Metrics Definition

When to activate

When building a metrics layer, creating KPI definitions for stakeholders, or reconciling dashboard metric discrepancies. Run before implementing metrics in dbt, a semantic layer, or dashboard tool. Use whenever a business metric needs a single source of truth.

When NOT to use

Not for ad-hoc metric exploration — this skill is for canonical, production metrics. Not without stakeholder alignment on what the metric should measure. Not for exploratory analysis without clear business context.

Metrics Definition Checklist

1. Identify the metric — What business outcome does it measure (revenue, churn, DAU, NPS)?
2. Define the formula — Explicit SQL or calculation steps; no ambiguity
3. Set the grain — Atomic level (daily, user-level, transaction-level)
4. Specify the window — MTD, YTD, rolling 30 days, point-in-time?
5. Document the source — Which fact table, dimension, or raw table?
6. List edge cases — Nulls, outliers, data quality issues, cancellations, refunds
7. Specify SLA — Latency requirement (real-time, hourly, daily, weekly)
8. Owner & stakeholders — Who owns the metric; who consumes it

Metric Types

Fact metric — Counts or sums directly from raw events (transactions, page views, sign-ups). Example: total_revenue, order_count.

Ratio metric — Quotient of two other metrics (conversion_rate = conversions / visitors). Risk: division by zero edge cases.

Derived metric — Calculated from other metrics or dimensions (net_revenue = gross_revenue - refunds - discounts).

Cohort metric — Grouped by cohort dimension, tracked over time (cohort_retention = active_users_this_month / users_in_cohort).

Derived time-series metric — Change over time (month_over_month_growth = (current_month - prior_month) / prior_month).

Metrics Registry Template

# Metrics Registry

**Team:** [Team name]  
**Last Updated:** [date]  
**Owner:** [Name, role]  

---

## Metric: [Metric Name]

**Business Definition:** [One sentence — what does this metric measure and why?]

**Formula:**
\`\`\`sql
SELECT 
  SUM(transaction_amount) AS metric_value
FROM fact_transactions
WHERE DATE(transaction_date) >= '2026-01-01'
  AND DATE(transaction_date) < '2026-02-01'
  AND transaction_status = 'completed'
\`\`\`

**Grain:** [Monthly / Daily / User-level / Transaction-level]  
**Calculation Window:** [Calendar month / Rolling 30 days / Fiscal quarter]  
**Data Source:** [fact_transactions] from [Data warehouse name]  
**Latency:** [Real-time / Hourly / Daily / Weekly]  
**Owner:** [Name]  
**Stakeholders:** [Teams/roles that consume this metric]

---

## Edge Cases & Handling

| Case | Value | Treatment |
|------|-------|-----------|
| Refunded transactions | Negative amount | Included in sum (net revenue) |
| Cancelled orders | NULL status | Excluded (WHERE status = 'completed') |
| Multi-currency | EUR, GBP, JPY | Converted to USD at transaction_date rate |
| Free trial users | $0 | Excluded (WHERE subscription_type != 'trial') |

---

## Calculation Example

**Period:** January 2026  
**Expected Output:** ~$2.4M

**Debug Query:**
\`\`\`sql
SELECT 
  transaction_date,
  COUNT(*) AS tx_count,
  SUM(transaction_amount) AS daily_revenue
FROM fact_transactions
WHERE DATE(transaction_date) >= '2026-01-01'
  AND DATE(transaction_date) < '2026-02-01'
  AND transaction_status = 'completed'
GROUP BY transaction_date
ORDER BY transaction_date;
\`\`\`

---

## Implementation

**In dbt:** Create a metric node in `schema.yml`:

\`\`\`yaml
metrics:
  - name: total_revenue
    label: Total Revenue
    model: fact_transactions
    description: "Sum of all completed transactions in USD"
    calculation_method: sum
    expression: transaction_amount
    filters:
      - field: transaction_status
        operator: '='
        value: 'completed'
    time_grains: [day, month, quarter, year]
\`\`\`

**In semantic layer:** Register metric in [Tool name] with refresh cadence and permissions.

---

## Reconciliation

**Dashboard vs. Report mismatch?** Check:
1. Calculation window (MTD vs. calendar month vs. rolling)
2. Data freshness (real-time vs. 1-day lag)
3. Filter differences (active users vs. all users)
4. Dimension slice (country filter, subscription type, etc.)

---

### [Metric 2 Name]

**Business Definition:** [definition]  
**Formula:** [SQL]  
**Grain:** [grain]  
... (repeat template)

---

Example

Metrics Registry

Team: Analytics
Last Updated: June 12, 2026
Owner: Alice Chen, Senior Analytics Engineer

Metric: Monthly Recurring Revenue (MRR)

Business Definition: Total predictable revenue from active subscriptions at month-end, excluding one-time purchases and trials.

Formula:

SELECT 
  DATE_TRUNC('month', DATE(snapshot_date))::DATE AS month_end,
  SUM(monthly_subscription_amount) AS mrr
FROM dim_subscriptions
WHERE snapshot_date = DATE_TRUNC('month', DATE(snapshot_date)) + INTERVAL '1 month' - INTERVAL '1 day'
  AND subscription_status = 'active'
  AND subscription_type NOT IN ('trial', 'free')
GROUP BY DATE_TRUNC('month', DATE(snapshot_date));

Grain: Monthly (aggregated across all active subscriptions at month-end)
Calculation Window: Calendar month (measured at last day of month)
Data Source: dim_subscriptions from Data warehouse
Latency: Daily (snapshot taken nightly)
Owner: Alice Chen
Stakeholders: Finance, Product, Executive Leadership

Edge Cases & Handling

Calculation Example

Period: May 2026
Expected Output: ~$1.8M

Debug Query:

SELECT 
  subscription_id,
  customer_id,
  monthly_subscription_amount,
  subscription_status,
  subscription_type
FROM dim_subscriptions
WHERE snapshot_date = '2026-05-31'
  AND subscription_status = 'active'
  AND subscription_type NOT IN ('trial', 'free')
LIMIT 100;

Implementation

In dbt:

metrics:
  - name: monthly_recurring_revenue
    label: MRR
    model: dim_subscriptions
    description: "Total monthly recurring revenue from active subscriptions"
    calculation_method: sum
    expression: monthly_subscription_amount
    filters:
      - field: subscription_status
        operator: '='
        value: 'active'
      - field: subscription_type
        operator: not_in
        value: ['trial', 'free']
    time_grains: [month, quarter, year]