Gregory Maxwell wrote: I have only seen this on some older dumps. Have not seen it for a while. I parse them out of the dumps. Jeff

-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 Gregory Maxwell wrote: [snip] Hypothetically, almost anyplace that crops strings or otherwise does internal string manipulation other than with Language::truncate() could end up spitting out bad UTF-8. :P Web POST and GET input is sanitized (normalized and bad character stripped) at the WebRequest level, but internal processing is not always pure, and in most cases output is not sanitized either. Incorrect cropping of long values in limited-length database fields is another possibility. (Note that XML dump generation specifically runs a UTF-8 cleanup step on output; the XML dump output is thus guaranteed to be UTF-8-clean and NFC.) Running UtfNormal::cleanUp() or similar should guarantee well-formed UTF-8 output. Any invalid bytes will be replaced with the Unicode replacement character U+FFFE which is reserved for this purpose. Thanks to UTF-8's structured value space, accidentally removing ASCII field delimiters in the cleanup step isn't possible -- the ASCII byte cannot be confused with a trailing byte for a longer sequence -- so in theory it should be safe to run a single step at the end instead of many individual cleanup checks. - -- brion vibber (brion @ pobox.com / brion @ wikimedia.org)

Brion Vibber wrote: This statement is true. I have never seen a Wikipeida XML dump with unicode errors. Jeff