

Note

In python3, the separation between the text type and the byte type are more clear. So in python3, there’s less need to have all APIs take both unicode and bytes.

Can you restrict the encodings?

If you determine that you have to deal with byte str you should realize that not all encodings are created equal. Each has different properties that may make it possible to provide a simpler API provided that you can reasonably tell the users of your API that they cannot use certain classes of encodings.

As one example, if you are required to find a comma (,) in a byte str you have different choices based on what encodings are allowed. If you can reasonably restrict your API users to only giving ASCII compatible encodings you can do this simply by searching for the literal comma character because that character will be represented by the same byte sequence in all ASCII compatible encodings.

The following are some classes of encodings to be aware of as you decide how generic your code needs to be.

Single byte encodings

Single byte encodings can only represent 256 total characters. They encode the code points for a character to the equivalent number in a single byte.

Most single byte encodings are ASCII compatible. ASCII compatible encodings are the most likely to be usable without changes to code so this is good news. A notable exception to this is the EBDIC family of encodings.

Multibyte encodings

Multibyte encodings use more than one byte to encode some characters.

Fixed width

Fixed width encodings have a set number of bytes to represent all of the characters in the character set. UTF-32 is an example of a fixed width encoding that uses four bytes per character and can express every unicode characters. There are a number of problems with writing APIs that need to operate on fixed width, multibyte characters. To go back to our earlier example of finding a comma in a string, we have to realize that even in UTF-32 where the code point for ASCII characters is the same as in ASCII, the byte sequence for them is different. So you cannot search for the literal byte character as it may pick up false positives and may break a byte sequence in an odd place.

Variable Width

ASCII compatible

UTF-8 and the EUC family of encodings are examples of ASCII compatible multi-byte encodings. They achieve this by adhering to two principles:

• All of the ASCII characters are represented by the byte that they are in the ASCII encoding.
• None of the ASCII byte sequences are reused in any other byte sequence for a different character.

Escaped

Some multibyte encodings work by using only bytes from the ASCII encoding but when a particular sequence of those byes is found, they are interpreted as meaning something other than their ASCII values. UTF-7 is one such encoding that can encode all of the unicode code points. For instance, here’s a some Japanese characters encoded as UTF-7:

>>> a = u'\u304f\u3089\u3068\u307f'
>>> print a
くらとみ
>>> print a.encode('utf-7')
+ME8wiTBoMH8-

These encodings can be used when you need to encode unicode data that may contain non-ASCII characters for inclusion in an ASCII only transport medium or file.

However, they are not ASCII compatible in the sense that we used earlier as the bytes that represent a ASCII character are being reused as part of other characters. If you were to search for a literal plus sign in this encoded string, you would run across many false positives, for instance.

Other

There are many other popular variable width encodings, for instance UTF-16 and shift-JIS. Many of these are not ASCII compatible so you cannot search for a literal ASCII character without danger of false positives or false negatives.