• Home
  • Features
  • Pricing
  • Docs
  • Announcements
  • Sign In

kubernetes-sigs / inference-perf / 40728476156
87%

Build:
DEFAULT BRANCH: main
Ran 28 Oct 2026 04:50PM UTC
Jobs 1
Files 137
Run time 1min
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst

28 Oct 2026 04:41PM UTC coverage: 85.239% (+0.1%) from 85.12%
40728476156

push

github

web-flow
feat(loadgen): bound a session-replay stage by duration (#768)

Implements #763.

## Problem

A `trace_session_replay` stage can only be bounded by session count. The
closest time-based option is the stage `timeout`, but that is a failure
mechanism, so using it to express an intended stop means:

- **The run is reported as a failure.** A stage that reaches its timeout
is recorded as `FAILED`, so a stop we asked for is indistinguishable
from something having gone wrong.
- **The sessions in flight at the cutoff drop out of the report.**
Session metrics are only built when a session completes, so those
sessions produce nothing at all. Longer sessions are the likeliest to be
running at any given moment, so the ones lost are disproportionately the
long ones.

## What this adds

- `duration` on `TraceSessionReplayLoadStage`, with a validator
requiring it to be shorter than `timeout` when both are set. If the
safety net could fire first, every duration-bounded run would report
`FAILED` and the field would silently do nothing.
- An admission check so no new session starts past the deadline, and a
loop exit that leaves the stage status `COMPLETED`.
- Sessions still running at the boundary are recorded instead of
dropped, marked `truncated: true` and counted under a new
`num_sessions_truncated`. They are deliberately kept out of the session
success/failure counts and the `session_duration_sec` percentiles — a
truncated session's events are incomplete by construction, so the normal
rule would report it as a server failure, and its duration is censored.
The requests those sessions did complete still count in the
request-level metrics.
- This accounting runs on **every** exit path, not just the duration
one: a timeout, an interrupt or an open circuit breaker drops sessions
the same way. `truncated` keeps them out of the success/failure counts
either way, so no existing metric changes meaning, though `num_sessions`
and the event totals grow for stages that previous... (continued)

211 of 229 new or added lines in 15 files covered. (92.14%)

1 existing line in 1 file now uncovered.

13097 of 15365 relevant lines covered (85.24%)

0.85 hits per line

Uncovered Changes

Lines Coverage ∆ File
8
85.59
0.22% inference_perf/loadgen/load_generator.py
6
88.5
0.82% inference_perf/datagen/replay/replay_graph_session_datagen.py
1
99.06
-0.94% inference_perf/config/config.py
1
93.72
-0.13% inference_perf/datagen/synthetic_agentic/synthetic_agentic_datagen.py
1
41.38
0.0% inference_perf/metrics/session_collector.py
1
96.77
-3.23% inference_perf/observability/context.py

Coverage Regressions

Lines Coverage ∆ File
1
85.59
0.22% inference_perf/loadgen/load_generator.py
Jobs
ID Job ID Ran Files Coverage
1 40728476156.1 28 Oct 2026 04:50PM UTC 137
85.24
GitHub Action Run
Source Files on build 40728476156
  • Tree
  • List 137
  • Changed 15
  • Source Changed 15
  • Coverage Changed 15
Coverage ∆ File Lines Relevant Covered Missed Hits/Line
  • Back to Repo
  • Github Actions Build #40728476156
  • b5af5e1c on github
  • Prev Build on main (#40727784604)
  • Next Build on main (#40729774242)
  • Delete
STATUS · Troubleshooting · Open an Issue · Sales · Support · CAREERS · ENTERPRISE · START FREE TRIAL · SCHEDULE DEMO
ANNOUNCEMENTS · TWITTER · TOS & SLA · Supported CI Services · What's a CI service? · Automated Testing

© 2026 Coveralls, Inc