AI Code Completion Service · View 11 of 21 · Data
Decisions
- RPO is 0 for policy, seats, consent and the model registry, and undefined for prefix caches — not because they do not matter, but because recomputation is the recovery mechanism.
- Retrieval indexes carry no backup: their recovery objective is a rebuild time against the repository that owns the code.
- The platform owns no source code. Every store that could hold it is ephemeral, opt-in, or a projection of a repository someone else owns.
Assumptions
- 40 TB of retrieval indexes across active workspaces; 6 TB/day of telemetry compressed.
- RTO 5 minutes to serve from another region with cold caches; 60 minutes to rebuild the largest workspace index.
Risks
- A rebuildable store with a 60-minute rebuild is a 60-minute degradation, not a non-event: during it, retrieval degrades to open-tabs context and acceptance measurably falls.