make use of time.clock_gettime_ns()

[socketpair]

It is faster and without floating point. Just time.clock_gettime_ns() // 1000000 to get milliseconds.

[akx]

On my machine, int(time.time() * 1000.0) seems to be slightly faster than time.clock_gettime_ns(time.CLOCK_REALTIME) // 1000000:

~ $ python3 -m timeit -s 'import time' 'time.clock_gettime_ns(time.CLOCK_REALTIME) // 1000000'
5000000 loops, best of 5: 71.7 nsec per loop
~ $ python3 -m timeit -s 'import time' 'int(time.time() * 1000.0)'
5000000 loops, best of 5: 63.3 nsec per loop

or more broadly

$ hyperfine "python3 -m timeit -s 'import time' 'int(time.time() * 1000.0)'" "python3 -m timeit -s 'import time' 'time.clock_gettime_ns(time.CLOCK_REALTIME) // 1000000'"
Benchmark 1: python3 -m timeit -s 'import time' 'int(time.time() * 1000.0)'
  Time (mean ± σ):      2.182 s ±  0.010 s    [User: 2.123 s, System: 0.057 s]
  Range (min … max):    2.167 s …  2.198 s    10 runs

Benchmark 2: python3 -m timeit -s 'import time' 'time.clock_gettime_ns(time.CLOCK_REALTIME) // 1000000'
  Time (mean ± σ):      2.513 s ±  0.025 s    [User: 2.448 s, System: 0.063 s]
  Range (min … max):    2.468 s …  2.536 s    10 runs

Summary
  python3 -m timeit -s 'import time' 'int(time.time() * 1000.0)' ran
    1.15 ± 0.01 times faster than python3 -m timeit -s 'import time' 'time.clock_gettime_ns(time.CLOCK_REALTIME) // 1000000'
~ $

[socketpair]

Hm, you are right. My measurements:

In [3]: timeit int(time.time() * 1000.0)
74.1 ns ± 0.951 ns per loop (mean ± std. dev. of 7 runs, 10,000,000 loops each)

In [4]: timeit time.clock_gettime_ns(time.CLOCK_REALTIME) // 1000000
93.9 ns ± 1.09 ns per loop (mean ± std. dev. of 7 runs, 10,000,000 loops each)

In [5]: qwe=time.CLOCK_REALTIME

In [6]: timeit time.clock_gettime_ns(qwe) // 1000000
87.1 ns ± 0.426 ns per loop (mean ± std. dev. of 7 runs, 10,000,000 loops each)

[socketpair]

It seems the only cause of slowness is the passing argument for the function. But anyway: reported here: python/cpython#111482

[akx]

Sure - should we use the old micro-optimization tricks of localizing global name lookups we'd probably be a bit faster, but at the expense of not being able to e. g. easily use freezegun or other time-mocking libraries.

If you need to generate gazillions of ULIDs as fast as possible, you'd probably want to use a native-code extension instead, eg https://github.com/bdraco/ulid-transform

[akx]

Looks like we could use clock.gettime_ns() in 3.13 as python/cpython@4fe22c7 was merged.

Thanks for the upstream bug report!