metrics-definition

The Core Problem

"47 dashboards and no answers." The failure mode is building metrics without agreeing on what they mean. "Revenue" means three different things to three different teams. "Active users" has no agreed-upon time window. Fix the definitions before building the dashboards.

Metric Definition Template

Every metric must specify:

• Name: Human-readable, follows naming convention (see below)
• Business definition: One sentence a non-technical person understands
• SQL logic: Exact calculation including numerator, denominator, and filters
• Grain: What level is this calculated at? (daily, weekly, per-user, per-account)
• Time window: What period does this cover? (trailing 7 days, calendar month, since signup)
• Dimensional attributes: What can you slice this by? (region, product line, customer segment)
• Owner: Who maintains this definition?
• Update cadence: How often does this refresh?
• Known edge cases: What situations produce unexpected results?

Never define a metric without specifying grain and time window. "Monthly active users" means nothing until you define what "active" means and whether "monthly" is calendar month or trailing 30 days.

Outcome Metric Trees

Connect metrics from business outcomes down to leading indicators:

Business outcome: Reduce hospital readmissions 10%
  Product outcome: Clinical decisions made 3x faster
    Feature outcome: Risk scores updated in real-time
      Leading indicator: Query latency under 1 second

Every metric at a lower level should causally influence the level above. If you can't draw the causal link, the metric doesn't belong in the tree.

Trust Metrics

Alongside performance metrics, track trust:

• Data accuracy rate: Percentage of values matching gold standard (target: 99.9%)
• Query response time: P95 latency for consumer queries (target: <3 seconds)
• Data freshness: Time between source update and availability (target: within SLA)
• Support response time: How fast you resolve data questions (target: <24 hours)
• Incident count: Zero algorithmic bias incidents, zero data breaches