Skip to content

ScreamingFace deployment decisions

This page is the developer-facing index of decisions that shape the current deployment. The linked ADRs and manifests remain authoritative; this page records their practical consequence for ScreamingFace.

Decision Practical consequence Authority
Each tenant receives a fenced Argo project and named namespaces ScreamingFace applications cannot deploy outside approved sf-* namespaces or create cluster-scoped resources ADR-0004
The platform owns public routing and identity enforcement Tenant charts cannot create or weaken Gateway or SecurityPolicy resources ADR-0005
The tenant chart owns workloads; PULSE owns the trust boundary Charts provide Deployments, Services, and Jobs; PULSE provides Secrets, storage, ServiceAccounts, RBAC, NetworkPolicy, and routes ADR-0006
Platform telemetry uses SigNoz as one logs, traces, and metrics backend ScreamingFace logs are collected without application changes; direct developer access waits for an explicit data-scoping decision ADR-0007
Charts are read over HTTPS with a GitHub App No personal token or deploy key is the steady-state Argo credential ADR-0006
Images are pushed by tenant CI over Azure OIDC No Azure client secret or registry password is stored; Kargo can discover the accepted artifact Delivery
Chart and image are pinned to the same exact tenant commit Kargo derives both pins from one Freight and required checks gate its automatic merge ADR-0008
Chart-generated runtime Secrets are disabled Credentials remain stable across Argo renders and arrive through Key Vault plus ESO Credentials
Credential record names are generation-scoped Rotation and parallel stacks use cutover, not deletion or in-place renaming ADR-0006
url4-cloud keeps an API token and exact namespaced RBAC It can create Runner Jobs but cannot read Secrets or widen its own permissions url4-cloud
aigateway and its migration have no API token They have no Kubernetes API use, so a projected token is unnecessary attack surface aigateway
aigateway trusts an edge-injected user header only on a private network It has no public route; caller and egress policies are part of authentication Architecture
/ws has no route timeout; REST is capped at 90 seconds Long-lived event streams survive proxy upgrades while slow REST fails clearly before the outer edge timeout url4-cloud
url4-cloud remains one replica in dev Its subscriber gate is process-local and is not yet safe across replicas url4-cloud
The old trial was retired after post-cutover proof Rollback is an immutable repin/redeploy, not a route switch to the trial Evidence and status
Pre-production and production are separate future environments Dev secrets, database posture, proof, and exceptions cannot be promoted as-is Quick rundown

Decisions still open

These are not hidden implementation details; they block stronger service claims or promotion:

  1. Align url4-cloud model declarations with aigateway's provider-prefixed canonical identifiers.
  2. Define how a verified user receives an approved provider profile and an enforceable spend cap.
  3. Make subscriber readiness depend on successful ai.url4.attach, adopt auto-attach, or move attach state out of process before scaling url4-cloud.
  4. Narrow the Azure federation to intended subjects and reduce ACR push scope before pre-production.
  5. Design pre-production and production data protection, availability, observability, capacity, promotion, and rollback gates.

When one of these decisions is resolved, update the governing ADR or create a new one, update the relevant tenant page, and attach fresh evidence. Do not silently convert an open constraint into an assumed behavior.

Reading order