View a markdown version of this page

Monitor workload resources - AWS Well-Architected Framework

Monitor workload resources

A workload gives up signals about its health through the logs and metrics its components emit. When you collect those signals and act on them, you detect degradation early and recover before customers feel it. The failure mode that monitoring exists to prevent is the silent failure, where a function stops working and nothing catches it until a customer reports the problem. Alerting on that condition, through a path that stays up even when the workload does not, is one of the main reasons you monitor at all.

This capability covers how an organization monitors the resources a workload depends on across four phases: generating telemetry from every component, aggregating it into metrics you can reason about, processing it in real time to notify responders and invoke automated recovery, and storing it for analysis over time. It also covers reviewing the monitoring strategy as the workload changes, and tracing requests end to end so you can find where errors and latency originate in a distributed system.

Capability intent

  • Every tier of the workload emits telemetry, including business-outcome metrics alongside technical metrics, and external endpoints are monitored independently of internal components.

  • Logs and metrics are aggregated to a central location and enriched with dimensions, so correlated failures surface instead of averaging away.

  • Critical signals have thresholds that invoke alarms, and each alarm routes to automation or to a person based on whether the response needs human judgment.

  • Automated responses correct well-understood failures in seconds and apply the same procedure every time, while operators follow documented procedures for the rest.

  • Historical logs are retained long enough to reveal trends and regressions, under a lifecycle policy that controls storage cost.

  • Monitoring scope is reviewed on a cadence and after significant changes, and requests are traced end to end across every component.

Common issues to watch for

  • Teams monitor only external interfaces and stay blind to internal component health, so a failure inside the workload surfaces as a customer complaint.

  • Teams track technical metrics only and ignore business outcomes, missing the drop in successful checkouts that a healthy CPU graph hides.

  • Teams set alarm thresholds too low, flood responders with noise, and train the team to dismiss the alert that mattered.

  • Teams route alarms that automation could handle to people, which generates notification overload and slows response.

  • Teams retain logs too briefly to see a slow regression, or retain everything forever and absorb rising storage cost with no lifecycle policy.

  • Teams instrument some components but not all, so latency from an untraced component hides inside the gap and looks like time spent elsewhere.

  • Teams set up a monitoring strategy at launch and never revisit it, so stale thresholds fire during normal operation and new features fail silently.