Upgrades and Rollback¶
Atlas upgrades should preserve two invariants:
- contract-owned surfaces remain understood and validated
- serving state stays recoverable if a rollout goes wrong
Upgrade Flow¶
flowchart TD
Validate[Validate runtime and contracts] --> Rollout[Roll out new runtime]
Rollout --> Observe[Observe health and load]
Observe --> Keep[Keep rollout]
Observe --> Rollback[Rollback if needed]
This upgrade flow keeps rollout discipline visible. Atlas upgrades should be validated, observed, and explicitly kept or rolled back rather than treated as one-way jumps.
Rollback Flow¶
flowchart LR
Problem[Operational problem] --> Scope[Determine runtime vs store scope]
Scope --> RuntimeRollback[Rollback runtime]
Scope --> StoreRollback[Rollback serving state if required]
This rollback flow explains one of the most important operator distinctions in Atlas: not every incident needs store-state rollback, and not every rollback should start there.
Operator Guidance¶
- separate runtime rollback from store-state rollback in your thinking
- verify health, readiness, and key query paths after rollout
- keep rollback paths explicit before you need them
- use compatibility and contract evidence as rollout input, not only hope and manual spot checks
What to Watch During Upgrade¶
- readiness instability
- unusual rejection or error patterns
- metrics or traces indicating saturation changes
- catalog or dataset discoverability regressions
Rollout Question That Saves Time¶
Ask first whether the change affected runtime behavior, serving-store state, or both. That answer usually determines the safest rollback path.
Purpose¶
This page explains the Atlas material for upgrades and rollback and points readers to the canonical checked-in workflow or boundary for this topic.
Stability¶
This page is part of the canonical Atlas docs spine. Keep it aligned with the current repository behavior and adjacent contract pages.