TenantAtlas/specs/051-entra-group-directory-cache/research.md
2026-01-11 22:02:06 +01:00

43 lines
2.1 KiB
Markdown

# Research: Entra Group Directory Cache (Groups v1)
## Decisions
### Decision: Full tenant scope sync (all groups)
- Rationale: Ensures browse/search is complete and name-resolution has the best chance of resolving IDs across modules.
- Alternatives considered:
- Only sync referenced groups (creates gaps, hard to debug, non-deterministic coverage).
### Decision: Start modes = manual + scheduled
- Rationale: Manual sync supports operator workflows (pre-restore/triage), while scheduled sync keeps cache fresh without relying on user action.
- Alternatives considered:
- Manual only (cache freshness depends on operator discipline).
- Scheduled only (harder to react quickly during ops).
### Decision: App-only (service principal) directory reads
- Rationale: Scheduled runs must not depend on an interactive user session; app-only is simpler to audit/lock down.
- Alternatives considered:
- Delegated tokens (breaks scheduled runs, unpredictable auth state).
### Decision: Cache metadata only (no membership/owners)
- Rationale: Solves the UX problem (name resolution + browse) with minimal data/PII exposure and lower API volume.
- Alternatives considered:
- Cache membership/owners (larger data surface, higher sensitivity, more Graph calls).
### Decision: Missing groups retention and purge
- Decision: Retain groups for 90 days after `last_seen_at`, then purge.
- Rationale: Preserves investigatory value and audit context while bounding DB growth.
- Alternatives considered:
- Immediate delete (breaks historical triage; creates sudden “unresolved” labels).
- Retain forever (unbounded growth).
### Decision: Graph access patterns (paging + retry)
- Decision: Use Graph list endpoint with paging; apply retry/backoff for 429/503; persist stable error categories in run record.
- Rationale: Full-tenant enumeration can be large; throttling is expected.
- Alternatives considered:
- Best-effort without retry (unreliable coverage).
- UI-time lookups (explicitly disallowed by spec).
## Open Questions (resolved for planning)
- Which specific Graph endpoint and permission names to use will be documented in the contract registry during implementation.