An experiment moves through a small set of states. This page covers every control on the detail page and the list, and exactly what each one does to the placement the experiment is attached to.
The lifecycle
- Draft: created, not started. The placement still serves its default flow.
- Running: live, enrolling users into variants.
- Paused: temporarily halted. The placement serves its default flow again, but assignments are remembered.
- Completed: ended, either with a declared winner or without one. This is a terminal state.
Controls
Most controls live on the experiment detail page header. The action buttons change with the status:
| Status | Buttons shown |
|---|
| Draft | Edit, Start Experiment |
| Running | Pause, Stop |
| Paused | Resume, Stop |
| Completed | Duplicate |
The experiments list also has a ... menu on each row with Pause (when running), Resume (when paused), and Delete.
What each transition does
Start (Draft to Running)
Starting begins bucketing. The experiment is marked running and bound to its placement as the active experiment. From this moment, qualifying users are assigned to variants and the placement serves variant flows instead of its default.
Starting an experiment changes what the placement serves to live users. Make sure both variant flows are the versions you intend to test before you start.
Pause and Resume
Pause stops enrollment without ending the test. The experiment is marked paused, and while it is paused the placement falls back to serving its default flow, the same as if no experiment were attached.
The assignments already made are not thrown away. They are remembered, so when you Resume, every user who was already bucketed returns to the same variant they had before. Pausing is safe and reversible; use it to freeze a test without losing your data or your buckets.
Stop (Running or Paused to Completed)
Stop ends the experiment with no winner. The experiment becomes Completed · No winner, it is detached from the placement, and the placement reverts to serving its default flow. Stopping is irreversible: you cannot resume a stopped experiment or collect more data on it. Its results are kept for reference.
To end an experiment by promoting one variant to the placement default instead, use Declare Winner, covered in Reading results.
Duplicate
Duplicate copies an experiment’s structure (variants, traffic split, metric, confidence) into a brand new draft named ... (Copy). Start dates, the winner, and live results are not carried over. Use it to rerun a similar test.
Delete
Delete removes an experiment entirely. Only draft experiments can be deleted. Once an experiment has started (running, paused, or completed) the backend keeps it so its history and results are preserved, and rejects the delete. Pause or stop a live test if you want it off the placement, but it stays in your list.
What the placement serves in each state
| Experiment state | What the placement serves |
|---|
| Draft | Default flow (experiment is not attached yet) |
| Running | A variant flow per assigned user |
| Paused | Default flow (assignments are remembered for resume) |
| Completed · No winner (stopped) | Default flow |
| Completed · Winner (declared) | The winning variant’s flow, which is now the new default |
SDK clients cache resolved flows, and the backend caches placement configuration. After you start, pause, stop, or declare a winner, allow a short propagation window before every device reflects the change. Sticky assignments mean already-enrolled users keep their variant regardless. Assignment is keyed off a stable per-install user id that the SDK generates and persists; see Variables and SDK context for how that id works and why a returning user stays in the same variant.
Common mistakes
- Treating Stop as Pause. Stop is final and clears the experiment off the placement. If you only want to take a breather, use Pause, which keeps assignments and lets you resume.
- Expecting Stop to apply a winner. Stopping reverts the placement to its original default flow. To promote a variant, use Declare Winner instead.
- Trying to delete a live or finished experiment. Only drafts are deletable. Stop a running test to take it off a placement; it remains in the list as a completed record.
Related pages