Skip to content

Metrics Analysis in Locust

Built-in Metrics

Locust tracks and reports the following out of the box:

Metric Description
RPS Requests per second (current)
Failures/s Failed requests per second
p50 / p90 / p95 / p99 Response time percentiles
Avg / Min / Max Response time summary
Total requests Cumulative count
Total failures Cumulative failures

Available in: web UI (http://localhost:8089), terminal output, CSV export.

Key Focus Areas

Latency — Watch p95 and p99

Average is a vanity metric. Focus on tail percentiles.

p50 = 45ms   → typical user: fine
p95 = 280ms  → SLA threshold: check against contract
p99 = 1800ms → 1 in 100 users: investigate

A p99 that grows as load increases signals a resource saturation problem.

Errors — Spikes Under Load

A healthy system has 0% errors at expected load. Error rate is the first indicator that something is breaking.

RPS:    100  200  300  400
Errors: 0%   0%   0.2% 8%   ← 300 RPS is the threshold

Throughput — Plateau Detection

When RPS stops growing despite adding more users, throughput has plateaued. This is the system's capacity ceiling.

Users: 10  50  100  200  400
RPS:   10  50  100  110  112  ← plateau at ~100 RPS

Bottleneck Detection

Locust metrics alone show symptoms. Correlate with system metrics (CPU, memory, DB) to find the root cause.

Symptom in Locust Likely Cause
Latency rises linearly with load CPU saturation
Latency spikes at specific RPS Thread pool / connection pool exhaustion
Error rate jumps at threshold Timeout / queue overflow
RPS plateaus despite more users Backend bottleneck (DB, app)
p99 much higher than p95 Occasional GC pause, lock, or slow query

Reading the Locust Report

After a headless run, export CSV:

locust -f locustfile.py --headless -u 200 -r 20 --run-time 120s \
  --csv=results/run1

Files generated: - run1_stats.csv — per-endpoint summary - run1_stats_history.csv — time-series data - run1_failures.csv — failure details

Import stats_history.csv into Grafana or plot with Python/pandas to visualize latency trends over time — far more useful than summary numbers alone.