The duplication prevention system is brilliant and brings to mind, though I’m not sure why, the fact that SL notecards are “recreated” meaning, I believe, that they’re assigned a new UUID every time they’re updated. That’s what I’ve been told is the complication in allowing scripts to write to them which is an interesting point you might add to your inventory article when you write it because I told you to so there.

UUIDs on the backend are actually fine when you arent using MySQL. MySQL optimises UUIDs very badly; treating them the same as any other text field (which arent good for indexes) – Postgres, MS-SQL and others either have a GUID class which treats it as a 128-bit integer, or handle text indexes better.

Inserting into MySQL; it’s actually best to convert 128-bit into 4×32-bit integers, then make one ‘unique’ key from the composite of all four fields. That hints MySQL enough to treat it as a 128-bit int and gets you the performance you want; albeit at the cost of not being able to read the UUIDs easily anymore.

@Flip: We dont use UUIDs for table joins anywhere – OpenSim is fairly naieve and doesnt need them; all the transactions we do are on a key/value store basis mostly. (which works well for scaling into distributed systems)

Table joins are only used for things like analytics — and we have to be careful with some of those, since it’s very easy to cause a nasty table lock which pauses the whole grid.

As for memcache – I can tell you OSGrid does this already; not memcache itself – but we do rely on a frequently accessed files cache to boost performance on the top requested 3GB of assets.

Perhaps it would be appropriate to have the backend database node keep an index of it’s internal (highly optimised and indexed) ids (straight INT/BIGINT UNSIGNED) and mapped to the GUIDs that the requesting application would use.

Again, I’m still interested in helping out with this side of things – but I think with my other commitments (for now) I’d need to be given a specific problem and scope. You have my number Adam

Incidentally, What would happen if a new asset server came online with objects with the same GUID (via chance or even a malicious user)? It’s vital that the platform can handle this gracefully, especially for a place such as OSGrid.

I haven’t looked at the DB schema for OpenSim in a while – while I agree UUIDs are great for unique identifiers, they are several orders of magnitude slower than integers when doing things within a DB implementation, especially when joining tables, etc. Are there plans to keep UUIDs for unique lookup, while using integers for things like joins within a DB instance? Great work, and way to break it down in English, Adam.

Regards,

-Tim / Flip