⚠️⚠️ UPDATE 2021: This proposal is now being pursued at: https://github.com/WICG/uuid ⚠️⚠️
Status: Stage 1
UUID generation is an extremely common software requirement
The ubiquitous nature of the
uuid module demonstrates that UUID generation is a common
Developers "re-inventing the wheel" is potentially harmful
Developers who have not been exposed to RFC 4122 might naturally opt to invent their own approaches
to UUID generation, potentially using
Math.random() (in TIFU by using
there's an in-depth discussion of why a Cryptographically-Secure-Pseudo-Random-Number-Generator
(CSPRNG) should be used when generating UUIDs).
Introducing a UUID standard library, which dictates that a CSPRNG must be used, helps protect developers from security pitfalls.
The UUID standard library provides an API for generating RFC 4122 identifiers.
The only export of the UUID library that is initially supported is
randomUUID(), a method which
version 4 "Algorithm for Creating a UUID from Truly Random or Pseudo-Random Numbers",
and returns the string representation (as described in RFC-4122).
// We're not yet certain as to how the API will be accessed (whether it's in the global, or a // future built-in module), and this will be part of the investigative process as we continue // working on the proposal. randomUUID(); // "52e6953d-edbe-4953-be2e-65ed3836b2f0"
Math.getRandomValues() exposes an identical API to the
recommendation. With the same guarantees, regarding the quality of randomness:
Implementations should generate cryptographically random values using well-established cryptographic pseudo-random number generators seeded with high-quality entropy, such as from an operating-system entropy source (e.g., "/dev/urandom"). This specification provides no lower-bound on the information theoretic entropy present in cryptographically random values, but implementations should make a best effort to provide as much entropy as practicable.
Math.getRandomValues() will act as the foundation for implementing UUID algorithms, providing a
single mockable (see #25) source of randomness.
Out of scope
Algorithms described in RFC 4122 other than version 4 are not initially supported.
Statistics we've collected (see analysis/README.md) indicate that the version 4 algorithm is most widely used:
|Algorithm Version||Repo Count||%||Weighted by Watch Count||%|
Regarding other UUID versions
While there is utility in other UUID versions, we are advocating starting with a minimal API surface that supports a large percentage of users (the string representation of version 4 UUIDs).
If research and/or user feedback later indicates that additional functionality, such as versions 1, 3, and 5 UUIDs, would add value, this proposal does not preclude these additions.
How do folks in the community use RFC 4122 UUIDs?
Creating unique keys for database entries
Generating fake testing data
Writing to temporary files
What are the advantages to uuid being in the standard library?
uuidmodule is relied on by
> 2,600,000repos on GitHub (June 2019). Guaranteeing a secure, consistent, well-maintained UUID implementation provides value to millions of developers.
- The 12 kb
uuidmodule is downloaded from npm
> 62,000,000times a month (June 2019); making it available in the standard library eventually saves TBs of bandwidth globally. If we continue to address user needs, such as
uuid, with the standard library, bandwidth savings add up.
How unique are v4 UUIDs?
If you ignore the challenges involved in random number generation, then v4 UUIDs are unique enough for all but the most stringent use cases. For example, the odds of a collision among 3.3 quadrillion version 4 UUIDs (equivalent to generating a million UUIDs/second for 104 years) is roughly one in a million (p = 0.000001). Source.
That said, the quality of the random number generator is vital to uniqueness. Flawed RNG implementations have led to UUID collisions in real-world systems. It is for this reason that this spec mandates that any random numbers used come from a "cryptographically secure" source, thereby (hopefully) avoiding such issues.
Why call the export
randomUUID() and not something like
UUIDs defined in IETF RFC 4122 are 128 bit numbers that follow a specific byte layout. All of them contain a "version" field comprising 4 bits and a "variant" field comprising 2 bits, meaning that 6 out of 128 bits are reserved for meta information.
v4 UUIDs are defined to have all remaining 122 bits set to random values, there cannot be
another UUID version that would contain more randomness.
While any name involving
v4 requires a rather deep understanding of the intricate meaning of the
term "version" in the context of the UUID spec, the term
randomUUID() appears to be much more
Aren't v1 UUIDs better because they are guaranteed to be unique?
As an oversimplification,
v1 UUIDs consist of two parts: A high-precision
timestamp and a
node id. IETF RFC 4122 contains several requirements that are supposed to ensure that
v1 UUIDs are unique.
- The timestamp has 100 nanosecond resolution and implementations are
required to throw an error or stall on
attempts to generate UUIDs at a rate higher than 10M/second on a single
node. Realistically that's only enforceable within a single thread/process on a single host. Enforcing this across multiple processes / hosts requires non-trivial architectures that run counter to the main thesis the UUID spec: "One of the main reasons for using UUIDs is that no centralized authority is required to administer them".
- The mechanism for generating
nodevalues preferred by the RFC is to use the host system's IEEE 802 MAC address. This made sense back when the RFC was authored and MAC addresses could reasonably be expected to be unique, but this is arguably no longer the case, not with the proliferation of virtual machines and containers where MAC addresses may not be unique by design.
So in practice, modern implementations will generate a random 48 bit
node value each time a
process is started leaving a probability of 1 in 248 for collisions in the
In the unlikely event of such a collision
it would take only 75 milliseconds
for a duplicate
v1 UUID to appear when generating UUIDs at a rate of 1M/second. So while also
unlikely, just like with
v4 UUIDs there is no practical guarantee
v1 UUIDs are unique.
Are there privacy concerns related to v1 UUIDs?
If implementations follow
the primary recommendations of RFC 4122 then
v1 UUIDs would indeed leak the hardware MAC address of the machine where they are being created.
As discussed above this
are either unavailable (browser, serverless functions) or not necessarily unique
(containers). However, there are
rumors that the presence of the MAC address lead to the arrest of the authors of the Melissa Virus
and according to the manual even
MySQL 8.0 still uses the hardware MAC address on some operating systems.
In any case the exact creation time of any
v1 UUID will be contained within the UUID. This alone
can be a privacy or data protection concern for many use cases (e.g. leaking the creation timestamp
of a user account) so it's yet another reason to be very careful when choosing to use
How do other languages/libraries deal with UUIDs?
Apart from that, UUID adoption across other languages/libraries seems to be rather inconsistent:
- deno added UUID to their standard
library, leaving out
v3. The code for UUID creation is essentially copied from the
uuidnpm module, hence method naming follows the
- Java provides methods for
UUID.randomUUID()) UUIDs but not
v5. It would be interesting to investigate further as to why these algorithms were chosen, given that on the one hand time-based UUIDs (
v1) appear to have much broader use than name-based (
v5) UUIDs and that on the other hand for name-based UUIDs the RFC already recommends
- C++ Boost
v3for name-based UUIDs but in its implementation anticipates that
v5(which uses SHA-1 for hashing) will be followed up by a newer name-based UUID version which will use a different hashing algorithm ("In anticipation of a new RFC for uuid arriving…").
- Google's implementation for go has chosen
v1to be the "default" export whose generator method is called
NewUUID(), whereas the other exposed methods have names closer to the abstraction we propose:
- Python provides methods for generating UUIDs named
after the version for all 4 versions (
uuid.uuid5()) plus a
UUIDclass to represent UUIDs and transform them into various representations.
- Rust provides methods for generating UUIDs named after the
version for all 4 versions (
Uuid::new_v5()) as static members of a
Uuidclass which is used to represent UUIDs and transform them into various representations.
- Identify champion to advance addition (stage-1)
- Prose outlining the problem or need and general shape of the solution (stage-1)
- Illustrative examples of usage (stage-1)
- High-level API (stage-1)
- Initial spec text (stage-2)
- Babel plugin (stage-2)
- Finalize and reviewer sign-off for spec text (stage-3)
- Test262 acceptance tests (stage-4)
- tc39/ecma262 pull request with integrated spec text (stage-4)
- Reviewer sign-off (stage-4)