Are GUID collisions possible?

Basically, no. I think someone went mucking with your database. Depending on the version GUID you're using the value is either unique (for things like version 1 GUIDs), or both unique and unpredictable (for things like version 4 GUIDs). SQL Server's implementation for their NEWID() function appears to use a 128-bit random number, so you're not going to get a collision.

For a 1% chance of collision, you'd need to generate about 2,600,000,000,000,000,000 GUIDs.

Basically they are not possible !, the chances are astronomically low.

But... I'm the only person I the world that I know of, that had a GUID colision once (yep!).

And I'm sure of it, and that it wasn't a mistake.

How did it happen, in a small application that was running on Pocket PC, at the end of an operation a command that has an generated GUID must be issued. The command after it was executed on the server it was stored in a command table on the server along with the execution date. One day when I was debugging I issued the module command (with the newly generated GUID attached) and nothing happened. I did it again (with the same guid, because the guid was generated only once at the beginning of the operation), and again, and nothing, finally trying to find out why the command isn't executing, I checked the command table, and the same GUID as the current one was inserted 3 weeks ago. Not believing this, I restored a database from 2 weeks backup, and the guid was there. Checked the code, the new guid was freshly generated no doubt about it. Pow guid collision, happened only once, but I really wish I would have won at lotto instead,the chance is greater :).

Edit: there are some factors that could have greatly increased the chance of this happening, the application was running on the PocketPC emulator, and the emulator has a save state feature, which means that every time the state is restored the local time is restored also and the guid is based on on the internal timer....also the guid generating algorithm for compact framework might be less complete than for example the COM one...

They are theoretically possible, but with 3.4E38 possible numbers, if you create tens of trillions of GUIDs in a year the chance of having one duplicate is 0.00000000006 (Source).

If two users ended up with the same GUID, I would wager that there is a bug in the program which is causing the data to be copied or shared.