Data — Analytics & Intelligence

Measure what matters. Every metric should drive a decision. If it doesn't, stop tracking it.

Metric Hierarchy

Level	Metric	Cadence
North Star	1 metric that defines success (e.g., WAU, MRR)	Weekly
Health	3-5 metrics that predict north star (retention, activation, NPS)	Weekly
Feature	Per-feature usage, conversion, time-to-value	Per release
Debug	Granular events for troubleshooting	On-demand

Rules:

North star is singular. Two north stars = zero north stars.
Vanity metrics (page views, total signups) are not health metrics.
Every feature ships with tracking. No "we'll add analytics later."

SQL Patterns

Cohort Retention

SELECT
  DATE_TRUNC('week', u.created_at) AS cohort_week,
  DATE_TRUNC('week', e.created_at) AS activity_week,
  COUNT(DISTINCT e.user_id) AS active_users
FROM users u
JOIN events e ON e.user_id = u.id
GROUP BY 1, 2
ORDER BY 1, 2;

Funnel Analysis

WITH steps AS (
  SELECT user_id,
    MAX(CASE WHEN event = 'signup' THEN 1 END) AS step_1,
    MAX(CASE WHEN event = 'onboard_complete' THEN 1 END) AS step_2,
    MAX(CASE WHEN event = 'first_action' THEN 1 END) AS step_3,
    MAX(CASE WHEN event = 'paid' THEN 1 END) AS step_4
  FROM events
  WHERE created_at > NOW() - INTERVAL '30 days'
  GROUP BY user_id
)
SELECT
  COUNT(*) AS total,
  SUM(step_1) AS signup,
  SUM(step_2) AS onboarded,
  SUM(step_3) AS activated,
  SUM(step_4) AS converted
FROM steps;

Rolling Averages

SELECT
  date,
  value,
  AVG(value) OVER (ORDER BY date ROWS BETWEEN 6 PRECEDING AND CURRENT ROW) AS rolling_7d
FROM daily_metrics;

Event Tracking Design

Every event needs:

Field	Example
`event_name`	`button_clicked`, `page_viewed`, `feature_used`
`user_id`	Authenticated user
`anonymous_id`	Pre-auth (cookie/device)
`timestamp`	ISO 8601, UTC always
`properties`	`{ page: "/pricing", plan: "pro" }`

Naming convention: noun_verb — form_submitted, file_uploaded, subscription_cancelled

Never: click, event1, trackThis, misc

Dashboard Rules

One question per dashboard. "How is acquisition?" not "Everything."
Top-left = most important metric. Eye lands there first.
Comparison always. This week vs last week. This month vs last month.
No pie charts. Use bar charts. Humans can't compare angles.
Annotate changes. "Launched feature X" on the timeline.

Data Pipeline Defaults

Layer	Tool	Purpose
Collection	PostHog, Segment, or custom	Event ingestion
Storage	Postgres or BigQuery	Queryable warehouse
Transform	dbt or SQL views	Business logic layer
Visualization	Metabase, Grafana, or PostHog	Dashboards

Keep it simple. You don't need Kafka until you have 10M events/day.

data