1
86%
main: 86%

Ran 19 Oct 2026 10:04AM UTC

Files 135

Run time 2s

Badge

Embed ▾

Committed 19 Oct 2026 09:57AM UTC coverage: 83.351% (-0.01%) from 83.364%

Job # 39233880775.1

Build Type

push

github

Committed by

web-flow

Commit Message

fix(loadgen): avoid holding worker permits during dependency waits (#901)

## Summary

Make `worker_max_concurrency` limit requests executing against the model
server, rather than also counting session events waiting on schedules or
predecessors.

## Changes

- Acquire session-replay worker permits only after schedule and
dependency waits.
- Preserve pre-dequeue admission behavior for non-session workloads;
apply the post-acquire teardown safeguard to all workloads.
- Recheck teardown signals after permit acquisition to prevent late
dispatch.
- Update trace-replay documentation.
- Add regressions for dependency-wait starvation and teardown races.

## Validation

- `pdm run validate`
- Focused semaphore regression tests
- Full test package: 2,108 passed in the sandbox; remaining failures
were environment/socket or unrelated terminal-rendering tests.
- Two independent reviews found no blocker or high-severity issue.

Known follow-up: dependency-aware admission for extremely large session
task backlogs.

Coverage Stats

12115 of 14535 relevant lines covered (83.35%)

0.83 hits per line

kubernetes-sigs / inference-perf / 39233880775 / 1
86%
main: 86%

README BADGES
x

Markdown

Textile

RDoc

HTML

Rst

Source Files on job 39233880775.1

kubernetes-sigs / inference-perf / 39233880775 / 1 86% main: 86%

README BADGES x

Markdown

Textile

RDoc

HTML

Rst

Source Files on job 39233880775.1

kubernetes-sigs / inference-perf / 39233880775 / 1
86%
main: 86%

README BADGES
x