How to filter (or replace) unicode characters that would take more than 3 bytes in UTF-8?

Question

I'm using Python and Django, but I'm having a problem caused by a limitation of MySQL. According to the MySQL 5.1 documentation, their utf8 implementation does not support 4-byte characters. MySQL 5.5 will support 4-byte characters using utf8mb4; and, someday in future, utf8 might support it as well.

But my server is not ready to upgrade to MySQL 5.5, and thus I'm limited to UTF-8 characters that take 3 bytes or less.

My question is: How to filter (or replace) unicode characters that would take more than 3 bytes?

I want to replace all 4-byte characters with the official \ufffd (U+FFFD REPLACEMENT CHARACTER), or with ?.

In other words, I want a behavior quite similar to Python's own str.encode() method (when passing 'replace' parameter). Edit: I want a behavior similar to encode(), but I don't want to actually encode the string. I want to still have an unicode string after filtering.

I DON'T want to escape the character before storing at the MySQL, because that would mean I would need to unescape all strings I get from the database, which is very annoying and unfeasible.

See also:

• "Incorrect string value" warning when saving some unicode characters to MySQL (at Django ticket system)
• ‘𠂉’ Not a valid unicode character, but in the unicode character set? (at Stack Overflow)

[EDIT] Added tests about the proposed solutions

So I got good answers so far. Thanks, people! Now, in order to choose one of them, I did a quick testing to find the simplest and fastest one.

#!/usr/bin/env python
# -*- coding: utf-8 -*-
# vi:ts=4 sw=4 et

import cProfile
import random
import re

# How many times to repeat each filtering
repeat_count = 256

# Percentage of "normal" chars, when compared to "large" unicode chars
normal_chars = 90

# Total number of characters in this string
string_size = 8 * 1024

# Generating a random testing string
test_string = u''.join(
        unichr(random.randrange(32,
            0x10ffff if random.randrange(100) > normal_chars else 0x0fff
        )) for i in xrange(string_size) )

# RegEx to find invalid characters
re_pattern = re.compile(u'[^\u0000-\uD7FF\uE000-\uFFFF]', re.UNICODE)

def filter_using_re(unicode_string):
    return re_pattern.sub(u'\uFFFD', unicode_string)

def filter_using_python(unicode_string):
    return u''.join(
        uc if uc < u'\ud800' or u'\ue000' <= uc <= u'\uffff' else u'\ufffd'
        for uc in unicode_string
    )

def repeat_test(func, unicode_string):
    for i in xrange(repeat_count):
        tmp = func(unicode_string)

print '='*10 + ' filter_using_re() ' + '='*10
cProfile.run('repeat_test(filter_using_re, test_string)')
print '='*10 + ' filter_using_python() ' + '='*10
cProfile.run('repeat_test(filter_using_python, test_string)')

#print test_string.encode('utf8')
#print filter_using_re(test_string).encode('utf8')
#print filter_using_python(test_string).encode('utf8')

The results:

• filter_using_re() did 515 function calls in 0.139 CPU seconds (0.138 CPU seconds at the sub() built-in)
• filter_using_python() did 2097923 function calls in 3.413 CPU seconds (1.511 CPU seconds at the join() call and 1.900 CPU seconds evaluating the generator expression)
• I did no test using itertools because... well... that solution, although interesting, was quite big and complex.

Conclusion

The RegEx solution was, by far, the fastest one.

Solution 1

Unicode characters in the ranges \u0000-\uD7FF and \uE000-\uFFFF will have 3 byte (or less) encodings in UTF8. The \uD800-\uDFFF range is for multibyte UTF16. I do not know python, but you should be able to set up a regular expression to match outside those ranges.

pattern = re.compile("[\uD800-\uDFFF].", re.UNICODE)
pattern = re.compile("[^\u0000-\uFFFF]", re.UNICODE)

Edit adding Python from Denilson Sá's script in the question body:

re_pattern = re.compile(u'[^\u0000-\uD7FF\uE000-\uFFFF]', re.UNICODE)
filtered_string = re_pattern.sub(u'\uFFFD', unicode_string)    

Solution 2

You may skip the decoding and encoding steps and directly detect the value of the first byte (8-bit string) of each character. According to UTF-8:

#1-byte characters have the following format: 0xxxxxxx
#2-byte characters have the following format: 110xxxxx 10xxxxxx
#3-byte characters have the following format: 1110xxxx 10xxxxxx 10xxxxxx
#4-byte characters have the following format: 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx

According to that, you only need to check the value of only the first byte of each character to filter out 4-byte characters:

def filter_4byte_chars(s):
    i = 0
    j = len(s)
    # you need to convert
    # the immutable string
    # to a mutable list first
    s = list(s)
    while i < j:
        # get the value of this byte
        k = ord(s[i])
        # this is a 1-byte character, skip to the next byte
        if k <= 127:
            i += 1
        # this is a 2-byte character, skip ahead by 2 bytes
        elif k < 224:
            i += 2
        # this is a 3-byte character, skip ahead by 3 bytes
        elif k < 240:
            i += 3
        # this is a 4-byte character, remove it and update
        # the length of the string we need to check
        else:
            s[i:i+4] = []
            j -= 4
    return ''.join(s)

Skipping the decoding and encoding parts will save you some time and for smaller strings that mostly have 1-byte characters this could even be faster than the regular expression filtering.

Solution 3

Encode as UTF-16, then reencode as UTF-8.

>>> t = u'𝐟𝐨𝐨'
>>> e = t.encode('utf-16le')
>>> ''.join(unichr(x).encode('utf-8') for x in struct.unpack('<' + 'H' * (len(e) // 2), e))
'\xed\xa0\xb5\xed\xb0\x9f\xed\xa0\xb5\xed\xb0\xa8\xed\xa0\xb5\xed\xb0\xa8'

Note that you can't encode after joining, since the surrogate pairs may be decoded before reencoding.

EDIT:

MySQL (at least 5.1.47) has no problem dealing with surrogate pairs:

mysql> create table utf8test (t character(128)) collate utf8_general_ci;
Query OK, 0 rows affected (0.12 sec)

  ...

>>> cxn = MySQLdb.connect(..., charset='utf8')
>>> csr = cxn.cursor()
>>> t = u'𝐟𝐨𝐨'
>>> e = t.encode('utf-16le')
>>> v = ''.join(unichr(x).encode('utf-8') for x in struct.unpack('<' + 'H' * (len(e) // 2), e))
>>> v
'\xed\xa0\xb5\xed\xb0\x9f\xed\xa0\xb5\xed\xb0\xa8\xed\xa0\xb5\xed\xb0\xa8'
>>> csr.execute('insert into utf8test (t) values (%s)', (v,))
1L
>>> csr.execute('select * from utf8test')
1L
>>> r = csr.fetchone()
>>> r
(u'\ud835\udc1f\ud835\udc28\ud835\udc28',)
>>> print r[0]
𝐟𝐨𝐨

Solution 4

And just for the fun of it, an itertools monstrosity :)

import itertools as it, operator as op

def max3bytes(unicode_string):

    # sequence of pairs of (char_in_string, u'\N{REPLACEMENT CHARACTER}')
    pairs= it.izip(unicode_string, it.repeat(u'\ufffd'))

    # is the argument less than or equal to 65535?
    selector= ft.partial(op.le, 65535)

    # using the character ordinals, return 0 or 1 based on `selector`
    indexer= it.imap(selector, it.imap(ord, unicode_string))

    # now pick the correct item for all pairs
    return u''.join(it.imap(tuple.__getitem__, pairs, indexer))

Solution 5

According to the MySQL 5.1 documentation: "The ucs2 and utf8 character sets do not support supplementary characters that lie outside the BMP." This indicates that there might be a problem with surrogate pairs.

Note that the Unicode standard 5.2 chapter 3 actually forbids encoding a surrogate pair as two 3-byte UTF-8 sequences instead of one 4-byte UTF-8 sequence ... see for example page 93 """Because surrogate code points are not Unicode scalar values, any UTF-8 byte sequence that would otherwise map to code points D800..DFFF is ill-formed.""" However this proscription is as far as I know largely unknown or ignored.

It may well be a good idea to check what MySQL does with surrogate pairs. If they are not to be retained, this code will provide a simple-enough check:

all(uc < u'\ud800' or u'\ue000' <= uc <= u'\uffff' for uc in unicode_string)

and this code will replace any "nasties" with u\ufffd:

u''.join(
    uc if uc < u'\ud800' or u'\ue000' <= uc <= u'\uffff' else u'\ufffd'
    for uc in unicode_string
    )

View more solutions

Comments

[John Machin]

Perhaps it should exclude surrogates. Also: uc <= u'\uffff' might be better than ord(uc) < 65536

[Ishbir]

Perhaps struct.unpack('<%dH' % (len(e)//2), e)?

[Philipp]

“However this proscription is as far as I know largely unknown or ignored.”—Hopefully not! At least Python 3 refuses to encode surrogate code points (try chr(55349).encode("utf-8")).

[John Machin]

@Philipp: Python 3 does seem to do the "right thing" -- however your example is a LONE surrogate which is a different problem; Python 2 passes that test but not this one: "\xed\xa0\x80\xed\xb0\x80".decode('utf8') produces u'\U00010000' instead of an exception.

[John Machin]

(1) The MySQL docs that I referred to declare the charset as part of the column definition: t character(128) character set utf8 ... are you sure that what you have is equivalent? (2) Try your UTF-16 stunt with Python 3.1 :-)

[Ignacio Vazquez-Abrams]

@John: (1) Retested with character set utf8 on 2.6. Results were the same. (2) That's just a limitation of the stock UTF-8 codec. It can be worked around with a custom codec. Or with MySQL doing the right thing in the first place.

[Denilson Sá Maia]

Hmmm... You forgot to add the u prefix to all strings! It should have been u'\ufffd'. ;)

[Flimm]

Note that the strings "[^\u0000-\uFFFF]" etc are not raw strings, that is, the string literals are not prefixed with r!

[Rolando Urquiza]

I had to change the the first range end in u'[^\u0000-\uD7FF\uE000-\uFFFF]' from '\uD7FF' to '\u07FF' because there where still some chars not beign cleaned.