-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 > I'm know. But you can implement locale-dependent version for single language > only.
- From the German example, you can't even do that (name order is different than dictionary order). I think we are agreed that the default SQLite implementation gets ASCII right and makes no attempt to deal specifically with non-ASCII locales. The ICU extension gets all the locales as right as possible which is why it is huge and slow. So that leaves a middle ground of somewhat right but being lighter-weight. The problem with having that as a default part of SQLite is that it will be giving the wrong answers, but rarely will developers realise just how wrong. And there are various tradeoffs that can be made between size/performance and correctness. Consequently it should be documented just where each implementation stands. The one you linked to is nice in that it documents in the code exactly how much bigger things become and you documented it being 4x faster than ICU. But what isn't documented is how accurate it is. If I took all the text from a leading newspaper in each locale, how well would it do. Would it deal correctly with the name of the prime minister or the capital city? An alternate approach would be working with the ICU folks to improve the size and performance of their library. For example the code could be refactored to have fast paths for the most common conversions to improve performance, or be able to omit various lesser used locales to improve size. Roger -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.9 (GNU/Linux) iEYEARECAAYFAknAKkkACgkQmOOfHg372QT++QCfdWoE0nW5Cu2MbuZKzI49+hlx PSMAoLuN79Zh3dcHxKxS1L/QJOCGEpH8 =kWCI -----END PGP SIGNATURE----- _______________________________________________ sqlite-users mailing list sqlite-users@sqlite.org http://sqlite.org:8080/cgi-bin/mailman/listinfo/sqlite-users