Skip to content
BytePatterns

Observability Basics

System Design: lesson 15 of 15

Metrics say something broke; traces say where.

Lesson 15 of 15 · 5 min

Observability Basics

Step 1 of 13

Metrics are cheap numbers over time: request rate, error rate, latency percentiles.

The Idea

Metrics are cheap numbers over time — request rate, error rate, latency percentiles — ideal for alerting and useless for detail. Logs record individual events with context, and get expensive at volume. A trace follows one request across every service it touched, which is the only view that finds the slow hop.

Real-World Example

A bus control room runs all three. The wall board shows headway and punctuality; each vehicle's fault log explains one breakdown; and when a complaint lands, somebody follows that single journey through every depot handover to see where the twelve minutes went.

The Tradeoff

All three cost money and attention, and a dashboard nobody reads is worse than none. Watch percentiles rather than averages, keep metric labels low-cardinality — a label per user ID will bankrupt you — and sample traces instead of recording every single request.

Your turn

Put the steps in the right order.

  1. Open the trace of a slow request to see which hop cost the time
  2. An alert fires when the error-rate metric crosses its threshold
  3. Read that one service's logs around the failing requests
  4. Check the dashboard to see which service the errors come from

Mini quiz

1 / 3

Alerting is best built on: