Test setup

fluxgo is evaluated on a chain of four Go microservices where every outbound call goes through a circuit breaker and a fault-injection proxy.

k6 load generator
      │
      ▼
 gateway :8080 ──Toxiproxy :8481──▶ auth :8081 ──Toxiproxy :8482──▶ user :8082 ──Toxiproxy :8483──▶ notification :8083
   [breaker]                          [breaker]                       [breaker]                       (terminal service)

• Three breakers, one per hop (gateway → auth, auth → user, user → notification). Each is either fluxgo or gobreaker, chosen at startup with CB_MODE.
• Toxiproxy sits on every link so latency, timeouts, bandwidth limits or a full outage can be injected between any two services.
• k6 drives the load against the gateway.
• Every service logs one JSON line per request and exposes its breaker state at GET /debug/cb (see Snapshot).

Hardware profiles

Each scenario runs under two resource limits per container, to see how the breakers behave when the services are starved.

Profile	CPU	Memory
Spec-A (standard)	2 vCPU	2048 MB
Spec-B (constrained)	0.5 vCPU	256 MB

Same parameters for both breakers

For a fair comparison, gobreaker’s static trip ratio equals fluxgo’s baseline threshold, and both share the same recovery timeout and half-open quota:

Setting	gobreaker	fluxgo
Trip threshold	failure ratio ≥ 0.6 or 5 consecutive failures	θₜ around ThetaBase = 0.6, adaptive
Open → Half-Open	Timeout = 60s	RecoveryTimeout = 60s
Half-Open probes	MaxRequests = 1	HalfOpenMaxRequests = 1

Scenarios

Scenario	Load	Fault	What it measures
Normal load	Steady, 5 min	None	False trips and overhead on a healthy system
Traffic spike	10 → 250 virtual users in 10 s	Bandwidth limit	False trips under a legitimate spike
Gradual degradation	Steady, 5 min	Latency +50 ms … +500 ms on user → notification	How early the breaker reacts to slow decay
Sudden failure	Steady, 3 min	Timeout at second 60	Detection delay on a hard outage

Each scenario is repeated for both breakers, alternating which one runs first to avoid thermal or caching bias. See Metrics for how runs are scored.