Observability — AWS
Service mapping
- CloudWatch Metrics — managed metrics + dashboards + CloudWatch Alarms for alerting. The default; integrates with most AWS services out of the box.
- CloudWatch Logs — managed log ingestion/query (Logs Insights). The pipeline
concern is
distributed-logging; here it's the metrics/alarm + log-metric-filter side. - X-Ray — managed distributed tracing; OTLP via the ADOT (AWS Distro for OpenTelemetry) Collector.
- Amazon Managed Service for Prometheus (AMP) + Managed Grafana — run the generic Prometheus/Grafana stack without self-hosting; pick this to stay PromQL-portable instead of CloudWatch-native.
- CloudWatch Synthetics / Route 53 health checks — synthetic + external
liveness probing (health-check gating is
load-balancing).
When to pick which
- CloudWatch for native, low-setup metrics+alarms across AWS services.
- AMP + Managed Grafana when you want the generic recipe managed (portability, PromQL, existing dashboards).
- X-Ray for tracing if you're CloudWatch-native; ADOT keeps you OTel-portable.
Limits / things that bite (verify against current docs)
- CloudWatch custom-metric and high-cardinality dimension costs add up fast — each unique dimension combination is a billable custom metric.
- Standard CloudWatch metric granularity is 1-minute (high-resolution down to 1s costs more) — alerting reaction time is bounded by this.
- PutMetricData / API throttling under bursty custom-metric publishing.
- X-Ray sampling is configured separately; default rules may under-sample errors.
Pitfalls
- Emitting high-cardinality CloudWatch dimensions (user/request IDs) — cost explosion; keep those in logs/traces.
- Assuming 1-minute metrics are fast enough for tight SLO burn alerts.
- CloudWatch-native lock-in: dashboards/alarms don't port; AMP+Grafana avoids this.
- Forgetting cross-account/cross-region aggregation cost when centralizing telemetry.