Course contentsShow
← System Design
Lesson 1185 of 1,91930. Observability: MetricsFree lesson

Histogram Metrics

Understanding bucketed distributions for measuring request latencies, sizes, and other continuous values.

Histogram Metrics

What you'll learn: How histograms bucket continuous values like latencies and response sizes to reveal distribution patterns that counters and gauges can't capture.

The Distribution Problem

Imagine measuring API request latency. A gauge shows the current value, a counter tracks totals—but neither tells you how latencies are distributed. Are most requests fast with occasional spikes, or consistently slow? Histograms answer this.

What Histograms Do

A histogram organizes continuous measurements into predefined buckets (ranges). Instead of storing every individual value, it counts how many observations fall into each bucket.

For example, tracking HTTP request duration with buckets:

  • 0-10ms: 850 requests
  • 10-50ms: 120 requests
  • 50-100ms: 25 requests
  • 100-500ms: 4 requests
  • 500ms+: 1 request

This reveals that most requests are fast, but you have outliers.

Key Features

Bucketing strategy matters. You define boundaries upfront (e.g., [10, 50, 100, 500, 1000] milliseconds). Choose buckets that match your SLAs—if 100ms is your target, ensure a bucket near that threshold.

Cumulative counts. Many implementations store cumulative counts: "≤10ms", "≤50ms", etc. This enables percentile calculation: if 970 of 1000 requests are ≤100ms, your P97 latency is ~100ms.

Trade-off: precision vs efficiency. You lose exact values but gain efficient storage and queryability across millions of requests.

When to Use Histograms

Use histograms for continuous values with meaningful distributions: request latencies, response payload sizes, database query durations, or message processing times. They're essential for understanding tail latency (P95, P99) that averages hide.

Key Takeaway: Histograms bucket continuous measurements into ranges, revealing distribution patterns and enabling percentile calculations—critical for understanding performance characteristics beyond simple averages.