PROGRESSIVE ROLLOUT · GATED EXPOSURE
Progressive rollout: expand candidate exposure
Step a defective candidate from 1% to 100% of traffic behind an error-rate gate, and measure how many users it hurts before it is stopped.
Event trace
Computed data
Final metrics
1,000 requests per minute, a baseline error rate of exactly 1%, and observed candidate errors equal to the whole number expected so far (rounded down): a deterministic stand-in for sampling noise. A step is extended until it has the minimum sample. Latency gates, segment-specific defects and statistical tests are not modeled.
A gate too loose to fire
The candidate adds 1.5 points of errors, a true rate of 2.5% against a 1.0% baseline. With a 2 min bake and a gate that fails only above baseline by 2 points, the steps observe 0%, 2%, 2.4%, 2.5% and 2.5%. None exceeds 3%, so the defect reaches 100%: PROMOTED WITH DEFECT, about 54.3 extra failed requests over 10 min.
Tighten the gate
Set Gate fails above baseline by to 1 point: the 25% step sees 12 errors in 500 requests, 2.4%, and rolls back after 6 min with 9.3 extra failures. At 0.5 points the 5% step already fails, 2 errors in 100, and only 1.8 extra failures reach users. Setting When the gate fails to Pause the rollout stops at the same step without reverting; Alert only and continue raises the alert at 25% and promotes anyway, ALERTED BUT PROMOTED, 54.3 extra failures.
Samples versus speed
Small steps see too few requests to measure 1.5 points: 20 requests at 1% expect 0.5 errors and observe none. Set Minimum candidate requests before judging to 500 with the 1-point gate: the 1% step bakes for 50 min, sees 12 errors in 500 requests and rolls back at 1% with 7.5 extra failures. Fewer users are hurt, but the rollout is slow. A longer Bake time per step alone does not help a loose gate: 10 min per step promotes the defect with 271.5 extra failures over 50 min.
Defect size
A Candidate extra error rate of 0 promotes cleanly. At 10 points the first step sees 2 errors in 20 requests, 10%, and rolls back after 2 min. The board shows each step's share and observed error rate; the table lists minutes, requests, errors and the gate result.
Replay and limits
Any change recomputes the run and rewinds it; the Complete-run preview states the outcome at once. Play and Pause walk through the events, Next event jumps one event, Restart rewinds, Speed runs from 0.25x to 2x and Inspection time seeks from 0 to 60 seconds, where the outcome and final metrics appear. Traffic is 1,000 requests per minute, the baseline is exactly 1%, and observed errors are the whole number expected so far, rounded down: a deterministic stand-in for sampling noise. Latency gates and statistical tests are not modeled.