Bosca / Experiments

Rollout Policies

Ramp it up. Stop on regression.

A rollout policy advances a winning variation toward full traffic on your terms — by hand, on a schedule, or automatically as the numbers hold. Guardrails watch the whole way and halt the ramp the moment a key metric slips.

How it advances

Four ways to widen exposure

Pick how much control you want to keep. Drive the ramp by hand, let the clock drive it, or let the results drive it — in discrete steps or one continuous climb.

Manual

You advance each step yourself, whenever you're ready to widen exposure.

Scheduled steps

Steps advance on a timer — a set weight after a set duration.

Adaptive steps

Each step waits until confidence clears your threshold before advancing.

Adaptive continuous

Traffic climbs by a fixed increment as long as the numbers hold up.

Steps & guardrails

Widen only when it's safe

A policy is a ladder of weight steps and a set of thresholds. The ramp only climbs to the next rung when the treatment clears the confidence you require — and a guardrail regression can hold it in place or halt it outright.

  • Ordered steps set the weight to serve and how long to hold it.
  • A minimum confidence gates each advance so you don't widen on noise.
  • Guardrail goals define the regression you won't tolerate — cross it and the ramp halts.
rollout · checkout-cta
5%
25%
50%
100%
guardrail: error rate ≤ +0.5% · min confidence 95%

Audit trail

Every move on the record

A rollout records what it did and why. Each advance, hold, halt, and completion is logged with the weight before and after and the confidence at the time — so an automated ramp is never a black box.

  • Four event kinds capture the ramp: advanced, held, halted, and completed.
  • Every entry keeps the before-and-after weight, so you can trace the whole climb.
  • A guardrail breach that halts the ramp shows up here with the metric that tripped it.
rollout history
advanced5% → 25%confidence 96%
advanced25% → 50%confidence 97%
held50% → 50%confidence 92% < 95%
advanced50% → 100%confidence 98%
completed100% → 100%fully rolled out

Keep exploring

More on Experiments

Experiments overview

Flags, A/B tests, rollouts, and results — one system.

Feature Flags

Typed flags with targeting rules and deterministic bucketing.

A/B Experiments

Measure one variation against another, with sticky assignment.

Results & Analysis

Real statistics from your analytics, with a ship / don't-ship call.