Metrics Aggregation

You run 50 instances of your API service. Each emits its own metrics: request count, error rate, latency histogram. That's 50 separate streams of data. The CEO doesn't want to see 50 dashboards, they want one number: "how is the API performing?"

Metrics aggregation is the process of combining raw metric data into meaningful summaries, across instances, across time windows, across dimensions. "Instance 17 served 340 requests in the last minute" becomes "the API fleet handled 18,000 requests per minute with a P99 latency of 120ms." The raw data tells you what individual machines are doing. The aggregated data tells you what the system is doing.

Every dashboard you've ever seen is built on aggregation. The monitoring system collects granular data and then slices, sums, averages, and percentiles it into the views that humans actually use for decisions.

Metrics can be aggregated across three dimensions:

Across instances (spatial aggregation)

• Sum: total requests across all instances
• Average: average CPU across all instances
• Max: highest memory usage among all instances
• Percentile: P99 latency across all instances (requires histograms)

Across time (temporal aggregation)

• Rate: requests per second over a 5-minute window
• Average: average latency over the last hour
• Max: peak error rate today
• Increase: total new errors since midnight

Across labels (dimensional aggregation)

• By status code: error rate for 5xx vs 4xx vs 2xx
• By endpoint: latency per API path
• By region: request volume per geographic region
• By customer tier: performance for free vs paid users

In PromQL (Prometheus's query language):

• sum(rate(http_requests_total[5m])), total request rate across all instances
• avg by (instance)(node_cpu_utilization). CPU per instance
• histogram_quantile(0.99, sum by (le)(rate(http_request_duration_seconds_bucket[5m]))). P99 latency across all instances