Analytics — data flow & API testing guide

Two hosts — don't mix them up

The platform runs on two separate prod hostnames. The Analytics page itself is a web UI page; the data behind it comes from the gateway API.

This trips people up: serverless.aies.scicom.dev = the web UI, serverlessgpu.aies.scicom.dev = the gateway API (note the extra gpu). Point curl / any OpenAI-style client at the gateway.

How the page gets its data

The page (page.tsx, admin-gated) renders a client component (analytics-view.tsx). On load it fires 4 parallel fetches to local Next.js API routes (/api/analytics/*). Those routes are server-side proxies: they inject auth and fan out to the real backends, so the browser never sees the gateway token or the SlurmUI secret.

Two data planes are kept deliberately separate: GPU Platform history (carries $ cost = cost_per_hr × duration) and SlurmUI reports (no cost — contributes job counts + GPU-hours).

The proxy routes (what the page calls)

The gateway endpoints (the real data source)

These are what you test directly with curl. All are admin-scoped admin only — a non-admin key gets 403 and an empty result. Base URL = https://serverlessgpu.aies.scicom.dev.

Example curls

1 · Set up

# gateway host + an ADMIN sgpu API key (create one in the web UI → Settings → API keys)
export GW=https://serverlessgpu.aies.scicom.dev
export TOKEN=sgpu_your_admin_api_key_here

# the window you want to query — ISO 8601, UTC
export SINCE=2026-06-01T00:00:00.000Z
export UNTIL=2026-06-15T23:59:59.000Z

# sanity check: list the available history kinds
curl -s "$GW/v1/history" -H "Authorization: Bearer $TOKEN" | jq

2 · Job history per kind (benchmark / training / compute / inference / endpoints)

# newest-first, max 1000 per page; bump offset to page through
curl -s "$GW/v1/history/benchmarks?since=$SINCE&until=$UNTIL&order=desc&limit=1000&offset=0" \
  -H "Authorization: Bearer $TOKEN" | jq '{count: (.jobs|length), has_more}'

# same shape for the other kinds
for kind in training compute inference endpoints; do
  echo "== $kind =="
  curl -s "$GW/v1/history/$kind?since=$SINCE&until=$UNTIL&order=desc&limit=1000" \
    -H "Authorization: Bearer $TOKEN" | jq '{count: (.jobs|length), has_more}'
done

3 · Inference summary (the chart numbers)

# exact per-day counts/tokens — tz controls day bucketing
curl -s "$GW/v1/history/summary?since=$SINCE&until=$UNTIL&tz=Asia/Kuala_Lumpur" \
  -H "Authorization: Bearer $TOKEN" | jq '.rows[:3]'

4 · Worker lifecycle events (GPU Timeline tab)

curl -s "$GW/admin/worker-events?since=$SINCE&until=$UNTIL&limit=20000" \
  -H "Authorization: Bearer $TOKEN" | jq '{count, sample: .events[:2]}'

5 · Source alias map

curl -s "$GW/v1/global-env" -H "Authorization: Bearer $TOKEN" \
  | jq '.[] | select(.key=="ANALYTICS_SOURCE_ALIASES")'

SlurmUI data is not directly testable by you. The /api/analytics/slurm proxy authenticates to SlurmUI with a server-only aura_… admin token that the browser never sees. You can only reach Slurm numbers through the web page (logged in as admin), not via the gateway.

Notes & gotchas

• Admin role required everywhere — both the web page and the API key. A non-admin gets a redirect (UI) or 403 + empty result (API).
• Pagination is capped. The web proxy pulls up to 5 pages × 1000 = 5000 records per kind and reports truncated for high-volume kinds (inference). For exact totals the charts use /v1/history/summary, not the raw paged list.
• Cost is computed client-side as detail.cost_per_hr × duration; inference/proxy carry no cost (they contribute activity counts + token totals).
• Time window: since/until are ISO 8601. The page sends until = end + 1s to make the range inclusive.
• Want to test the proxy routes instead of the gateway directly? Hit https://serverless.aies.scicom.dev/api/analytics/... from a browser session logged in as admin (they read your session cookie, not a Bearer token).

Internal testing reference · GPU Platform analytics · generated from main.