In high-reliability systems a criterion for identifier design is easy detection of defective identifiers. This includes buggy systems and adversarial manipulation.
The problem with UUIDs that rely on entropy sources is that it is computationally expensive to detect if the statistical distribution of identifiers is diverging from what you would expect from a random oracle. I've written systems that can detect entropy source anomalies but you'll want to turn it off in production.
It is pretty cheap to sanity check most non-probabilistic identifier schemes. UUIDs that use broken hash algorithms (e.g. UUIDv3/5) or leak state (e.g. UUIDv7) are exposed to adversarial exploitation.
The identifier scheme is dependent on the use case. Does the uniqueness constraint apply to the instance of the object or the contents of the object? Is the generation of identifiers federated across untrusted nodes? How large is the potential universe of identifiers?
The basic scheme I've seen is a 128-bit structured value that has no probabilistic component. These identifiers can be encrypted with AES-128 when exported to the public, guaranteeing uniqueness while leaking no internal state. The benefit of this scheme is that it is usually drop-in compatible with standard UUID even though it is technically not a UUID and the internal structure can carry useful metadata about the identifier if you can decrypt it.
Federated generation across untrusted nodes requires a more complex scheme, particularly if the universe of identifiers is extremely large. These intrinsically have a collision risk regardless of how the identifiers are generated.
All of the standardized UUID really weren't designed with the requirements of scalable high-reliability systems in mind. They were optimized for convenience and expedience which is a perfectly reasonable objective. Most people don't need an identifier system engineered for extreme reliability, even though there is relatively little cost to having one.
How does a high-reliability system have a broken /dev/random? You're better off fixing it rather than trying to fix every downstream component that uses it. You can put your AES-128 counter there if you can count reliably.
The latest UUID (7?) Uses half random gen, half timestamp. This not only makes it sortable by creation, but would also make a collision like this impossible.
It's still possible in most implementations of UUIDv7.
UUIDv7 assigns the first 48 bits for the timestamp in milliseconds. You can generate a lot of UUID's in a millisecond though!
Then you have another 12 bits that you can use as you wish; "rand_a". The spec has a few methods they suggest on how to use these bits including 12 bits of random data, using it for sub-millisecond timestamps, or creating a monotonic counter, but each have their downsides:
- Purely random data means you can still run into collisions and anything within the same millisecond is unordered
- Sub millisecond you can run into collisions; there's nothing stopping you from generating two UUID's with the same 62 bits of rand_b data in the same sub-millisecond timestamp.
- Monotonic counters can overflow before the next tick, then what? Rollover? Once you roll over it's no longer monotonic and you can generate the same random data within the same monotonic cycle. Also; it's only monotonic to the system that's generating the UUID. If you have a distributed system and they each have their own monotonic cycles then you'll be generating UUID's with the same timestamp + monotonic counter, and again, are relying on not generating the same random data.
You can steal some of the 62 bits in rand_b if you want as well; you can use rand_a for sub-millisecond accuracy, and then use a few bits of rand_b for a monotonic counter. There's still a chance of collision here, but it's exceedingly low at the expense of less truly random data at the end.
If you want truly collision free, you'd also need to assign a couple of bits to identify the subsystem generating the UUID so that the monotonic counter is unique to that subsystem. You lose the ordering part of the monotonic counter this way though, but I guess you could argue that in nearly 100% of cases the accuracy of sub-millisecond order in a distributed system is a lie anyways.
I think by the time you're building a system that needs to generate (and persist!) billions of identifiers per millisecond, you're solidly past the point where all your design decisions need to be vetted for whether they make sense on your extremely exotic setup.
But 12 bits is not "billions of identifiers" -- it's 4096. Once you exhaust that counter in the same millisecond, you are still relying on a gamble that your random source will not generate the exact same bit sequence for the previous same counter value. And this thread started out with the OP explaining that random collisions are much more common than we'd like them to be, for various reasons.
We have a dedicated snowflake id generator service that returns batch ids. It's also distributed, each service adds its own instance number to the id. When it overflows it just blocks for the next ms. For our traffic, it's never a bottleneck.
Something I use on my own distributed system (where I wanted 64-bit IDs), is use 32 bits for the time in seconds (with an epoch from 2020, so good until 2088), 8 bits for the device ID and 24 bits for a serial number (resets to 0 every time the seconds increments).
That's generally enough IDs per second for most of my edge nodes, but the central worker nodes need more, so I give them a different split and use 4 bits for the device ID and 28 bits for serial number instead.
If a node overflows its serial number that second, I kind of cheat and increment the seconds field early. Every time this happens, I persist the seconds field to the database, and when the app restarts, it starts its seconds count at the last persisted seconds plus one. If the current time in seconds is greater than the last used seconds, I also update it and reset the serial number. Works remarkably well for smoothing out very occasional spikes in ID generation while still approximately remaining globally sortable.
I also "waste" a bit of the 32-bit time field by considering it to be signed, even though it's not really because I don't expect this system to last long enough to reach times where the MSB gets set. But if I ever change my system, I'll set that bit and everything will stay ordered. I'll probably reset the epoch at that point too.
The inclusion of a timestamp in v7 makes collisions impossible unless the generating systems think that the time is the same down to the millisecond, which makes the temporal distance quite relevant.
Plenty of systems end up generating multiple UUID's in a single millisecond.
The issue with UUIDv7 is that you also have significantly less entropy since you only have a 62 bits (sometimes less, depending on implementation) of "random" data. So while the time aspect of format lowers the chances of collisions, generating two UUIDv7's in the same millisecond (depending on implementation) have a significantly higher chance of collision than two UUIDv4's.
It's still incredibly unlikely, but it's also incredibly unlikely you generate two matching UUIDv4's, but it does happen.
TLDR; It's possible to generate matching UUIDv7's, don't assume otherwise.
But essentially, using UUID v7 you actually have less risk of collisions than with UUID v4.
Because of the birthday paradox, if you have N bits of randomness, you can expect a collision approximately after (2^((N/2)-1)) random numbers.
With v4, you have 122 bits of entropy over all time, so will see a collision after 2^60 allocations, approx 1.2 x 10^18.
With v7, you sacrifice 48 bits of entropy to give you 74 bits of entropy every millisecond, so you will see a collision after approximate 2^36 allocations per millisecond, approx 6.8 x 10^10 per millisecond.
You could argue that the risk of a collision is too high per millisecond because it's likely that 68 billion UUIDs are generated every millisecond. And maybe I'd agree. But the counter argument is that with v4 you'd expect a collision after 2^24 milliseconds, or 280 minutes, allocating at the same rate of 68 billion UUIDs per millisecond.
Obviously "all time" is longer than "280 minutes", so v7 is actually statistically less likely to cause collisions than v4, even though it seems counter-intuitive because it has a smaller space devoted to entropy. The key insight is that the time provides bits that are guaranteed to be unique, so only collisions within the same timestamp are significant, and every bit used to provide known-unique values is worth 2 bits of entropy.
Sorry if I worded poorly but you’re definitely less likely to run into a collision with v7, but it’s not impossible, which is what I was trying to point out.
Almost impossible, it depends on how fast they're being generated and the precision of the timestamp. The real problem is two years later when someone finds and removes that usleep(10000); /* sleep 10 µs */ that was the hard speed brake needed for the UUID generator, and suddenly duplicate IDs start showing up a few times per day or something similar.
That depends on your definition of high-availability. If high availability includes distributed writers, (global) sequences are not the best solution because generating unique sequence values requires synchronisation between all writers. In those cases, you might need to explicitly partition the ID space so that individual writers are guaranteed not to get in each others' hair.
