feat(cost): expose priced-hours coverage and estimated queue rates

[khluu]

What

On /cost (2026-08-20 → 2026-09-03), Total Cost showed $31,176 "Known queues only" while only ~20% of compute hours were priced — h200_35gb (23.3k h) and amd_mi300_1 (20.5k h) had no rate, making the headline total misleading.

• API (src/app/api/cost/route.ts): adds pricedHoursShare / estimatedHoursShare (priced hours / total hours) to the response, plus an estimated flag per row in byQueue.
• Pricing (src/lib/queue-costs.ts): adds clearly-labeled estimated: true rates with source comments for the large partner/self-hosted queues — h200_35gb (scaled from the known $0.30/hr h200_18gb slice), full H200, AMD mi250/mi300/mi355 (market rental ~$1.80/$2.50/$3.00 per GPU/hr × queue GPU count), mithril-h100-pool, b200-k8s, l4-k8s, a100_queue, gh200_queue, medium_cpu_queue_premerge (assumed r6in.4xlarge), arm64_cpu_queue_premerge (assumed same as postmerge). Queues with no defensible rate (intel-cpu, amd-cpu, arm-cpu, intel-gpu, intel-hpu, ascend, dgx-spark) stay unpriced.
• Cost page (src/app/cost/page.tsx):
  • Total Cost card detail now reads e.g. Covers 21% of compute hours (+75% est.) (StatCard API unchanged).
  • Estimated rows in the Cost by Queue table get an est. suffix on $/hr and Est. Cost (no new colors).
  • The Unpriced Queues card is now expandable (<details>) and lists the remaining unpriced queue names.
• New computePricingCoverage() helper in queue-costs.ts with unit tests (src/lib/queue-costs.test.ts).

Coverage on the example range goes from 20.5% priced → 20.5% priced + 75.3% estimated (95.8% covered), 4.2% still unpriced (7 small queues, now listed in the card).

Verification

• npm test (tsx --test): 65/65 pass, including new queue-costs.test.ts (5 tests: estimated rates have sources, coverage split math, empty input, previously-unpriced big queues now priced).
• npm run lint: clean. tsc --noEmit: clean.
• Coverage numbers above verified by running computePricingCoverage against the live /api/cost byQueue payload.

Daily chart investigation (unverified)

The "Daily cost by queue" chart reportedly rendered empty in two headless captures despite 105 dailyCostByQueue rows with total_cost > 0. I could not reproduce in a real browser: npm run dev needs DATABRICKS_*/DATABASE_URL credentials that I don't have access to, so no confirmed bug fix here. Static review of the chart path found no defect: series keys (cost__<queue>/hours__<queue>) contain only -/_ (no Recharts path-separator characters), data flows correctly through the stacking useMemo, and ResponsiveContainer has an explicit height. Most likely the captures were taken before SWR resolved (the page early-returns "Loading cost data...") or before ResponsiveContainer/bar animation settled in the headless environment. Worth a follow-up capture with network-idle wait if it still reproduces.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
vllm-ci-dashboard	Ready	Preview	Sep 3, 2026 11:33am UTC