I would still do a benchmark, for these reasons:
1) IIRC rawdata might be the whole document (or at least everything that has not been parsed yet);
2) the '>' is very likely to be found;

This situation is fairly different from the one presented in #17170, where the strings are shorts and the character is not present in the majority of the strings.

Profiling and improving html.parser (and hence _markupbase) was already on my todo list (even if admittedly not anywhere near the top :), so writing a benchmark for it might be useful for further enhancements too.

(Note: HTMLParser is already fairly fast, parsing ~1.3MB/s according to http://www.crummy.com/2012/02/06/0, but I've never done anything to make it even faster, so there might still be room for improvements.)