kubernetes-sigs / inference-perf

79%

push

github

fix: preserve microsecond precision in Azure trace replay (#782)

Azure trace replay truncates fractional timestamps to two digits.
Requests at 1, 2, 9 and 11 milliseconds are therefore scheduled at
relative offsets `[0, 0, 0, 0.01]`, collapsing distinct arrivals into a
burst instead of preserving `[0, 0.001, 0.008, 0.010]`.

Retain up to six fractional digits, matching datetime's microsecond
resolution. Keep the existing accepted quoted, UTC `T`/`Z`, whole-second
and shorter-fraction formats, and truncate higher precision only at the
microsecond boundary.

The regression test loads an actual CSV through AzurePublicDatasetReader
and TraceReplayLoadTimer, checking millisecond, microsecond and
cross-second offsets. Both new tests fail before the fix and pass
afterward.

Validation:
- Trace replay tests: 5 passed.
- Full suite: 1365 passed, 7 skipped, and one analysis test initially
failed because the optional matplotlib dependency was absent. After
installing the analysis dependency, that test passed (1366 total passing
tests).
- Coverage regression check: 77.83% versus 77.80% on the same main
commit; both the absolute floor and delta check passed.
- Strict mypy: 260 source files passed.
- Ruff 0.15.20 (matching pdm.lock), formatting and CLI flag
documentation sync passed.
- Manual benchmark: real CLI replays a four-row Azure timing trace
against a local HTTP completion endpoint; all 4 requests succeed. Exact
submillisecond scheduling offsets are asserted in the timer regression
test rather than relying on wall-clock network timing.

---------

Signed-off-by: git-jxj <65210887+git-jxj@users.noreply.github.com>

5 of 5 new or added lines in 1 file covered. (100.0%)

9426 of 11945 relevant lines covered (78.91%)

0.79 hits per line