Question

A SaaS company runs a high-traffic order-processing service on ECS Fargate. The operations team wants per-customer and per-API-endpoint latency and error metrics so they can build dashboards and alarms broken down by both dimensions. There are roughly 8,000 active customers and 40 API endpoints. The team initially implemented this by calling the CloudWatch PutMetricData API from the application on every request with custom dimensions, but the monthly CloudWatch bill has grown dramatically and the API is being throttled during peak load. Which approach improves cost and reliability while preserving the ability to query metrics by customer and endpoint?

Accepted Answer

Emit structured logs using the CloudWatch Embedded Metric Format (EMF) to stdout, letting CloudWatch Logs asynchronously extract metrics, and use CloudWatch Logs Insights or Contributor Insights for high-cardinality per-customer analysis.

Answer

Continue using PutMetricData but batch up to 1,000 metrics per API call and add exponential backoff to eliminate throttling and reduce cost.

Answer

Create a separate custom CloudWatch metric namespace per customer and publish metrics with PutMetricData to isolate cardinality and reduce charges.

Answer

Send all request telemetry to Amazon Managed Service for Prometheus with per-customer labels and visualize it in Amazon Managed Grafana.