Adaptive threshold

A static breaker compares the failure rate with a constant. fluxgo compares it with θₜ, a threshold recomputed from live latency.

The formula

R_ewma  = St / Sbase           EWMA ratio vs the healthy baseline
R_p     = Pt / Pbase           P95/P99 ratio vs the healthy baseline
Dt      = W1·R_ewma + W2·R_p   weighted degradation ratio
theta_t = clamp(ThetaBase / Dt, ThetaMin, ThetaMax)

Symbol	Meaning
St	Current EWMA latency, smoothed with Alpha
Pt	Current P95 (or P99) over the rolling window
Sbase, Pbase	Healthy-state baselines captured during warmup
W1, W2	Weights of the two signals; they must sum to 1
Dt	Degradation ratio: 1 means “as healthy as the baseline”
ThetaBase	Threshold used when latency equals the baseline

What it does in practice

• Latency doubles (Dt = 2): θₜ halves. With the default ThetaBase = 0.6 the circuit opens at a 30% failure rate instead of 60%.
• Latency improves (Dt < 1): θₜ rises, so the circuit tolerates more failures before opening.
• Clamping keeps θₜ inside [ThetaMin, ThetaMax] (default [0.05, 0.9]), so a single outlier cannot make the breaker either trigger-happy or deaf.

The circuit opens when the failure rate in the outcome window reaches θₜ, and only once the window holds at least MinRequests outcomes.

Why two signals

• EWMA reacts smoothly to a shift of the average latency.
• P95/P99 reacts to the tail, which is where a degrading dependency shows up first.

W1 = 0.4 and W2 = 0.6 weight the tail slightly more. Setting W1 = 1 uses only the average; W2 = 1 uses only the percentile.

Watching it live

Snapshot() exposes St, Pt, Dt, ThetaT and the baselines, so you can chart θₜ against the failure rate in your own monitoring.