Guides
Scheduled changes and ramps
Put a change live at a time, or ramp a test's allocation up in steps, without being at your desk.
A scheduled change is a change request with a time. Put a launch at Monday 09:00, raise a test's allocation overnight, or ramp one up in steps, without being at your desk.
Scheduling a change
Forms that save what an environment serves — a test's or flag's settings and options, an audience's rules — have Schedule… next to save. Pick a date and time, to the minute, at most a year ahead.
- In an environment that doesn't need approval, anyone who may make the change may schedule it.
- In a reviewed one, an admin schedules directly; everyone else asks, an admin approves it and may change the time, and the approving admin becomes its scheduler.
The time is entered and shown in the time zone named in the hint: the zone Cloudflare places your request in, not your browser's (UTC when you run locally). Hours that a clock change skips or repeats are handled the same everywhere: a skipped time moves forward by the gap, a repeated one takes the first occurrence.
It goes live within about two minutes of its time, not to the second: Improve's minute cron runs it, and sites pick it up from the datafile after that.
Ramps
A ramp is one scheduled change in steps: a test's allocation, or a flag's share of one option.
- Up to 12 steps, each later than the one before, in the future and within a year.
- Each step runs on its own. The request stays Scheduled until the last step, and shows the step it is on and the next.
- A manual change to the subject in between stops the ramp: the next step finds it changed and fails, and the subject says why.
- Pausing a ramp cancels its remaining steps; what already ran stays. Rolling back is another change: allocation or share 0, or stop the test.
A ramp of a grouped test steps its group share, so its buckets grow as its allocation grows.
Running, and failing
A scheduled change runs as whoever scheduled it, with their role at that moment: someone who left the organization, or lost the role, doesn't change production. An organization over its plan's limit can't change anything, so the change fails instead of applying.
If the subject changed since the change was made, it does not run: it is marked failed with "changed since it was scheduled". That is deliberate — applying a change on top of someone else's would be worse than not applying it. It sends test.schedule_failed or flag.schedule_failed if a webhook is subscribed, and the asker and scheduler see the notice on the subject until they dismiss it.
An unexpected error rolls the run back and is retried when the cron next runs; after five tries the change is marked failed with the error.
Cancelling
The scheduler or an admin cancels a waiting change, or pauses a ramp, from the subject. Nothing changes; what a ramp already ran stays. Ask again to schedule something else — a subject has one scheduled change or ramp at a time.
What you see
- Lists show Scheduled · time on tests, flags and audiences, with a ramp's next step.
- The subject shows its schedule, the time, who scheduled it, a ramp's steps with the one reached, and Changed since when it no longer applies.
- The Scheduled page lists everything waiting in the environment.
Agents and MCP
scheduleChange, listScheduledChanges and cancelSchedule in the agent API, and the same as MCP tools. askForChange and approveChange take an applyAt. A time that has passed needs a new one on approval: an already-past ramp step would otherwise run in a burst.