[xwiki-devs] [Proposal] Reduce risk of duplicated document id (XWIKI-4396)

Since nobody seems to have remark nor opinion, I have started to do A), but it was still insufficient for my requirements. So I have now done B) without fallback using MD5, and if no one stop me, I will probably do C) soon. Since the storage is entirely broken by this change, I need the startup to be stopped on migration failure and migration to be mandatory. Migration error was ignored until now, but this is wrong IMO. It would means that the code was both backward compatible with old data and failsafe to any migration failure. Moreover, Migration could also be disabled by config, but this is not really safe IMO. Currently I have added code that throws if migrations fails, preventing the wiki to be started. Should I improve further ? could we expect user to take care to migration requirement, or should we have mandatory migration and optional one ? I am also wondering if I should target to commit in release 3.2 or 3.3 since the name of the migrator and the db version will depends on this. Hoping to have some comments this time. On Mon, Sep 19, 2011 at 12:54, Denis Gervalle <dgl(a)softec.lu> wrote: improve this. String.hashcode(fullName:lang). Since this is a really poor hashing method for small changes, the risk that two documents with similar names of same length are colliding is really high. This Id is used by the Hibernate store for document unicity and really needs to be unique, and at most a 64bit-numeric at the same time. Currently we use only 32bits since java hashcode are limited to 32bits integers. fullName:lang, but converting the current implementation is not really easy. This is probably why XWIKI-4396 has never been fixed. Therefore, my current goal is to reduce the likeliness of a collision by choosing a better hashing method and taking into account the fact that document fullname are short string and the number of unique ids required are very limited (since the unicity is only expected for a given XWiki database) compared to the 64bits integrer range. options: runs on 64bits, I was thinking about using the algorithm produced by Professor Daniel J. Bernstein (DJB) since it is well-know, wildly used, easy to implement algorithm with a good distribution on small strings. than now; require db migration of all document keys truncating to the lower 64bits. Note that oldcore already use MDA5 for hashcoding a whole XWikiDocument to provide the API with a getVersionHashcode(), and for the validation hash used by the persistent login manager. The first use Object.hashcode() as a fallback, which is really bad and defeat the purpose. The second does not provide any fallback and may fail unexpectedly. For our case, if we really want a fallback, we needs to store the hashing algorithm used in the database at creation time, and anyway, fail when it was not available. really flexible since it would be possible to choose the algorithm at creation time and does not require full migration for existing database. probably as a column of xwikidbversion; if this config value is broken, the whole DB is broken was thinking about integrating something like http://twmacinta.com/myjava/fast_md5.php (non-native version) which is LGPL. This will ensure availability of the hashing algorythm while having a rather strong one. persistent manager found it insufficient. Caleb ? overkill ? flawless -- Denis Gervalle SOFTEC sa - CEO eGuilde sarl - CTO -- Denis Gervalle SOFTEC sa - CEO