New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
v7 should have a counter option? #717
Comments
Hi @rogusdev 👋 I'd be keen to go catch up on some of the discussion around that. Those sections look like recommendations rather than requirements, so I think It seems a bit overkill to me to need a clock, a source of randomness, and a monotonic counter to generate V7 UUIDs, but we could consider adding |
The test vector for UUIDv7 appears to use a fully random value for I think we should consider supporting this, but don't think we need to block stabilizing the current APIs on it. |
From my perspective, while the main point of uuid v7 is to put the clock at the front, so that db indexes can btree on them properly, counters add additional safety mechanisms that are quite to my taste: if you are generating ids in a single thread, you are guaranteed that no duplicates can happen, if there are less than the counter limit per millisecond. While the degree of randomness involved otherwise is certainly extremely unlikely to have duplicates, I am a big fan of guarantees. Which is in fact why they put it into the RFC, as that is also a feature in many other popular id libraries. |
That said, "SHOULD" is indeed a recommendation, rather than requirement, and they list multiple alternatives. So I support having a separate set of ctor functions to call for the counter version(s). |
Hi Ashley, I'm Sergey Prokhorenko, a contributor to rfc4122bis and a counter enthusiast. Some developers implement the counter:
UUIDv7 with the counter for PostgreSQL is currently being developed in C by Andrey Borodin: But it seems to me that Rust is also a good tool for such development. |
We've currently actually got access to a counter in our
An alternative for 2. would be to add |
Like the vast majority, I am convinced that only the seventh version is worthy of attention. There is no need to waste time and effort implementing other versions.
I'm against. A counter initialized to zero increases the collision probability. Worse, the counter value may be the same as the random value used instead of the counter at the beginning of the millisecond. I prefer a counter that is initialized to a random value at the beginning of each millisecond. But this may reduce the actual capacity of the counter. Therefore, the leftmost bit of the counter should be initialized to zero and/or the timestamp should be incremented when the counter overflows.
The timestamp in the seventh version is shorter, and you missed the ver and var segments.
If we are talking about initializing the counter every millisecond with a random value and incrementing within a millisecond, then I am for it. |
This library already implements |
I highly recommend looking at the UUIDv7 implementation in PostgreSQL v17 by Andrey Borodin |
Thanks @sergeyprokhorenko 👍 Having a good reference implementation to point at will definitely be helpful |
Another good implementation: It is mentioned in this benchmark. Due to the excessively long counter (initialized with a random number every millisecond), less entropy is generated, which requires a lot of resources. But it's worth trying to replace rand::rngs::OsRng with openssl-rand for better performance. |
This patch has already been successfully tested: https://ardentperf.com/2024/02/03/uuid-benchmark-war/ |
Thought/Question: Why not add a builder pattern with defaults? That doesn't eliminate the backwards compatibility and can improve clarity around generating the UUID. |
The API in the Rust ULID crate provides monotonicity using an If you know that you are generating a batch then an Here's is an example: use ulid::Ulid;
use uuid::Uuid;
fn main() {
let mut uuids = Vec::with_capacity(10);
for _ in 0..10 {
let uuid = Uuid::now_v7();
uuids.push(uuid);
}
for uuid in uuids.iter() {
println!("{:<12} {uuid}", "uuidv7:");
}
println!("");
let mut ulids = Vec::with_capacity(10);
let mut ulid = Ulid::new();
for _ in 0..10 {
ulids.push(ulid);
ulid = ulid.increment().unwrap();
}
for ulid in ulids.iter() {
println!("{:<12} {ulid}", "ulid:");
}
} Is taking blatant inspiration from the excellent work in the ULID crate ok? Of course. ULID is the very first referred to time-based unique ID referenced in the new RFC. Note that this also gives more control over creation of batches. If one creates a batch of 1000 UUIDs, using an internal counter you have no control over whether they will be in the same millisecond. This means you will have part of the batch spread across different timestamps and the random component will jump around for each millisecond. Using an increment you can keep the millisecond and the random component stable. It's not required in the specification but seems useful to be able to trivially see if UUIDs are part of the same batch. |
This is not true |
New functions for generating UUIDv7 with a counter in the ClickHouse DBMS: link The structure is exactly the same as https://crates.io/crates/uuid7 and https://www.npmjs.com/package/uuidv7 |
I've started working on an implementation over in #755 that supports respecting the counter value when constructing v7 UUIDs in a way that lets a caller decide how wide they want that counter to be so the guarantees about ordering and uniqueness are configurable. It's still just a draft so will ping once it's ready. |
How might you allow the caller to control whether the UUIDs are generated for the same millisecond if the timer is internal to the uuid generator? |
UUIDs generated in different milliseconds have different timestamp values. UUIDs generated at the same millisecond have the same timestamp values. |
RFC9562 "Universally Unique IDentifiers (UUID)" has already been published and entered into force. Differences between the implementation for DBMS ClickHouse and the implementation for DBMS PostgreSQL:
In both implementations in case the counter overflows, the timestamp field is incremented by 1 and the counter is reset to a random new start value (guard bit to zero). I would also recommend to add the ability to shift the timestamp by the value specified by the formal parameter
This is allowed by RFC9562:
It is quite difficult to implement the exclusion (in the unpredictable future) of leap seconds required by RFC9562:
|
In order to maintain the invariant that UUIDs are sortable by the order they’re generated in this would need to increment the timestamp for all subsequently generated UUIDs in that millisecond, right? |
Yes that's right. Moreover, this may continue for several consecutive milliseconds of high generation rate until the real time catches up with the timestamp. But with a guard bit and a counter 18 bits or longer, a counter overflow is an impossible event (until a quantum computer came along). |
UUIDv7 extension written in RUST, implemented in PostgreSQL |
Long story short, I would propose that the only “safe” option for UUID v7 is to use ALL of its Here is a detailed explaination: Due to the Birthday Problem , the length of random bits is crucial to its security requirements, and the frequency matters too. The approximation formula goes like: Where We can see that For example, our UUID v4 has 122 random bits in total, which means its "random space size" is ULID has 80 random bits in total, but it's 1ms separated. Its chance of having a collision in 1ms is about UUID v7 has 74 random bits, and is also 1ms delimited. Its chance is From the calculation above we may already understand that the random bits are super important, so we should not omit any of them, and we should limit the number of randomizations. Suppose that we are using the Then, given that we are generating Thus, in a year we only have If overflow happens, I would say by default we may simply let it overflow and then trim off the new 0b1 bit, but we do need to provide an option to This approach is not entirely unguessable, but still conforms to the "Monotonic Random" method in RFC 9562 Section 6.2. It is also the tradition how ULID handles monotonic counters while balancing unguessability. Additionally, it would be purely an evil and insecure approach to add up a random value other than 1, because this will be a significant damage of the random bit size and make the entire method vulnerable to collisions. A |
All of these calculations make the mistaken implicit assumption that all generated UUIDs end up in a common database table where they can collide. The absurdity of these calculations can be confirmed by the following example: UUIDv7s with a counter generated by one generator will never collide, even if random parts are completely removed from them. This is just as true as integer keys generated by autoincrement do not collide. The random part is necessary for the uniqueness of UUIDs generated by different generators (when merging database tables or at distributed UUID generation). By the way, the counter is initialized with a random number every millisecond. Therefore, it is incorrect to say that the counter reduces the random part. |
This is not a mistaken assumption - the collision resistance analysis itself is based on how many UUIDs are generated in parallel and put together into a single storage collection, it won't exist if mitigations could be taken in downstream services. This is a too-good-to-be-true implication that I dare not to assume.
This is also the most ideal scenario. In the real world we usually have multiple workers and pods working together, and some of them are potentially lack of time synchronizations. In the worst case, UUIDs are actually generated on mobile/wasm client side - which means collision is still a thing to carefully consider even with the counter. Essentially I believe we are making a very similar proposal - all This however is not all methods RFC 9562 suggested - I was trying to analyze all methods in the RFC and conclude the best option we should implement. Here are all methods mentioned and accepted in the RFC:
|
The authors of RFC 9562 did not intend to reject any of the contributors' proposals, but simply to combine all of these proposals. This does not mean that all these proposals were justified. But in real implementations Method 1 is used, which is not at all divided into two opposing strategies, but is single. Your proposal differs only in the absence of the random part generated for each individual UUID. It kills the unguessability without giving anything in return. |
True. Utilizing high precision timestamp in the counter is not a good option at all, I wouldn't prefer this strategy either in real world implementations.
I believe it is the nature of using a counter. No matter how we decide to add per-UUID random bits, it will always put exponential damages to the collision resistance, even just a small amount of bits. As most UUID users use V4 nowadays, they really need the same or at least equivalent collision resistance over years and billions of generations. In real world scenarios, having randomized entrypoint for each millisecond still provides basic unguessibility. My proposal is exactly how ULID handle counters, and UUID v7 has pretty much the same structure compared to ULID - which means, using all random bits as a randomly initialized plus-one counter is a battle-tested production proof approach. And in fact, predictability won't be not a concern if you choose not to use the counter at all, and it truly maximized the level of unguessibility. This is also my recommendation in scenarios where sub-millisecond predictability really matters more than avoiding collisions. Using a counter should always be optional, because there are simply no silver bullets to cope both without compromises. |
Per the new draft "With this method rand_a section of UUIDv7 SHOULD be utilized as fixed-length dedicated counter bits that are incremented by one for every UUID generation."
That's in the "Fixed-Length Dedicated Counter Bits (Method 1)" section under https://datatracker.ietf.org/doc/html/draft-peabody-dispatch-new-uuid-format-04#monotonicity_counters which is directly linked from https://datatracker.ietf.org/doc/html/draft-peabody-dispatch-new-uuid-format-04#section-5.2 under the rand_a explanation.
To my reading, that means it should have a counter, like v6 (and v1), but instead the current v7 builder implementation https://docs.rs/uuid/latest/src/uuid/v7.rs.html#47-53 is just random for rand_a. Am I missing something on this one? If not, and this is a desirable feature, I might look into putting up a PR to add the counter support.
The text was updated successfully, but these errors were encountered: