Gradual Ramp-Up Schedules
What you'll learn: How to systematically increase traffic to a new model using time-based or metric-based schedules to minimize risk during deployment.
The Core Concept
A gradual ramp-up schedule defines a controlled plan for incrementally increasing the percentage of traffic routed to your new model during deployment. Rather than switching all users at once (even in a canary deployment), you follow a predetermined schedule that gradually expands exposure.
Think of it like opening a new restaurant: you don't invite the entire city on day one. You might start with friends and family (1% traffic), then soft-open to neighbors (5%), gradually building to a grand opening (100%) only after validating each stage works smoothly.
Time-Based Schedules
With time-based schedules, you increase traffic at fixed intervals:
- Hour 0: 1% traffic to new model
- Hour 2: 5% traffic
- Hour 6: 10% traffic
- Day 1: 25% traffic
- Day 3: 50% traffic
- Day 7: 100% traffic
This approach provides predictable rollout timing but doesn't automatically respond to issues.
Metric-Based Schedules
Metric-based schedules condition progression on meeting quality thresholds. You only advance to the next stage if key metrics (error rate, latency, business KPIs) remain healthy:
- If error rate < 0.5% for 2 hours → increase to 10%
- If p95 latency remains acceptable → proceed to 25%
- If guardrail metrics pass → continue escalation
This is adaptive and safer but requires robust monitoring infrastructure.
Combining Both Approaches
Most production systems use hybrid schedules: time-based progression with metric-based gates. You advance on schedule only if health checks pass, providing both structure and safety.
Key Takeaway: Gradual ramp-up schedules define how traffic percentage increases during deployment, either on fixed timelines (time-based), conditional on performance metrics (metric-based), or both, minimizing risk by validating each stage before expanding exposure.