You want version 1 UUIDs. They don't collide but they leak the generating computer's MAC address and require a monotonic clock.

From RFC4122 we have this format

8 bytes time-low
2 bytes time-mid
2 bytes time-high & version (high nybble is a 1)
2 bytes clock-seq-high & reserved (high two bits are 1 and 0)
2 bytes clock-seq-low
6 bytes MAC address

Clock-sequence is actually the subsecond portion of the clock now.

The thing about the two byte fiedls is they are not two one byte fields. If you mess the code up, you get different results on little endian and big endian machines and you have a chance of collision again. The string form looks like this no matter what your endian is:

????????-????-1???-[89AB]???-?[02468ACE]??????????

The thing about this is when it works, it works. If you don't have MAC address collisions (the general rule is if you get colliding MAC addresses, RMA both cards) and if you actually have a monotonic clock, this works 100% of the time.

There's a work around for not having a monotonic clock. That is, do it the old way. Serialize all UUID generations and use clock-seq as a counter that increments with each generation.

The thing about this is you will have bad generation and most likely collisions if your PC battery dies on any node that generates.

If you can't tolerate leaking the MAC addresses, you're stuck with lesser means. V4 UUIDs are really good for all terrestrial use under modern hardware assumptions (has thermal diode or other hardware random number generator). V5 UUIDs can do the job under specific circumstances (they're a hash of something else; the main problem is they're subject to adversarial attack).

You want version 1 UUIDs. They don't collide but they leak the generating computer's MAC address and require a monotonic clock.

From RFC4122 we have this format

8 bytes time-low
2 bytes time-mid
2 bytes time-high & version (high nybble is a 1)
2 bytes clock-seq-high & reserved (high two bits are 1 and 0)
2 bytes clock-seq-low
6 bytes MAC address

Clock-sequence is actually the subsecond portion of the clock now.

The thing about the two byte fiedls is they are not two one byte fields. If you mess the code up, you get different results on little endian and big endian machines and you have a chance of collision again. The string form looks like this no matter what your endian is:

????????-????-1???-[89AB]???-?[02468ACE]??????????

The thing about this is when it works, it works. If you don't have MAC address collisions (the general rule is if you get colliding MAC addresses, RMA both cards) and if you actually have a monotonic clock, this works 100% of the time.

There's a work around for not having a monotonic clock. That is, do it the old way. Serialize all UUID generations and use clock-seq as a counter that increments with each generation.

The thing about this is you will have bad generation and most likely collisions if your PC battery dies on any node that generates.

It's been pointed out in comments that VMs don't really get globally unique MAC addresses, and it has also been pointed out that the solution to this is to hand out the node IDs yourself. When you do this you are supposed to set the broadcast bit (the one marked [02468ACE] so that it would match [13579BDF] instead).

If you can't tolerate leaking the MAC addresses and can't issue node IDs, you're stuck with lesser means. V4 UUIDs are really good for all terrestrial use under modern hardware assumptions (has thermal diode or other hardware RNG). V5 UUIDs can do the job under specific circumstances (they're a hash of something else; the main problem is they're subject to adversarial attack).

You want version 1 UUIDs. They don't collide but they leak the generating computer's MAC address and require a monotonic clock.

From RFC4122 we have this format

8 bytes time-low
2 bytes time-mid
2 bytes time-high & version (high nybble is a 1)
2 bytes clock-seq-high & reserved (high two bits are 1 and 0)
2 bytes clock-seq-low
6 bytes MAC address

Clock-sequence is actually the subsecond portion of the clock now.

The thing about the two byte fiels is they are not two one byte fields. If you mess the code up, you get different results on little endian and big endian machines and you have a chace of collision again. The string form looks like this no matter what your endian is:

????????-????-1???-[89AB]???-?[02468ACE]??????????

The thing about this is when it works, it works. If you don't have MAC address collisions (the general rule is if you get colliding MAC addresses, RMA both cards) and if you actually have a monotonic clock, this works 100% of the time.

There's a work around for not having a monotonic clock. That is, do it the old way. Serialize all UUID generations and use clock-seq as a counter that increments with each generation.

The thing about this is you will have bad generation and most likely collisions if your PC battery dies on any node that generates.

If you can't tolerate leaking the MAC addresses, you're stuck with lesser means. V4 UUIDs are really good for all terrestrial use under modern hardware assumptions (has thermal diode or other hardware RNG). V5 UUIDs can do the job under specific circumstances (they're a hash of something else; the main problem is they're subject to adversarial attack).

You want version 1 UUIDs. They don't collide but they leak the generating computer's MAC address and require a monotonic clock.

From RFC4122 we have this format

8 bytes time-low
2 bytes time-mid
2 bytes time-high & version (high nybble is a 1)
2 bytes clock-seq-high & reserved (high two bits are 1 and 0)
2 bytes clock-seq-low
6 bytes MAC address

Clock-sequence is actually the subsecond portion of the clock now.

The thing about the two byte fiedls is they are not two one byte fields. If you mess the code up, you get different results on little endian and big endian machines and you have a chance of collision again. The string form looks like this no matter what your endian is:

????????-????-1???-[89AB]???-?[02468ACE]??????????

The thing about this is when it works, it works. If you don't have MAC address collisions (the general rule is if you get colliding MAC addresses, RMA both cards) and if you actually have a monotonic clock, this works 100% of the time.

There's a work around for not having a monotonic clock. That is, do it the old way. Serialize all UUID generations and use clock-seq as a counter that increments with each generation.

The thing about this is you will have bad generation and most likely collisions if your PC battery dies on any node that generates.

If you can't tolerate leaking the MAC addresses, you're stuck with lesser means. V4 UUIDs are really good for all terrestrial use under modern hardware assumptions (has thermal diode or other hardware random number generator). V5 UUIDs can do the job under specific circumstances (they're a hash of something else; the main problem is they're subject to adversarial attack).