hyperuuid

Python 3.14 finally added uuid.uuid7() to stdlib, with a real monotonic counter — genuinely well done. If you're stuck on 3.11-3.13 like most production code still is, stdlib has no v6/v7 at all, and this package gives you both today without waiting for a runtime upgrade — and on any version, the native extension below outruns stdlib outright.

RFC 9562 UUID v4 (random), v5 (deterministic), v6 and v7 (time-sortable) generation. A native extension built with PyO3 — the Rust core linked directly into the CPython extension module, no dlopen, no C-ABI hop, no ctypes marshalling, no runtime bridge.

Install

pip install hyperuuid

Real platform-specific wheels, so it lands at native speed with nothing to compile, the same way numpy or cryptography does — no compiler needed and no dependencies at all; the PyO3 extension is the package. The wheels are abi3, so one per platform covers every CPython from the 3.11 floor up — eight native ones, plus a ninth for Pyodide in the browser (below):

Only wheels are published — there is no sdist. An interpreter none of the wheels matches (a glibc older than 2.28, PyPy, free-threaded CPython) gets No matching distribution found from pip, not a source build.

On Alpine the musl wheel needs nothing beyond the base image — it carries its own copy of libgcc_s — and the whole suite passes on a bare python:3.14-alpine.

Not yet covered: free-threaded (no-GIL) CPython (3.13t/3.14t). An abi3 wheel is ignored by a free-threaded interpreter (it's a genuinely separate ABI, not a compatibility flag), so closing this gap means building and shipping additional version-specific cp313t/cp314t wheels alongside the existing ones, not just a build-flag change. PyO3 itself has supported free-threading (opt-in, gil_used = false) since 0.23; the cleaner long-term fix — PEP 803's abi3t stable ABI, one build covering both GIL and no-GIL — needs Python 3.15+, not yet released. Revisiting once that lands or free-threaded adoption justifies the extra wheel legs. The stance in the code today: the extension module does not declare gil_used = false, so a build from a checkout for a free-threaded interpreter imports with the GIL switched back on (CPython says so in a RuntimeWarning) — correct, just not parallel. And no call releases the GIL: each is nanoseconds to microseconds of pure Rust, less than the handoff would cost.

The package is typed: it ships py.typed and a stub for the extension module, so mypy and pyright see uuid.UUID and datetime.datetime rather than Any.

Usage

import uuid
import hyperuuid

hyperuuid.new_v4()
hyperuuid.new_v5(uuid.NAMESPACE_DNS, "example.com")
hyperuuid.new_v6()
id4 = hyperuuid.new_v7()

hyperuuid.v7_timestamp(id4) # recover the embedded UTC datetime.datetime
hyperuuid.v7_unix_millis(id4) # the same instant as Unix-epoch milliseconds, an int
hyperuuid.get_timestamp(id4) # None instead of assuming id4 is v6/v7
hyperuuid.is_rfc(id4, 7) # True: the RFC 9562 variant and version 7, in one call
stored = hyperuuid.v7_to_sql_order(id4) # byte order SQL Server's uniqueidentifier needs to sort by creation order
hyperuuid.v7_unix_millis(stored, hyperuuid.Layout.SQL_SERVER) # read in place, no conversion back

# One native call, one random-bytes fetch, one counter reservation for the whole batch:
batch = hyperuuid.new_v7_batch(1000)

Returns stdlib uuid.UUID objects — built through the fastuuid-style fast path (UUID.__new__ plus object.__setattr__ of the int/is_safe slots), since UUID.__init__'s own validation costs more than the entire native call; the test suite pins constructor indistinguishability so this can't silently drift from a real UUID(bytes=...) construction. For v5's namespace argument, use the RFC 9562 Section 6.6 well-known namespaces already in the standard library — uuid.NAMESPACE_DNS, NAMESPACE_URL, NAMESPACE_OID, NAMESPACE_X500 — no need for this package to redefine them. hyperuuid.NIL/MAX are the RFC 9562 §5.9/§5.10 special-value UUIDs. hyperuuid.v7_timestamp(id) recovers the embedded UTC datetime.datetime from a version 7 UUID (raises OverflowError past year 9999 — the RFC's 48-bit field holds values up to year 10889, but datetime.datetime cannot); hyperuuid.v6_timestamp(id) does the same for version 6, and can never raise that way — v6's 60-bit tick count, offset from the 1582 UUID epoch, tops out around the year 5236. hyperuuid.v6_unix_millis(id) and v7_unix_millis(id) return the same instant as an int of Unix-epoch milliseconds, with no datetime built around it — the cheaper call, and for version 7 the one that carries the whole 48-bit field. new_v6/new_v7 also accept a datetime.datetime directly in place of a raw millisecond count — converted in exact integer arithmetic and truncated to the millisecond, never rounded into the next one; a naive datetime is read as local time, the way datetime.timestamp() reads it. get_timestamp(id) is the version-agnostic counterpart to v6_timestamp/v7_timestamp — it checks the value itself, in one native call, and returns None for anything that isn't an RFC 9562 version 6 or 7 UUID, instead of assuming the caller already knows. The variant is part of that check: a 6 or 7 in the version nibble under another variant carries no timestamp. hyperuuid.new_v6_batch(count)/ new_v7_batch(count) generate count UUIDs sharing one timestamp capture and one native call, instead of count of each. hyperuuid.v7_to_sql_order(id)/ v7_from_sql_order(id) convert a version 7 UUID to and from the byte order SQL Server's uniqueidentifier needs on the wire to sort by creation order — computed once in the native Rust core rather than reimplemented in Python, and verified there (and independently against the real System.Data.SqlTypes.SqlGuid comparator in the C# binding's test suite). v6_to_sql_order(id)/ v6_from_sql_order(id) do the same for version 6, though same-millisecond v6 UUIDs aren't guaranteed to sort correctly afterward — v6 has no counter, so clock_seq/node (not the timestamp) decide ties, the same pre-existing RFC 9562 v6 limitation plain order already has.

A caller's bug is an exception: an argument of the wrong type is a TypeError; a timestamp no field can hold (negative, or past v7's 48 bits or v6's 60), a batch count outside 0 to 4294967295 (0 to MAX_V7_BATCH for v7) and a layout that is not a Layout are a ValueError; a batch too large to allocate is a MemoryError, never a batch of some other size.

Version 7 batches

A v7 batch (new_v7_batch, fill_v7) is always in strictly increasing order. The 26-bit counter that orders UUIDs within a millisecond is one process-wide sequence that wraps every 2^26 values, so a batch can straddle the wrap; the UUIDs from the wrap on carry unix_millis + 1 rather than sorting before the ones ahead of them, so an embedded timestamp can be one millisecond past the one supplied, never more. For the same reason one batch holds at most hyperuuid.MAX_V7_BATCH (67,108,864) UUIDs: a larger count, or a fill_v7 buffer of more than that many, is a ValueError naming the limit, raised before anything is allocated or written. Version 6 has no counter and no such limit.

The roll-forward orders one batch, not the stream. The next batch or new_v7 call in the same real millisecond starts its counter just past the wrap and carries the supplied timestamp, so it sorts before the previous batch's tail, which was stamped a millisecond later; two single new_v7 calls either side of the wrap in one millisecond sort in reverse the same way. It happens at most once per MAX_V7_BATCH UUIDs the process mints.

Inspecting a UUID

import hyperuuid
from hyperuuid import Layout, Variant

hyperuuid.version(id4)                     # 7: the version nibble, 0-15
hyperuuid.variant(id4)                     # Variant.RFC9562
hyperuuid.is_rfc(id4, 7)                   # True: the guard before trusting v7 fields

stored = hyperuuid.v7_to_sql_order(id4)
hyperuuid.version(stored, Layout.SQL_SERVER)       # 7
hyperuuid.is_rfc(stored, 7, Layout.SQL_SERVER)     # True
hyperuuid.get_timestamp(stored, Layout.SQL_SERVER) # the embedded datetime, read in place

version, variant and is_rfc are answered by the native core, the same functions every other binding calls, so no bit-reading is duplicated here. version(id) is the version nibble whatever the variant (stdlib's id.version is None for a non-RFC variant; this is not), Nil reads as 0 and Max as 15, and is_rfc(id, n) is the check that means "an RFC 9562 UUID of version n" — False, never an error, for an n outside 0 to 15. variant(id) is a Variant — NCS (Nil included), RFC9562, MICROSOFT or FUTURE (Max included) — read in RFC 9562 order only. It has no layout form, because in SQL Server order the variant sits at a different byte for each version, so don't hand it a SQL-ordered value or a uniqueidentifier read-back; is_rfc(id, n, Layout.SQL_SERVER) is the check for one of those, and it tests the variant where that version puts it.

Layout says which byte order a uuid.UUID is held in: Layout.RFC9562, what every other function here takes and returns and the default everywhere, or Layout.SQL_SERVER, the order v6_to_sql_order/v7_to_sql_order return (and a uniqueidentifier column holds). version, is_rfc, v6_unix_millis/v7_unix_millis, v6_timestamp/v7_timestamp and get_timestamp all take one as their last argument, and read a SQL-ordered value's permuted bytes in place, with no conversion back first. In Layout.SQL_SERVER only versions 6 and 7 have an order: version is 6 or 7 for bytes that form a SQL-ordered v6 or v7, and 0 for bytes that don't, and the core checks the variant where each version puts it, so a SQL-ordered v6 never reads as a v7 or the other way round. The bytes alone cannot say which layout a value is in — an RFC-ordered random v4 forms a valid SQL-ordered v7 one time in 16 — so the caller must keep track of the layout, carry it with the value, and pass it at every call site that can see a SQL-ordered value. With the default layout, version and is_rfc read RFC order, so is_rfc(sql_ordered, 7) reads the wrong bytes; there is deliberately no layout-agnostic check.

Layout and Variant are IntEnums holding the native core's codes (Layout.RFC9562 == 1, Layout.SQL_SERVER == 2; Variant.NCS == 1 through Variant.FUTURE == 4), and a plain int equal to a layout code is accepted where a Layout is. Neither has an "unspecified" member: the layout defaults to Layout.RFC9562, and anything else — an unknown code is a ValueError, a non-integer a TypeError — is refused rather than guessed at. There is no raw-bytes form; wrap 16 bytes in uuid.UUID(bytes=...) first, exactly as they are held.

hyperuuid.native_version() names the core actually loaded, "major.minor.patch", decoded from the same packed hyperuuid_version export every other binding probes. hyperuuid.BACKEND is always "native", the PyO3 extension.

In the browser (Pyodide)

The same PyO3 extension, compiled for Pyodide's Emscripten target, is published to PyPI as hyperuuid-X.Y.Z-cp311-abi3-pyemscripten_2026_0_wasm32.whl (about 74 KB), so micropip finds it the way pip finds the native wheels:

import micropip
await micropip.install("hyperuuid")

import uuid, hyperuuid
hyperuuid.new_v7()
hyperuuid.new_v5(uuid.NAMESPACE_DNS, "example.com")

It is the native backend (hyperuuid.BACKEND == "native"), with the same API and no JavaScript bridge: randomness comes from Emscripten's getentropy (the browser's crypto.getRandomValues), and the clock is Date.now(), so v6/v7 timestamps have millisecond resolution there, as everywhere. CI installs the wheel into Pyodide and runs this package's whole pytest suite in it twice, under Node and in headless Chrome.

The wheel is for Pyodide 314.x (Python 3.14, platform pyemscripten_2026_0). It is abi3, but a Pyodide ABI is one Python minor built with one exact Emscripten, so each Pyodide ABI year needs its own wheel: earlier Pyodide lines (0.29.x and older) find no wheel, and the next line is covered by the release that adds its build.

Bulk generation into a buffer

fill_v7(buffer) and fill_v6(buffer) write raw RFC 9562-ordered bytes — 16 per UUID — straight into a bytearray you already own, in one native call, without constructing a single uuid.UUID:

import hyperuuid

buf = bytearray(1000 * 16)
hyperuuid.fill_v7(buf)                      # 1000 v7 UUIDs, one timestamp capture
first = bytes(buf[0:16])                    # ready for a BYTEA / uniqueidentifier parameter

This is roughly 35x faster than new_v7_batch, and the reason is worth understanding, because it decides whether you should use it at all:

new_v7_batch does not spend its time in the native call — it spends it building a thousand uuid.UUID instances. Skip that and Python lands at 3.7 µs, which is the same native ceiling the Go (3.6 µs) and C# (3.7 µs) bindings hit for identical work.

The third row is the catch, and it inverts the advice: if you need uuid.UUID objects, keep using new_v7_batch. Filling bytes and constructing UUIDs from them in Python is more than six times as slow, because the extension builds them through a much faster path internally than you can from Python. Reach for fill_v7 only when bytes are the destination — a database parameter, a wire format, a bulk COPY — not a step on the way to objects.

len(buffer) must be a multiple of 16, a zero-length buffer writes nothing, and fill_v7 takes at most MAX_V7_BATCH UUIDs' worth (above). Both functions take an optional datetime or Unix-epoch millisecond timestamp, same as the rest of the API.

One deliberate limitation: these take a bytearray, not any writable buffer. Supporting memoryview, mmap or NumPy arrays needs Py_buffer, which entered CPython's stable ABI in 3.11. That is this extension's floor as of this release (abi3-py311), so the wider buffer protocol is no longer ruled out; it is simply not built yet.

Why not stdlib uuid?

This is the one binding where the honest answer genuinely depends on which Python you're running — this package supports 3.11+, and stdlib's own v6/v7 story changed dramatically partway through that range:

• Python 3.11-3.13: stdlib has uuid1/uuid3/uuid4/uuid5 — no v6, no v7, at all. This package is the only way to get either without a third-party dependency, and the native extension outruns stdlib's v4/v5 on top of that.
• Python 3.14+: stdlib added uuid.uuid6()/uuid7()/uuid8(), and uuid7() genuinely implements RFC 9562 §6.2's monotonic counter (42 bits of it) — this isn't a naive random-bits implementation, it's a real, well-built addition, and it's what the benchmarks below measure against. This package still wins across the board there — see Benchmarks — plus:
  1. Cross-language consistency. The same Rust core mints v5 namespace UUIDs for Go, C#, Ruby, and every other binding in this repo — verified in CI to match stdlib's own uuid.uuid5 byte-for-byte. If your stack isn't Python-only, or you need every service minting IDs from the literal same engine rather than N independent (if individually correct) implementations, that's not something stdlib can offer regardless of version.
  2. Batch generation. new_v7_batch(count) shares one timestamp capture, one random-bytes fetch, and one counter reservation across the whole batch — stdlib's uuid7() has no bulk-generation entry point, so a loop of individual calls is the only option there.
  3. One behavior across your whole supported range. If your package needs to run on 3.11 and 3.14, this avoids sys.version_info-gated code paths for v6/v7 support.

Benchmarks

Measured with pyperf (linux-x64 on an Intel Core i9-11900H, CPython 3.14.7 as Fedora builds it, python bench_uuid.py --fast; see bench_uuid.py). Linking the core directly into the extension module — no ctypes boundary to cross — puts every generator ahead of stdlib's own C-accelerated implementations:

Most of what a call costs is the uuid.UUID it hands back, so the extension builds that through the C API's own entry points — an instance allocated and its two slots set, no Python-level call — all of it inside the stable ABI, so one wheel still covers every CPython from 3.11.

Batch generation amortizes the rest of the per-call cost:

Timestamp extraction vs. stdlib's .time property

CPython 3.14's uuid.UUID.time has real version-aware extraction logic of its own (branches on version, computes the right thing for v6/v7, not just a v1-only stub), so this is a genuine head-to-head — each call measured against a UUID generated once outside the timed loop, so only the extraction itself is timed:

.time returns an int, so the *_unix_millis rows are the like-for-like ones; the *_timestamp rows hand back a timezone-aware datetime and are still ahead. Reach for the integer form when the value is going to be stored, compared or forwarded as a number.

Worth noting: stdlib's .time for v6 returns raw Gregorian-epoch 100ns ticks, not Unix milliseconds like hyperuuid.v6_timestamp — different units if you actually need the value, but a fair timing comparison of "the cost of pulling the embedded time out" either way.

Reproduce: maturin develop --release (from python/, inside a virtualenv — pyproject.toml already points maturin at ../rust/Cargo.toml and the python feature) to build the release extension — pip install -e ".[bench]" alone builds debug by default and will understate every number above — then pip install pyperf and python bench_uuid.py --fast -o results.json.

Verifying provenance

Every wheel PyPI serves, the Pyodide one included, carries a GitHub build-provenance attestation. The wheels are built, installed and attested in CI by the shared hyper-build-wheels.yml workflow in SkunkWerkx/.github, and release.yml verifies each one before publishing it unchanged, so the verify command names that signer:

pip download hyperuuid==X.Y.Z --no-deps -d .
gh attestation verify hyperuuid-X.Y.Z-*.whl \
  --repo SkunkWerkx/HyperUuid --signer-repo SkunkWerkx/.github
# or: gh attestation verify hyperuuid-X.Y.Z-*.whl --owner SkunkWerkx

This is a separate thing from the PEP 740 attestations gh-action-pypi-publish already sends to PyPI itself, which PyPI-side tooling checks on its own — this is the GitHub/Sigstore transparency-log route, checked with gh attestation verify, the same route every other artifact in this project uses. See csharp/README.md's provenance section for why an artifact signed by the shared workflow needs --signer-repo.

License

MIT