A computational performance comparison of several Python libraries to generate unique identifiers.

The benchmarks are quite simple and consist of generating IDs using each library's raw interface as much as possible, e.g. the UUID module is benchmarked by measuring the execution time of the uuid.uuid4 function in the test_generate benchmark. These methods generally return a custom ID object for each implementation. Serializing this object into a string or some other primitive data time is not benchmarked for now.

Measurements are taken using Pytest's benchmark plugin.

Converting a primary data type representation of an ID into an ID object is also measured for libraries that support it in the test_parse benchmark.

• Python's own UUID module to generate v4 UUIDs with uuid.uuid4().
• KSUID implementations:
  • svix-ksuid 0.6.2 with the standard second-precision implementation
  • cyksuid 2.1.0
• python-ulid 2.2.0
• timeflake 0.4.3
• snowflake-id 1.0.2
• cuid2 2.0.1
• epyxid 0.3.3

Full results are available in here.

----------------- benchmark 'test_generate': 8 tests -----------------
Name (time in ns)                 Mean                StdDev          
----------------------------------------------------------------------
generate[epyxid]              204.0224 (1.0)         36.7382 (1.0)    
generate[snowflake]           395.8639 (1.94)        39.1105 (1.06)   
generate[cyksuid]           1,215.6683 (5.96)       301.1302 (8.20)   
generate[python-ulid]       1,815.1159 (8.90)       664.9570 (18.10)  
generate[uuid4]             2,300.7282 (11.28)      778.5865 (21.19)  
generate[timeflake]         2,333.4728 (11.44)      517.6361 (14.09)  
generate[svix]              2,264.9832 (11.10)      543.5821 (14.80)  
generate[cuid2]           54,413.7716 (266.70)   20,030.3508 (545.22) 
----------------------------------------------------------------------

The fastest library was epyxid, followed by snowflake-id and cyksuid. The rest of the libraries were within the same order of magnitude, except for cuid2 which was around 267 times slower.

--------------- benchmark 'test_parse': 6 tests ---------------
Name (time in ns)            Mean              StdDev          
---------------------------------------------------------------
parse[epyxid]            206.5733 (1.0)      102.7942 (1.0)    
parse[cyksuid]           393.3274 (1.90)     151.0228 (1.47)   
parse[uuid4]             990.0393 (4.79)     304.1820 (2.96)   
parse[snowflake]       1,176.4304 (5.69)     320.3776 (3.12)   
parse[timeflake]       2,813.7338 (13.62)    579.4188 (5.64)   
parse[svix]           17,005.8076 (82.32)  1,631.0898 (15.87)  
---------------------------------------------------------------

For libraries that allowed serializing and parsing primitive representations, results were very similar to generation.

Measurements were performed with Python 3.12.2 running on macOS 24.3.0. Hardware was a MacBook Pro with an Apple Silicon M1 chip.

To reproduce the results, install Python 3.12.2 and then install all required packages with uv sync (it is recommended to use a virtual environment created with uv venv).

Use GNU make to run the benchmark suite with

make

Alternatively, run pytest manually with pytest bench.py.

The data and scripts used to perform benchmarks are dedicated to the public domain. Libraries used in this project have their own licenses which are linked to above.