# Split-Horizon DNS Implementation Guide

Last updated: 2026-03-20

## Problem

All DNS queries for `*.vish.gg`, `*.thevish.io`, and `*.crista.love` currently resolve to Cloudflare proxy IPs (104.21.x.x), even when the client is on the same LAN as the services. This means:

1. **Hairpin NAT** — LAN traffic goes out to Cloudflare and back in through the router
2. **Internet dependency** — if the WAN link goes down, LAN services are unreachable by domain
3. **Added latency** — ~50ms roundtrip through Cloudflare vs ~1ms on LAN
4. **Cloudflare bottleneck** — all traffic proxied through CF even when unnecessary

## Solution

**Status: IMPLEMENTED (2026-03-20)**

Use AdGuard Home on Calypso (primary) and Atlantis (backup) as **split-horizon DNS resolvers** that return local IPs for homelab domains when queried from the LAN, while external clients continue to use Cloudflare.

```
                    ┌──────────────────────────────────┐
                    │         DNS Query for             │
                    │         nb.vish.gg                │
                    └───────────────┬──────────────────┘
                                    │
                    ┌───────────────▼──────────────────┐
                    │      Where is the client?         │
                    └───────┬───────────────┬──────────┘
                            │               │
                   LAN Client          External Client
                            │               │
                            ▼               ▼
                 ┌──────────────┐   ┌──────────────┐
                 │ AdGuard Home │   │  Cloudflare  │
                 │ (Calypso +  │   │   DNS        │
                 │  Atlantis)  │   │              │
                 │ Returns:    │   │ Returns:     │
                 │100.85.21.51 │   │ 104.21.73.214│
                 │(NPM Tailscale)│  │ (CF proxy)   │
                 └──────┬───────┘   └──────┬───────┘
                        │                  │
                        ▼                  ▼
                 ┌──────────────┐   ┌──────────────┐
                 │ NPM (local)  │   │  Cloudflare  │
                 │ matrix-ubuntu│   │  → WAN IP    │
                 │ :443 ~1ms   │   │  → NPM       │
                 └──────┬───────┘   │  ~50ms       │
                        │           └──────┬───────┘
                        ▼                  ▼
                 ┌─────────────────────────────────┐
                 │        Backend Service           │
                 │   (same result, faster path)     │
                 └─────────────────────────────────┘
```

## Prerequisites

NPM is now on matrix-ubuntu (192.168.0.154) listening on standard ports 80/443/81. The migration from Calypso was completed on 2026-03-20.

| Port | Status |
|------|--------|
| 80:80 | **Active** |
| 443:443 | **Active** |
| 81:81 | **Active** (Admin UI) |

## Implementation Steps

### Step 1: Move NPM to Standard Ports -- DONE

NPM migrated from Calypso to matrix-ubuntu (192.168.0.154) on 2026-03-20. Compose file: `hosts/vms/matrix-ubuntu/nginx-proxy-manager.yaml`. Host nginx on matrix-ubuntu has been disabled (`systemctl disable nginx`); NPM now handles mastodon.vish.gg, mx.vish.gg, and mm.crista.love directly.

Router port forwards updated:
- `WAN:443 → 192.168.0.154:443`
- `WAN:80 → 192.168.0.154:80`

### Step 2: Configure AdGuard DNS Rewrites -- DONE

AdGuard DNS rewrites configured on both Calypso (http://192.168.0.250:9080) and Atlantis (http://192.168.0.200:9080). Wildcard entries point to NPM's Tailscale IP:

| Domain | Answer | Notes |
|--------|--------|-------|
| `*.vish.gg` | `100.85.21.51` | All vish.gg domains → NPM Tailscale IP |
| `*.thevish.io` | `100.85.21.51` | All thevish.io domains → NPM Tailscale IP |
| `*.crista.love` | `100.85.21.51` | All crista.love domains → NPM Tailscale IP |

These three wildcards cover all 36 proxy hosts. AdGuard resolves matching queries locally instead of forwarding to upstream DNS.

**Exceptions** — these domains need direct IPs (not NPM), added as specific overrides:

| Domain | Answer | Reason |
|--------|--------|--------|
| `mx.vish.gg` | `192.168.0.154` | Matrix federation needs direct access on port 8448 |
| `derp.vish.gg` | `192.168.0.250` | DERP relay — direct IP, no CF proxy |
| `derp-atl.vish.gg` | `192.168.0.200` | Atlantis DERP relay |
| `headscale.vish.gg` | `192.168.0.250` | Headscale control — direct access |
| `turn.thevish.io` | `192.168.0.200` | TURN/STUN needs direct UDP |

**.tail.vish.gg overrides** — specific rewrites to override the wildcard for Tailscale-specific subdomains.

Specific entries take priority over wildcards in AdGuard.

### Step 3: Set AdGuard as LAN DNS Server -- DONE

Router (Archer BE800) DHCP configured with dual AdGuard DNS:

1. **Primary DNS:** `192.168.0.250` (Calypso AdGuard)
2. **Secondary DNS:** `192.168.0.200` (Atlantis AdGuard, backup)

### Step 4: Configure Atlantis AdGuard (Backup DNS) -- DONE

Same DNS rewrites added to Atlantis's AdGuard instance (http://192.168.0.200:9080) as backup:

- Same wildcard rewrites as Calypso (pointing to `100.85.21.51`)
- Reachable at `192.168.0.200`

### Step 5: Test

```bash
# Verify local resolution
dig nb.vish.gg @192.168.0.250
# Expected: 192.168.0.250 (NPM local IP)

# Verify external resolution still works
dig nb.vish.gg @1.1.1.1
# Expected: 104.21.73.214 (Cloudflare proxy)

# Test HTTPS access via local DNS
curl -s --resolve "nb.vish.gg:443:192.168.0.250" https://nb.vish.gg/ -o /dev/null -w "%{http_code} %{time_total}s\n"
# Expected: 200 in ~0.05s (vs ~0.15s through Cloudflare)

# Test all domains resolve locally
for domain in nb.vish.gg gf.vish.gg git.vish.gg sso.vish.gg dash.vish.gg; do
    ip=$(dig +short $domain @192.168.0.250 | tail -1)
    echo "$domain → $ip"
done
```

## SSL Considerations

**Resolved:** NPM now uses **Let's Encrypt wildcard certificates** (DNS challenge via Cloudflare API) instead of Cloudflare Origin certs. This means:

- Certs are trusted by all browsers, whether traffic comes through Cloudflare or directly via LAN
- No browser warnings for split-horizon DNS LAN access
- Certs auto-renew via NPM's built-in Let's Encrypt integration

## What Changes for Each Path

### LAN Client
```
Browser → nb.vish.gg
  → AdGuard DNS: 100.85.21.51 (NPM Tailscale IP)
  → NPM (matrix-ubuntu:443) → SSL termination (LE wildcard cert)
  → Proxy to backend (192.168.0.210:8443)
  → Response (~1ms total DNS+proxy)
```

### External Client
```
Browser → nb.vish.gg
  → Cloudflare DNS: 104.21.73.214
  → Cloudflare proxy → WAN IP → Router
  → NPM (matrix-ubuntu:443) → SSL termination
  → Proxy to backend (192.168.0.210:8443)
  → Response (~50ms total)
```

### Internet Down
```
Browser → nb.vish.gg
  → AdGuard DNS: 100.85.21.51 (cached/local)
  → NPM (matrix-ubuntu:443) → SSL termination
  → Proxy to backend
  → Response (services still work!)
```

## Current NPM Proxy Hosts (for reference)

All 36 domains that would benefit from split-horizon:

### vish.gg (27 domains)
| Domain | Backend |
|--------|---------|
| actual.vish.gg | calypso:8304 |
| cal.vish.gg | atlantis:12852 |
| dash.vish.gg | atlantis:7575 |
| dav.vish.gg | calypso:8612 |
| docs.vish.gg | calypso:8777 |
| gf.vish.gg | homelab-vm:3300 |
| git.vish.gg | calypso:3052 |
| headscale.vish.gg | calypso:8085 |
| kuma.vish.gg | rpi5:3001 |
| mastodon.vish.gg | matrix-ubuntu:3000 |
| mx.vish.gg | matrix-ubuntu:8082 |
| nb.vish.gg | homelab-vm:8443 |
| npm.vish.gg | calypso:81 |
| ntfy.vish.gg | homelab-vm:8081 |
| ollama.vish.gg | atlantis:11434 |
| ost.vish.gg | calypso:3000 |
| paperless.vish.gg | calypso:8777 |
| pt.vish.gg | atlantis:10000 |
| pw.vish.gg | atlantis:4080 |
| rackula.vish.gg | calypso:3891 |
| retro.vish.gg | calypso:8025 |
| rx.vish.gg | calypso:9751 |
| rxdl.vish.gg | calypso:9753 |
| scrutiny.vish.gg | homelab-vm:8090 |
| sf.vish.gg | calypso:8611 |
| sso.vish.gg | calypso:9000 |
| wizarr.vish.gg | atlantis:5690 |

### thevish.io (5 domains)
| Domain | Backend |
|--------|---------|
| binterest.thevish.io | homelab-vm:21544 |
| hoarder.thevish.io | homelab-vm:3482 |
| joplin.thevish.io | atlantis:22300 |
| matrix.thevish.io | matrix-ubuntu:8081 |
| meet.thevish.io | atlantis:5443 |

### crista.love (2 domains)
| Domain | Backend |
|--------|---------|
| crista.love | guava:28888 |
| cocalc.crista.love | guava:8080 |
| mm.crista.love | matrix-ubuntu:8065 |

## Rollback

If something breaks:
1. Change router DHCP DNS back to `1.1.1.1` / `8.8.8.8`
2. Or remove the DNS rewrites from AdGuard
3. All traffic reverts to Cloudflare path immediately

## Related Documentation

- [NPM Migration](npm-migration-jan2026.md) — Reverse proxy configuration
- [Authentik SSO](authentik-sso.md) — Forward auth depends on NPM routing
- [Cloudflare DNS](cloudflare-dns.md) — External DNS records
- [Image Update Guide](../admin/IMAGE_UPDATE_GUIDE.md) — Mentions Gitea/NPM as bootstrap dependencies