Prompts need versions, gated promotion and rollback. What fits?
Why D is the answer
MLflow prompt versions with aliases
Aliases point to versions, so rollback is simple.
In detail
- Versions keep a history.
- An alias points to the version in use, so promotion and rollback only move the alias.
- Serve the alias, not a fixed version, in your app.
Use case
An illustrative scenario, not a real incident.
A team tests prompt version 4, then moves the production alias to it. When problems appear, the alias goes back to version 3 in seconds.
See it in different ways
flowchart LR
A["Prompt version 4"] --> B["Automated tests"]
B --> C{"Approved?"}
C -->|Yes| D["Move the production alias"]
C -->|No| E["Keep the alias on version 3"]sequenceDiagram participant C as pyfunc chain participant M as MLflow participant U as Unity Catalog participant S as Model Serving C->>M: Log the model M->>U: Register a version U->>S: Deploy to an endpoint
stateDiagram-v2 state "Version 3 live" as s0 state "Version 4 tested" as s1 state "Version 4 live" as s2 [*] --> s0 s0 --> s1: new version s1 --> s2: move alias s2 --> s0: roll back
Key point: Move an alias to promote or roll back.
Why the other options are wrong
- A. A JSON file overwritten on each run
- Overwriting a file loses the history.
- B. A Delta table overwritten on each deployment
- Overwriting a table also loses the history.
- C. A prompt pasted into a notebook
- A pasted prompt cannot be promoted or rolled back.