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

kubernetes-sigs / inference-perf / 42755379419
87%

Build:
DEFAULT BRANCH: main
Ran 10 Nov 2026 04:33AM UTC
Jobs 1
Files 141
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

10 Nov 2026 04:21AM UTC coverage: 86.269% (+0.3%) from 86.011%
42755379419

push

github

web-flow
Report requests cancelled at stage teardown (#859)

Fixes #858

Old behavior: a request still in flight when
`stage_teardown_grace_seconds` ran out was cancelled and recorded
nowhere. Not a success, not a failure, not in the report. New behavior:
it is recorded as its own outcome, `cancelled`, beside `successes` and
`failures`. The server neither answered nor refused it, so it counts as
neither.

Why not a failure: #768 reports a planned `duration` stop as COMPLETED
and keeps the sessions it cuts short out of the session failure counts.
Counting their in-flight requests as failures would put server failures
back at the request level, for a stop the config asked for.

Changes:
* `process_request` catches `CancelledError`, records the request with
error type `StageTeardownCancelled`, then re-raises. Covers a cancel
during an attempt and during a retry backoff. The mock client does the
same.
* `summarize_requests` partitions requests three ways. The `cancelled`
block carries `count`, `in_flight_sec` (how long each had been waiting,
a lower bound on its latency) and `prompt_tokens`. Omitted when nothing
was cancelled, the way `retries` is, so an unaffected report is
unchanged.
* `failures.count`, `failures.by_label`, the error rate and every
latency distribution exclude cancelled requests.
* A cancelled request never becomes its session's error, in the
collector and in the report generator.
* `inference_perf_requests_total` gets `status="cancelled"`.
`inference_perf_request_errors_total` does not count it.
* A cancelled request's OTel span gets no status and
`inference_perf.request.cancelled = true`, not ERROR.
* The llm-d benchmark report partial counts cancelled requests under
`requests.incomplete`, not `requests.failures`. Unset when nothing was
cancelled.
* Report validation knows the `cancelled` block. The count checks add it
to successes and failures, and the partial's `incomplete` must match
`cancelled.count`.
* The CLI error table gets a "Cancelle... (continued)

73 of 74 new or added lines in 10 files covered. (98.65%)

155 existing lines in 8 files now uncovered.

13703 of 15884 relevant lines covered (86.27%)

0.86 hits per line

Uncovered Changes

Lines Coverage ∆ File
1
43.33
1.95% inference_perf/metrics/session_collector.py

Coverage Regressions

Lines Coverage ∆ File
77
89.89
1.35% inference_perf/datagen/replay/replay_graph_session_datagen.py
31
95.25
-0.05% inference_perf/reportgen/base.py
15
92.31
1.48% inference_perf/apis/anthropic_messages.py
9
92.75
0.51% inference_perf/apis/base.py
9
96.67
0.05% inference_perf/apis/chat.py
9
93.79
0.36% inference_perf/reportgen/summary/validation.py
4
95.6
0.31% inference_perf/reportgen/per_request/validation.py
1
43.33
1.95% inference_perf/metrics/session_collector.py
Jobs
ID Job ID Ran Files Coverage
1 42755379419.1 10 Nov 2026 04:33AM UTC 141
86.27
GitHub Action Run
Source Files on build 42755379419
  • Tree
  • List 141
  • Changed 19
  • Source Changed 18
  • Coverage Changed 19
Coverage ∆ File Lines Relevant Covered Missed Hits/Line
  • Back to Repo
  • Github Actions Build #42755379419
  • 68d55da1 on github
  • Prev Build on main (#42664894421)
  • Next Build on main (#42889203792)
  • 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