Percentage rollouts¶
Percentage rollouts express batches as fractions of the target group instead of fixed counts. They scale gracefully from 50-agent groups to 50 000-agent groups.
When to use¶
- The target group will grow or shrink over the life of the rollout (autoscaling, dynamic groups).
- You want consistent risk exposure across rollouts of different group sizes.
- You want to leave the operator-cognitive-load level low — "20 % is 20 %" is easier than "27 of 134 means 20 %".
Configuring¶
In the rollout wizard:
- Strategy:
Percentage. - Steps: a list of percentages (
5, 25, 50, 100is the default). The last step must be100. - Auto-advance: off, or on with a dwell timer.
Behaviour¶
- Step 1 (e.g. 5 %) computes the agent count from the current group size at the moment the step starts. If new agents join the group mid-step, they are not added to the in-flight batch.
- Each step proceeds like a batch — push, wait for
Applied, gate if anything goes wrong. - Steps move forward to the next on advance or auto-advance.
Re-evaluation between steps¶
Each step re-reads the group's membership at start-of-step. If the group grew between steps 1 and 2, step 2 covers the cumulative N % of the new total (minus the agents already covered).
This re-evaluation matters for dynamic groups: a label change that adds agents during a rollout is naturally absorbed.
Random vs deterministic ordering¶
By default, agents are randomised within a step — different rollouts on the same group hit different agents first. For repeatability you can pick by host name or by last seen.
A host name ordering plus the same step list across two configurations means "the same agents get the new behaviour first both times" — useful for aligning rollouts on the same canary cohort.
Compared to batch¶
- Batch is N agents per batch — easy to reason about for small fleets.
- Percentage is N % of the group — easy to reason about for large fleets.
Both share the underlying state machine and gates.
Tips¶
- The default
5, 25, 50, 100is a sensible starting point. For high-risk changes, prefer1, 5, 25, 100and bias dwell time on the early steps. - Percentage rollouts and dynamic groups compose well — let the selector pick the cohort, let the percentages pace the rollout.