public/docs/architecture-overview.md

📐 Architecture Overview

Note

This document defines the invariant movement of traffic between infrastructure components — directional flow and authority boundaries, not configuration syntax, ports, or IP addresses. Eight distinct traffic categories, each stated as a Law (what must always be true), a Flow (the actual path), and a Failure Behavior (what happens when the target is unreachable).

Dependency graph

flowchart LR
    Internet(["Internet"]) --> Edge["Edge Firewall"]
    Edge --> LAN["LAN"]

    LAN --> DNS["Internal DNS"]
    LAN --> Vault["Vault"]
    DNS --> Vault

    Vault --> Apps["App Host<br/>~40 stacks"]
    Vault --> Automation["Automation Host"]
    LAN --> FileServer["File Server"]
    DNS --> FileServer
    LAN --> Automation

    Apps --> InternalProxy["Internal Proxy Zone"]
    Apps --> SSO["SSO"]
    Automation --> GitHost["Git Hosting"]

    Vault --> EdgeProxy["User-Facing Proxy<br/>mTLS"]
    DNS --> EdgeProxy
    Apps -.proxied.-> EdgeProxy
    SSO -.auth.-> EdgeProxy

    Apps -.-> Monitoring["Monitoring"]
    Automation -.-> Monitoring
    FileServer -.-> Monitoring
    Vault -.-> Monitoring
    DNS -.-> Monitoring
    EdgeProxy -.-> Monitoring

    Apps -.-> SecMon["Security Monitoring"]
    Automation -.-> SecMon
    FileServer -.-> SecMon
    Vault -.-> SecMon
    DNS -.-> SecMon
    EdgeProxy -.-> SecMon

Tip

Solid arrows are real dependencies — the target has to be reachable and healthy for the source to function. Dotted arrows are asynchronous reporting, not a blocking dependency — the distinction Law 5 below makes explicit.

The eight traffic laws

1. Ingress — two independent authorities

Law: Public traffic and LAN-facing user traffic terminate at different places and are never conflated.

	Public	LAN-facing (user apps)
Path	Client → WAN → Edge Firewall → Backend	Client → LAN → User-Facing Proxy (mTLS + SSO) → Backend
TLS terminates at	The edge firewall's own reverse-proxy package	The user-facing proxy
Enforcement	No backend is ever directly exposed to WAN — only HTTPS/HTTP and one admin SSH port pass the WAN interface	Mutual TLS (client cert required) + SSO

Failure behavior: if the edge firewall fails, public access is unavailable and everything else keeps running. If the user-facing proxy fails, all browser-facing access is unavailable and everything else keeps running. The two failures are structurally independent — one can't cascade into the other.

2. East-West — two genuinely distinct patterns

Law: App-to-app HTTP/S traffic and non-HTTP infra-to-infra traffic are not interchangeable, and neither uses a flat unauthenticated mesh (see "what changed," below).

Pattern	Path	Trust
(a) App-to-app HTTP/S	Service → internal proxy zone → trusted-IP bypass or SSO forward-auth → Service	IP-allowlist based (docker bridge subnets + every fleet host), not mTLS — mTLS is reserved exclusively for Law 1's LAN-facing ingress
(b) Infra-to-infra, non-HTTP	Service → firewall rule (explicit per-flow allow) → Service, direct, no proxy involved	Metrics/logs, MQTT, security-agent reporting, NTP, Vault Agent auth all take this path

Failure behavior: if internal DNS fails, app-to-app resolution fails and that path breaks with it. Direct IP communication may still work for pattern (a) but bypasses the SSO gate entirely if the source IP happens to be trusted — a real, accepted risk surface, not an oversight. Core service runtime doesn't depend on the public-facing edge (Law 1) at all — completely separate plane.

3. Secrets — one hop, never direct

Law: Applications do not communicate directly with Vault.

Flow: Service → local Vault Agent → Vault Server. Vault Agent performs AppRole authentication and token renewal; in practice it renders local files (an env file consumed via Compose, or a full rendered config file where multi-line config is needed) that the application reads. Applications never call the Vault API themselves, and Vault is never exposed directly to application containers.

Failure behavior: if Vault becomes unavailable, secret renewal stops but services continue operating until their current token expires — no immediate disruption, a slow-building one only if the outage outlasts every live token's TTL.

4. DNS — resolution determines the routing plane

Law: Which internal DNS zone a hostname resolves in determines which traffic plane (Law 1 or Law 2) a request enters.

Full model — the two zones, the enforcement that makes this a real boundary rather than a convention, and the audit that caught real drift in how it was documented: docs/dns-naming.md.

5. Observability — asynchronous, never inline

Law: Logs and metrics push asynchronously to monitoring infrastructure. Monitoring systems are not inline with application traffic.

Flow: Service → observability agent → Monitoring (logs, metrics). Service → security agent → Security Monitoring. Every managed host runs both agents, enforced by automation that self-heals on drift rather than being applied once and left alone.

Important

Monitoring infrastructure does not proxy or mediate application traffic, and service runtime does not depend on monitoring availability — this is why these are the dotted, not solid, arrows in the dependency graph above. Failure behavior: if monitoring fails, log/metric collection stops; application services continue operating normally, completely unaffected.

Full detail on what's actually collected and why push beats pull here specifically: docs/observability.md and docs/security-monitoring.md.

6. Notifications — two branches, one non-critical destination

Law: Notifications converge on one flow and are non-critical to service operation — neither branch is part of any service's runtime dependency chain.

Branch	Path
MQTT	Service → local notify script → message broker (per-host topic) → central subscriber
HTTP webhook	App → shared webhook endpoint (App Router + per-app mapper) or a dedicated endpoint for non-JSON payloads → same downstream formatter

Both branches converge on the same formatting step before delivery, but reach it differently — the MQTT branch derives its destination channel from the topic name, the webhook branch's per-app mapper functions set it explicitly. Failure behavior: notification failures never impact core workloads — this is event-driven, not persistent inline execution, so a dropped notification is lost, not queued against a retry that could back up into the sender.

Full mechanics: docs/notifications.md.

7. CI/CD — deployment traffic is separate from runtime traffic

Law: Build and deploy activity never sits inline with application runtime.

Flow: Scheduler (weekly) → build automation → self-hosted registry → target host redeploy.

Invariants: CI/CD components always initiate connections outward, never accept inbound triggers from the apps they build. Runtime services do not depend on CI/CD availability. Deployment failure doesn't affect already-running services — enforced by the redeploy step's own health-check/rollback logic, not just incidentally true. This is schedule- and upstream-digest-triggered, not push-triggered — editing a stack's own build file doesn't by itself trigger a rebuild until the next scheduled sweep finds the upstream image has actually changed.

Full mechanics and the reasoning behind schedule-triggered over push-triggered: docs/registry-and-build.md.

8. Management — administrative access is its own separate plane

Law: Administrative access is explicitly controlled and restricted to approved management interfaces, separate from every other plane above.

Flow: Administrator → automation host → a dedicated, key-based service identity → target host, on a non-standard port.

Invariants: management interfaces are restricted to approved, non-standard ports fleet-wide, never the service's default port. Access is limited by explicit per-host-pair firewall rules sourced from the automation host only — not a blanket LAN allow. Administrative access originates only from the automation host, via one dedicated identity, never from arbitrary internal sources. Application services do not depend on management interfaces being reachable.

Failure behavior: if management access fails, administration becomes unavailable but every running service continues operating normally — the same "control plane failure doesn't touch data plane" property that holds across every law above.

Global architectural laws

These apply across every category above, not just one:

Law	Statement
Routing authority	Three routing authorities exist, each scoped to a distinct plane — the edge firewall (public ingress), the user-facing proxy (LAN ingress), the internal proxy zone (app-to-app east-west). No internal service ever acts as its own router.
Trust boundary	Trust is segmented across four boundaries: WAN→edge (public), LAN browsers→user-facing proxy (mTLS+SSO), app-to-app→internal proxy (allowlist+SSO), LAN infra-to-infra→firewall rule (explicit per-flow). Backend services trust proxy headers, never raw client IP, on any path.
Plane separation	Two structurally independent traffic planes — Ingress (public and LAN-facing sub-paths) and East-West (app-to-app and infra-to-infra sub-paths) — with no crossover between them.
Observability independence	Observability never sits inline in any request path; it's asynchronous and push-based, full stop (Law 5).

flowchart LR
    WAN["WAN"] -->|public| EdgeFW["Edge Firewall<br/>public proxy"]
    Browser["LAN browsers"] -->|mTLS + SSO| UserProxy["User-facing proxy"]
    Service["App-to-app"] -->|allowlist + SSO| InternalProxy["Internal proxy"]
    Infra["LAN infra-to-infra"] -->|per-flow rule| HostFW["Host firewall"]

Two real architecture changes, not a static design

Important

This shape isn't how the fleet started — it's evolved twice, deliberately. The public-facing reverse proxy used to run as its own dedicated VM; it was consolidated directly onto the edge firewall as a package instead, removing an entire host from the public-ingress path without losing any functionality — one less thing that can fail, one less thing to patch and monitor independently. And Law 2's internal service-to-service model used to be a flat, unauthenticated mesh network with no trust boundary at all; it was fully replaced by the two-zone split now described in docs/dns-naming.md, trading a simpler-but-weaker model for one with a real trust boundary between browser-facing and app-to-app traffic. Both changes reduced the attack surface at the cost of more moving parts to reason about — a tradeoff made deliberately, not stumbled into.

What this map deliberately excludes

Not a firewall rule listing, not a port inventory, not a service configuration reference — those live in docs/firewall-and-network.md and the other docs linked throughout each law above. This is strictly the stable, directional shape of how trust and traffic actually move — the eight laws above are meant to still be true even after the specific tools implementing them change.

---

Related: docs/firewall-and-network.md, docs/dns-naming.md, docs/secrets-and-pki.md, docs/notifications.md, docs/registry-and-build.md