RED metrics: rate, errors, duration

If you can only keep three application metrics, keep these. Tom Wilkie’s RED method — rate, errors, duration — is the application-facing half of SRE’s golden signals. Saturation stays on the host. RED stays on the request.

The problem

Teams drown in custom dashboards because every layer emits something. CPU, GC, queue depth, cache hit ratio — all useful in context, none of them the first answer to “is checkout OK?” The first answer is: are people hitting it, are they failing, and how long are they waiting?

What it is

• Rate (throughput) — sampled transactions per minute or second for a route or service.
• Errors — the fraction of those transactions with span status error.
• Duration — a latency distribution. Averages lie; p95 and p99 show the tail users actually feel.

Apdex is a companion score, not a fourth golden signal. It classifies requests as satisfied, tolerating, or frustrated against a threshold, then folds that into 0–1. It is a conversation tool (“we dropped below 0.9 after the deploy”), not a substitute for looking at p95.

What it gives a team

• A one-screen health check for the app, not the cluster.
• A way to see whether a deploy changed traffic shape, failure rate, or latency — three different incidents.
• A threshold language for performance alerts: error rate, p95, or throughput crossing a line you chose.

How Talaria does it

ClickHouse materialized views aggregate root spans into span_red_1m (count, errors, p50/p95/p99 by project, service, route, environment, minute). The dashboard calls SpansEndpoint.redSeries and renders stat tiles plus an area chart you can toggle across throughput, p95, and errors. Time range controls cover 7 / 14 / 30 days.

The Performance overview is prebuilt so the path stays short: read the three signals, then open a waterfall.

How to read it

• Throughput down + errors up — dependency or deploy failure. Open error traces.
• Throughput flat + p95 up — something on the critical path got slower. Open a slow transaction.
• Errors up + p95 flat — failures that return quickly (4xx/5xx without waiting). Check issue grouping and the failing peer on Dependencies.
• Apdex down with a quiet p95 — a larger share of requests crossed the “tolerating” line even if the extreme tail did not move much.

Because sampling is head-based, RED describes sampled traffic. Error transactions are always kept once tracing is on, so error rate is not under-sampled the way successful latency is. Read duration with that in mind.

Turn it on

RED series fill after sampled transactions arrive. Enable tracing on an official SDK. Without ClickHouse, the overview is empty — span storage is not optional for this view.