# Metrics — Derived Rules

> Derives from `domains/observability/first-principles.md` P1 (Structured by Default), P4 (Cardinality Discipline), P8 (SLI/SLO Awareness).

## The Four Golden Signals

| Signal | What |
|--------|------|
| Latency | Time to serve a request (p50, p95, p99) |
| Traffic | Request rate (req/s) |
| Errors | Error rate (errors/s, or % of traffic) |
| Saturation | How full is the system (CPU, memory, queue depth) |

- All four are needed. Missing one is a blind spot.
- Latency is percentiles, not average. Average hides the long tail.

## Cardinality (P4)

- Labels have bounded cardinality. `user_id` as a label = unbounded cardinality = unbounded bill.
- High-cardinality dimensions belong in traces, not metrics.
- A metric with `user_id` as a label is a 1M-series metric. That is a budget bomb.

## Counter vs Gauge vs Histogram

| Type | What | Example |
|------|------|---------|
| Counter | Monotonically increasing | `http_requests_total` |
| Gauge | A value at a point in time | `active_connections` |
| Histogram | Distribution of values | `http_request_duration_seconds` |

- A counter never decreases. Use `rate()` over time to get the rate.
- A gauge can go up and down. Use it for saturation.
- A histogram gives percentiles. Use it for latency.

## SLI/SLO (P8)

- SLI (Service Level Indicator): a metric of good/total (e.g., 99.9% of requests < 500ms).
- SLO (Service Level Objective): the target for the SLI (e.g., 99.9% over 30 days).
- Error budget: 1 - SLO. If SLO is 99.9%, error budget is 0.1%. Spend it on feature risk, not bugs.
- When the error budget is exhausted, freeze features. Fix reliability.

## What Violates Metrics Discipline

| Violation | Principle |
|-----------|-----------|
| `user_id` as a label | P4 Cardinality |
| Average latency only | P8 (hides the tail) |
| No error rate metric | P8 (no SLI) |
| 1000 metrics, no SLO | P8 (no objective) |
| A counter that decreases | (type error) |