Interpreting Results

Healthy System

A system behaving correctly under load shows:

Signal	What It Looks Like
Stable RPS	Flat line at target throughput
Flat latency	p95/p99 barely changes as users increase
Low error rate	0% or < 0.1% throughout the test
Stable memory	Memory usage plateaus, no upward drift
Consistent CPU	CPU proportional to RPS, not spiking

RPS:     ──────────────────── (flat at target)
p95:     ────────────────────
Errors:  ────────────────────
          0s       60s      120s

---

Degrading System

Early warning signs before a system fails:

Signal	Interpretation
Latency slowly increasing	Resource is being consumed — find which one
Occasional p99 spikes	GC pauses, lock contention, slow queries appearing
RPS variance increasing	Processing becomes inconsistent
Memory creeping up	Potential memory leak — confirm with soak test

p99:     ___
            \___
                \___________
                             \___
                                  \__________________ (slowly growing)
          0s       60s      120s     180s     240s

This pattern demands investigation before production load reaches these levels.

---

Overloaded System

System operating beyond capacity:

Signal	What Happens
Latency spikes sharply	Queue backs up, requests wait for workers
Error rate jumps	Timeouts, 503s, connection refused
Throughput plateau	Adding users doesn't add RPS — system is saturated
RPS drops despite high user count	System spending time recovering, not processing

RPS:     ___________
                    \_____________________  ← plateau then drop
p99:                       _______________
                           /               ← spike
Errors:                    ______________
                           /               ← spike
          0s       60s      120s     180s

The knee point is where this transition begins — the critical finding of any load test.

---

Reading the Numbers

Scenario: is 200ms p95 acceptable?

It depends on the SLO: - If SLO is p95 < 300ms → pass - If SLO is p95 < 100ms → fail

Always evaluate results against the SLO, not against intuition.

Scenario: RPS is 100 but users are 200

Effective RPS = users × (1 / (response_time + wait_time))

200 users × 1 / (0.4s + 0.6s wait) = 200 RPS possible
If actual RPS = 100 → latency is 1.4s, not 0.4s

High user count with low RPS signals the system is slow, not that Locust is misconfigured.

---

Result Classification Matrix

p95	Error Rate	Conclusion
Within SLO	< 0.1%	Pass — healthy
Within SLO	0.1–1%	Warning — investigate errors
Within SLO	> 1%	Fail — reliability issue
Exceeds SLO	< 0.1%	Fail — latency issue
Exceeds SLO	> 1%	Fail — system overloaded