sglang-ci-stats

Auto-scraped per-test elapsed-time history for sgl-project/sglang's PR Test workflow.

What's here

Path	Contents
runs/<YYYY-MM-DD>T<HH-MM-SS>Z__<run_id>.json	Per-run archive. One file per scraped CI run; jobs grow append-only as reruns succeed. Each file also carries a time_stats block (this run's wall-clock data point).
model.json	Derived partition model: per-(file, suite, backend) est (p90) and per-suite (coeff, bias, r_squared, method) fitting wall_clock = coeff * sum(elapsed) + bias. OLS is emitted only when identifiable and explanatory (coeff > 0, r_squared >= 0.5); otherwise the constant-overhead model (coeff=1.0, bias=median(wall - sum_elapsed), method="overhead"). Deterministic function of runs/ within the same UTC day. git log model.json is the historical archive.
scrape.py	The scraper. Pulls completed PR Test runs on main (events: schedule + workflow_dispatch) within a 48h rolling window and writes runs/*.json, each stamped with a per-run time_stats summary. No cross-run aggregation.
derive.py	The deriver. Reads runs/*.json and emits one fresh models/*.json snapshot per invocation.
demo.py	Trend viewer. Walks every run's time_stats and renders a single-file demo.html (Chart.js via CDN) with line charts of CI time over time (total / per-stage / per-runner). The generated demo.html is gitignored.
.github/workflows/scrape.yml	Auto-runs scrape.py + derive.py every 6h via GitHub Actions and commits any new files.

How to consume

Each runs/<...>.json looks like:

{
  "run_id": 25792056628,
  "started_at": "2026-05-13T09:59:44Z",
  "event": "schedule",
  "head_sha": "...",
  "jobs": [
    {
      "job_id": 75759852398,
      "name": "stage-c-test-4-gpu-h100 (2)",
      "suite": "stage-c-test-4-gpu-h100",
      "backend": "cuda",
      "started_at": "2026-05-13T10:05:10Z",
      "completed_at": "2026-05-13T10:26:28Z",
      "labels": ["4-gpu-h100"],
      "timings": [
        {"file": "test/registered/foo.py", "elapsed": 120, "passed": true},
        ...
      ]
    }
  ],
  "time_stats": {
    "total_wall_seconds": 42772,
    "per_stage":  {"base-b-test-1-gpu-large": 15504, ...},
    "per_runner": {"1-gpu-h100": 21800, ...}
  }
}

time_stats is this run's one data point in the time series. wall-clock per job is completed_at - started_at; parallel jobs are summed (runner-time consumed, not end-to-end latency). per_stage (by suite) and per_runner (by runner type) are two partitions of the same total_wall_seconds.

Clone the repo and walk runs/ directly, or curl individual files via raw.githubusercontent.com.

Time trend

python demo.py --open   # writes demo.html (gitignored) and opens it

Line charts of total / per-stage / per-runner CI time across all archived runs.

Filter / parser

• workflow: PR Test
• branch: main
• event: schedule or workflow_dispatch (PR-triggered runs excluded)
• run-level: status=completed (any conclusion)
• job-level: conclusion=success
• log parser: primary = TIMINGS BEGIN/END JSONL block (sglang#25232); fallback = legacy filename=..., elapsed=N, regex with keep-last retry dedup

Updating manually

gh workflow run scrape.yml -R sgl-project/sglang-ci-stats

or trigger from the GitHub Actions UI.