Security Documentation Index¶
Maps the ten documents ENAC-IT requires for this service to where each one actually lives. Most were already written as ordinary engineering docs; this page is what makes them findable as a set. Start here when answering a security questionnaire or an audit.
Reading time: ~6 minutes
The ten required documents¶
| # | Required document | Where it lives | State |
|---|---|---|---|
| 1 | Encryption keys management | Encryption and Key Management | ✅ Complete |
| 2 | Operating procedures | Operations, Release runbook, Infra overview | ✅ Complete |
| 3 | Change management | Guardrails § Workflow, Release management, Workflow guide | ✅ Complete |
| 4 | Malicious code detection | ADR-014 §1, §2, §6 — Dependabot, CodeQL, secret scanning, Trivy | ✅ Complete |
| 5 | Vulnerability monitoring | Security policy (CodeAuditor), ADR-014 §1, §2, CI/CD workflows | ✅ Complete |
| 6 | Third-party list | Third parties below; dependency-level detail in the repository's GitHub dependency graph | ✅ Complete |
| 7 | Incident response | Incident Response — severity tiers, communication timeframes, confidentiality, and the personal-data notification; responder roster in the Disaster Recovery Plan (private ops repository) | ✅ Complete |
| 8 | Business continuity plan | Disaster Recovery Plan (private ops repository) — recovery team, namespace and bucket re-provisioning, secret recovery, GitOps and ArgoCD restore, database restore, manual build path, monitoring recovery | ✅ Complete |
| 9 | Maintenance and restore | Release runbook, Disaster Recovery Plan (private ops repository), backup section of Infra overview, Recovery objectives | ✅ Complete |
| 10 | Compliance procedures | EPFL compliance mapping, EPFL constraints | ✅ Complete |
What is still open¶
- Recovery timeframes (9) — best effort for what we control, documented under Recovery objectives. Detection is measured and fast; database RPO/RTO is now covered by EPFL DSI's own SLA (see below), but deleted objects are still not recoverable at all — that is a decision to take, not code to write.
- Object storage support queue — every other DSI dependency has a
named
SI_queue in the service map; object storage does not. Decided 2026-09-07: open the ticket at1234@epfl.chwith the bucket name. Replace with theSI_queue if DSI ever publishes one.
That gap is not closed by a code change. Track it as an issue rather than leaving it implied by an unticked box.
Recovery objectives¶
Best effort for what we control. Detection, notification and application restore have no contractual SLA — that's ours to own. Database recovery follows EPFL DSI's own terms: see the EPFL PostgreSQL SLA for backup frequency, RPO, RTO and the point-in-time-recovery window — link it rather than copying its figures here, since DSI can revise them without this doc knowing. What is measured on our side:
| Stage | Actual capability |
|---|---|
| Detection | 1–10 minutes. PodOOMKilled fires at 1 min; PodNotReady, PVCAlmostFull and ImagePullBackOff at 5 min; BackendMetricsAbsent, the total-outage detector, at 10 min |
| Application symptoms | 5 minutes — latency P50/P95/P99 and error-rate alerts; upload and job SLO breaches at 15 min |
| Notification | ~30 s after an alert fires, by email to the sysadmin group; re-sent every 4 h until resolved, and again on recovery |
| External check | Icinga probes HTTPS, certificate validity and the database from outside the cluster |
| Restore, application | Minutes — ArgoCD auto-sync and self-heal reconcile a bad deploy; revert the GitOps commit and it rolls back |
| Restore, database | Requested from EPFL DSI by a ticket in SI_POSTGRESQL (SVC1757). RPO, RTO and the PITR window are set by the EPFL PostgreSQL SLA, not by us |
| Restore, files | Not possible, by decision — buckets have versioning disabled. Object storage stages ingestion files; the durable record is in PostgreSQL, and a lost CSV is re-uploaded rather than restored |
Detection and application recovery are ours and are fast. Database recovery follows the EPFL PostgreSQL SLA — refer to it directly for current figures rather than a value duplicated here.
Object storage is deliberately not versioned. It holds ingestion files
in transit (tmp/, processing/, processed/); the emission data they
produce lives in PostgreSQL and is backed up there. Paying for version
history on a staging area would protect a copy, not the record.
Alert rules and routing live in the private ops repository under
overlays/<env>/monitoring/.
Service map and DSI queues¶
Moved to Operations § Symptom to owner,
the page an operator opens during an incident. It holds the service map,
the DSI service numbers and support queues, and the note that object
storage has no published SI_ queue.
Third parties¶
Services this deployment depends on:
| Service | Role |
|---|---|
| Microsoft Entra ID | Authentication and authorization (OIDC) |
| EPFL DSI PostgreSQL | Application database |
| EPFL S3 | Object storage for uploads and ingestion files |
| ENAC-IT Infisical | Secret and encryption-key vault, all environments |
| EPFL Tableau | Source of travel and headcount data |
| European Central Bank | Exchange-rate reference data |
| GitHub | Source hosting, CI/CD, image registry, scanning |
| OpenShift / Kubernetes | Runtime platform |
DSI-run services carry a service number and a support queue — see the service map for which one to ticket.
Library and package dependencies are not listed here — they change too often for a hand-maintained list to stay honest. The GitHub dependency graph is the source of truth, and Dependabot watches it.
Ownership and review cadence¶
| Document | Owner | Reviewed |
|---|---|---|
| This index | Lead developer | Annually, and on any change |
| Encryption and key management | Lead developer | On any encryption change |
| Operating and restore docs | Lead developer | Before each major release |
| Third-party list | Project manager | Quarterly |
| Incident response, BCP | Project manager | Annually |
A document that changes without this index changing is the failure mode this page exists to prevent — update both in the same PR.