Talk to sales
Workbook

Data Analyst

You own: answering business questions with data — including questions about model performance, cost, and drift.

What's new for you

WasNow in GeneFlow
ML metrics in slide decksAll in the warehouse via gf_* exports
Cost-per-prediction was opaquegf_inference_logs.cost_usd is queryable
No way to answer "which model serves this product?"Lineage + gf_endpoints joined

Day-1 setup

You don't write code against GeneFlow directly — you query the gf_* tables in your warehouse (replicated via CDC). Talk to your AE if they aren't there yet.

Workflow 1 — Per-product model health

SELECT
  e.name      AS endpoint,
  e.model_name,
  e.model_version,
  AVG(l.latency_ms)::INT AS avg_latency_ms,
  PERCENTILE_CONT(0.95) WITHIN GROUP (ORDER BY l.latency_ms) AS p95_latency_ms,
  100.0 * AVG(CASE WHEN l.status_code >= 400 THEN 1 ELSE 0 END) AS error_rate_pct,
  COUNT(*) AS reqs
FROM gf_endpoints e
LEFT JOIN gf_inference_logs l
       ON l.endpoint_id = e.id AND l.ts >= now() - INTERVAL '24 hours'
WHERE e.tenant_id = 'ACME' AND e.status = 'READY'
GROUP BY e.name, e.model_name, e.model_version
ORDER BY reqs DESC;

Workflow 2 — Cost vs. revenue per model

SELECT
  e.model_name,
  SUM(l.cost_usd)                AS daily_inference_cost,
  SUM(l.cost_usd) * 365          AS annualized_cost,
  COUNT(*)                        AS daily_inferences
FROM gf_endpoints e
JOIN gf_inference_logs l ON l.endpoint_id = e.id
WHERE l.ts >= now() - INTERVAL '1 day'
GROUP BY e.model_name
ORDER BY daily_inference_cost DESC;

Join to your revenue tables for cost-of-prediction-as-% of revenue.

Workflow 3 — Drift incidents over time

SELECT date_trunc('day', fired_at) AS day,
       severity,
       COUNT(*) AS firings
FROM gf_drift_alerts
WHERE tenant_id='ACME' AND fired_at >= now() - INTERVAL '30 days'
GROUP BY 1, 2 ORDER BY 1 DESC;

Workflow 4 — Compare model versions in production

-- Yesterday's inferences split by version
SELECT model_version,
       COUNT(*) AS reqs,
       AVG(latency_ms)::INT AS avg_ms,
       SUM(cost_usd) AS cost
FROM gf_inference_logs
WHERE endpoint_id = 'ep_abc...' AND ts >= now() - INTERVAL '1 day'
GROUP BY model_version;

Useful during canary rollouts.

Workflow 5 — Cohort: prompts in Production

SELECT name,
       MAX(version) FILTER (WHERE current_stage='Production') AS prod_version,
       MAX(version) AS latest_version
FROM gf_prompt_versions
WHERE tenant_id='ACME'
GROUP BY name;

Common gotchas

  • gf_inference_logs is sampled — production endpoints with high QPS sample at 1% by default. Don't trust raw counts for revenue calcs; trust rates.
  • Tenant filter on every query — multi-tenant. Mistakes mix tenants.
  • feature_snapshot is JSONB — feature_snapshot->>'amount' returns text; cast to numeric before aggregating.

Where to go next