การแจ้งเตือนและ SLO สำหรับความน่าเชื่อถือของเอพีไอ

Logs, metrics, and traces tell you what happened. Alerting turns that data into a page when something needs attention. Done well it catches problems early; done poorly it drowns you in noise.

An SLI is a measurable signal of user-facing health, such as:

• Request success rate
• P99 latency
• Availability

Good SLIs reflect what users experience, not internal trivia.

An SLO is a target for an SLI over a window, for example:

99.9% of requests succeed over 30 days.

It sets the line between acceptable and not.

If your SLO is 99.9%, you are allowed 0.1% failures — that is your error budget. Spend it on risky deploys; when it runs low, slow down and stabilize.

Alert on symptoms users feel (errors, slowness), not every internal cause. A high CPU alert may be harmless; a spike in 500s is not. Symptom alerts reduce false pages.

The simplest alert fires when a metric crosses a line for a duration.

Better than raw thresholds: alert on how fast you are burning the error budget. A fast burn pages immediately; a slow burn raises a ticket. This balances urgency and noise.

Too many alerts and on-call ignores them all. Keep alerts:

• Actionable — every page needs a response
• Deduplicated — group related fires
• Routed — page only for urgent, ticket the rest

Attach a runbook link to every alert: what it means, how to diagnose, and how to mitigate. The responder should never start from zero at 3 a.m.

Alerts say something is wrong; dashboards show why. Pair each SLO with a dashboard that breaks the SLI down by endpoint, region, and version so triage is fast.

After an incident, run a blameless postmortem: focus on the systemic causes, not the person who pushed the change. The output is a list of concrete fixes — better alerts, guardrails, runbook updates — that prevent recurrence.

You learned to alert on what matters:

• SLIs measure user-facing health
• SLOs set targets and define an error budget
• Alert on symptoms and burn rate, not every cause
• Keep alerts actionable with runbooks and dashboards