Compare timings in nanoseconds for sub-microsecond precision by jaraco · Pull Request #23 · jaraco/pytest-perf

jaraco · 2026-07-13T22:31:35Z

Closes #18

Problem

Result parsed timeit's output into a datetime.timedelta, whose resolution bottoms out at one microsecond. For sub-microsecond exercises — e.g. the name-normalization work in python/importlib_metadata#533 — both the control and experiment rounded to timedelta(0), collapsing variance = (experiment - control) / control to 0/0 and rendering any regression or improvement invisible.

Fix

Parse with the new tempora.parse_nanoseconds (jaraco/tempora#49, released in tempora 5.11), keeping durations as a Decimal count of nanoseconds so the delta and variance stay meaningful at any scale. Since the values are no longer timedeltas, Result.__str__ now formats the experiment and delta with timeit's own units (nsec/usec/msec/sec) instead of timedelta's H:MM:SS rendering.

Result

The suite's own perf exercises now report meaningful sub-microsecond numbers where they previously would have been timedelta(0):

check_perf_isolated:   553 nsec  (+7 nsec, 1%)
with deps and extras:  10.5 nsec (+0 nsec, 0%)
diff_from_oh_nine_two: 2.95 nsec (-0.02 nsec, -1%)
simple test:           525 nsec  (-35 nsec, -6%)

Bumps the tempora floor to >=5.11 for parse_nanoseconds. Full suite green locally against tempora 5.11.0.

🤖 Generated with Claude Code

Result parsed timeit's output into a datetime.timedelta, whose resolution bottoms out at one microsecond. For sub-microsecond exercises -- e.g. the name-normalization work in python/importlib_metadata#533 -- both the control and experiment rounded to timedelta(0), collapsing the variance calculation to 0/0 and rendering any regression or improvement invisible. Parse instead with tempora.parse_nanoseconds, keeping durations as a Decimal count of nanoseconds so the delta and variance stay meaningful at any scale. Because the values are no longer timedeltas, format the reported experiment and delta with timeit's own units (nsec/usec/msec/sec) rather than timedelta's H:MM:SS rendering. Closes #18. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

….signed Rework the previous commit's approach in response to review. Parsing to a bare Decimal count of nanoseconds retained the needed precision but reduced a duration to a scalar with implied units, and left the unit formatting -- which isn't pytest-perf's concern -- living here. Parse instead into a tempora.Duration (5.12): a nanosecond-resolution timedelta that carries its own magnitude and units, divides by another Duration to yield the variance ratio directly, and renders itself in timeit's natural units (nsec/µsec/msec/sec). The reported delta gets its sign from jaraco.functools.signed (4.6), replacing a boolean flag on the formatter. variance returns to a try/except, and all duration formatting leaves this module. Requires tempora 5.12 (Duration) and jaraco.functools 4.6 (signed). Closes #18. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Add a doctest exercising Result: parsing, delta, variance (including the zero-control branches that return inf or 0), and rendering. Inline the signed delta formatter into __str__ rather than binding it at module scope. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>