The Four Golden Signals, Revisited for 2026

Latency, traffic, errors, saturation. The original four still hold; the way you measure them has evolved. The 2026 update with concrete metric definitions.

Latency: percentiles, not averages

p50, p95, p99 are the working set. Averages hide tail behaviour and are useless for SLOs.

Track per-endpoint and per-method. A single latency number per service hides the bad endpoints behind the good ones.

Errors: rate, not count

Errors per request, not errors per minute. Rate is comparable across services; raw count is not.

Distinguish 4xx (client error) from 5xx (server error). They have different fixes and different urgencies.

Saturation: the leading indicator

CPU, memory, connection pool utilisation. The signal that fires before the user-visible failure.

Saturation alerts let you act preventively. Most teams under-instrument it; the cost is missed early warnings.