uuideal 0.2.0

Vectorcall patches for stdlib uuid powered by Rust

uuideal

Makes Python uuid fast.

import uuideal. uuideal.install(). That's it.

[!CAUTION] uuideal.install() patches CPython global state (vectorcall slots). Don't use it unless you know what you're doing. Do not call uuideal.install() from reusable packages implicitly; leave that decision to applications.

uv add uuideal

import uuid
import uuideal


same_function = uuid.uuid4
same_class = uuid.UUID

uuideal.install()  # This makes uuid fast.

assert uuideal.installed()
assert uuid.uuid4 is same_function
assert type(uuid.uuid4()) is same_class
assert uuideal.uuid7().version == 7

uuideal.uninstall()  # This makes uuid slow.

UUID generation is mostly backed by the uuid Rust crate, with two exceptions:

• On Python builds where uuid.uuid1() would generate multiprocessing safe uuids ( UUID.is_safe), uuideal will preserve that behavior.
• uuid.uuid8() will always use random.getrandbits() instead of Rust CSPRNG to generate randomness. It will remain cryptographically unsafe and compatible with seeded random module.

uuideal.uuid6(), uuideal.uuid7() and uuideal.uuid8() are also available on Python 3.12 and 3.13.

Comparison

Feature	stdlib	fastuuid	uuid_utils.compat	uuid_utils	stdlib + uuideal
Uses exact uuid.UUID	✅	✗	✅	✗	✅
No call-site changes	✅	✗	✗	✗	✅
safe uuid1() support	✅	✗	✗	✗	✅
Fast generation*	✗	✗	✗	✅	✅
Fast parsing*	✗	✅	✗	✅	✅
Beautiful codebase	✗	✅	✅	✅	✗

* At least 15x faster than stdlib for both uuid4() and uuid7() and at least 3x faster for UUID('<hex>').

Generation

Operation	stdlib	uuid_utils.compat	uuid_utils	stdlib + uuideal
uuid1()	1,752 ns	443.4 ns (3.95×)	72.5 ns (24.18×)	74.1 ns (23.64×)
uuid1() safe	737.7 ns	N/A	N/A	131.4 ns (5.62×)
uuid3()	1,007 ns	669.7 ns (1.50×)	204.6 ns (4.92×)	176.9 ns (5.69×)
uuid4()	1,195 ns	259.0 ns (4.62×)	60.5 ns (19.76×)	62.4 ns (19.17×)
uuid5()	993.7 ns	653.7 ns (1.52×)	192.5 ns (5.16×)	167.3 ns (5.94×)
uuid6()	868.5 ns	324.8 ns (2.67×)	71.4 ns (12.17×)	73.3 ns (11.84×)
uuid7()	1,445 ns	300.9 ns (4.80×)	94.8 ns (15.25×)	86.6 ns (16.70×)
uuid8()	720.1 ns	N/A	N/A	166.1 ns (4.33×)
Speedup (geomean)*	1.00×	2.85×	11.45×	12.03×

Conversions

Operation	stdlib	uuid_utils	stdlib + uuideal
UUID('<hex>')	492.8 ns	99.4 ns (4.96×)	101.2 ns (4.87×)
str(value)	351.4 ns	93.0 ns (3.78×)	58.7 ns (5.99×)
pickle.dumps(value)	1,134 ns	1,189 ns (0.95×)	1,083 ns (1.05×)
pickle.loads(payload)	829.7 ns	653.9 ns (1.27×)	700.3 ns (1.18×)
Speedup (geomean)*	1.00×	2.18×	2.45×

Access

Operation	stdlib	uuid_utils	stdlib + uuideal
value.int	40.3 ns	61.3 ns (0.66×)	40.7 ns (0.99×)
value.bytes	89.7 ns	53.8 ns (1.67×)	52.6 ns (1.71×)
value.bytes_le	324.5 ns	55.2 ns (5.88×)	52.6 ns (6.17×)
value.hex	132.2 ns	95.2 ns (1.39×)	62.0 ns (2.13×)
value.fields	334.4 ns	101.8 ns (3.29×)	98.3 ns (3.40×)
value.time_low	78.3 ns	50.0 ns (1.57×)	48.4 ns (1.62×)
value.time_mid	96.0 ns	50.2 ns (1.91×)	48.5 ns (1.98×)
value.time_hi_version	95.8 ns	50.8 ns (1.88×)	48.6 ns (1.97×)
value.clock_seq_hi_variant	96.1 ns	46.2 ns (2.08×)	44.5 ns (2.16×)
value.clock_seq_low	96.4 ns	45.6 ns (2.11×)	44.3 ns (2.18×)
value.node	78.9 ns	54.3 ns (1.45×)	53.2 ns (1.48×)
value.time	550.0 ns	54.1 ns (10.16×)	53.2 ns (10.34×)
value.clock_seq	181.5 ns	50.1 ns (3.62×)	48.8 ns (3.72×)
value.urn	413.0 ns	101.4 ns (4.08×)	64.2 ns (6.43×)
value.variant	115.1 ns	59.9 ns (1.92×)	43.6 ns (2.64×)
value.version	175.3 ns	46.1 ns (3.80×)	43.7 ns (4.01×)
value.is_safe	40.2 ns	94.3 ns (0.43×)	40.3 ns (1.00×)
sorted(values)	494,154 ns	260,985 ns (1.89×)	146,656 ns (3.37×)
Speedup (geomean)*	1.00×	2.15×	2.60×

* Geomean uses only operation groups where every displayed candidate has valid timing data.

Ran on Apple M1 macOS 15.0.1 CPython 3.14.5 best of 5 repeats after autoranging each case to at least 100ms using 2 worker process(es), nice adjusted by -20, thread QoS set to USER_INTERACTIVE.