A $29.60 Fly bill traced to ~1.25 TB/30d egress on forge.eblu.me (99.95% of
all proxy egress), ~71% of it AI scrapers (Meta meta-externalagent, OpenAI
GPTBot, Amazonbot, Bytespider) crawling the public mirror repos' infinite
git-history URL space and timing out Forgejo. robots.txt already disallowed
/mirrors/ but those agents ignore it, so enforce at the edge: return 403 (^~
to beat the regex asset locations), served as a roll-of-dishonour page with an
X-Naughty-Scrapers header. Mirrors stay reachable on the tailnet via
forge.ops.eblu.me. Tier 2 (UA denylist + Anubis) and the Cloudflare rejection
are documented in docs/explanation/ai-scraper-mitigation.md.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
## Summary
Brings up the Adelaide / Heidi / Addie baby shower app on ringtail k3s with the public/private split that the app's hosting contract calls for: `shower.eblu.me` (public, via Fly proxy) and `shower.ops.eblu.me` (tailnet). App is consumed as a wheel from the Forgejo PyPI index — source lives at [`adelaide-baby-shower-app`](https://forge.eblu.me/eblume/adelaide-baby-shower-app).
### What's included
- **ArgoCD app + manifests** under `argocd/manifests/shower/` (deployment, service, ProxyGroup ingress, ConfigMap for `DJANGO_DEBUG`/`DJANGO_ADMIN_URL`, ExternalSecret for `DJANGO_SECRET_KEY` from 1Password item `Shower (blumeops)`, NFS PV on sifaka, RWX media PVC, RWO local-path data PVC for SQLite). Recreate rollout because SQLite is single-writer.
- **Public surface** (`fly/`): new `shower.eblu.me` server block proxying to `shower.ops.eblu.me`. `/admin/` returns 403 at the edge except `/admin/login/` and `/admin/logout/`, which are rate-limited via a new `shower_auth` zone. `X-Clacks-Overhead` on. GNU Terry Pratchett.
- **fail2ban** filter (`shower-admin-login.conf`) matching 401/403/429 on `/admin/login/` and jail (`shower.conf`) with `maxretry=5/findtime=600/bantime=3600`. The `nginx-deny` action was generalized to take a per-jail `nginx_deny_file` so the shower has its own deny list (forge keeps using the legacy default).
- **Caddy** route on indri (`shower.ops.eblu.me` → `https://shower.tail8d86e.ts.net`).
- **Pulumi** Gandi CNAME `shower.eblu.me → blumeops-proxy.fly.dev.`.
- **Grafana** APM dashboard `configmap-shower-apm.yaml` (request rate, error rate, failed admin login count, latency percentiles, bandwidth, access logs) mirroring `docs-apm.json` with a `host="shower.eblu.me"` filter.
- **Container** `containers/shower/default.nix` — `dockerTools.buildLayeredImage` with a nixpkgs Python and a startup wrapper that creates `/app/data/.venv`, pip-installs `adelaide-baby-shower-app==1.0.0` from the forge PyPI index on first boot, runs migrations + collectstatic, and execs gunicorn. A `local_settings.py` shim pins `DATABASES.NAME`/`MEDIA_ROOT`/`STATIC_ROOT` to absolute paths so they don't end up in site-packages.
- **Docs** runbook at `docs/how-to/operations/shower-app.md` linked from the apps registry, plus changelog fragments.
### Defense layers on the public surface
1. fly nginx geo+fail2ban `$shower_banned` (per-service deny list)
2. fly nginx `limit_req zone=shower_auth` (3 r/s per Fly-Client-IP)
3. django-axes (5 fails / 1h, keyed on username+ip_address)
4. edge `/admin/` block (returns 403 for anything that isn't login/logout)
## Prerequisites for the user to do (NOT in this PR)
Halted on these per request — they touch shared/manual systems:
- [x] **NFS share** on sifaka: `/volume1/shower`, NFS rule for ringtail RW, `chown 1000:1000`
- [ ] **1Password item** `Shower (blumeops)` in the blumeops vault with a freshly minted `secret-key` field (`openssl rand -base64 48`) — do NOT reuse anything that has lived in git
- [ ] **Container build**: `mise run container-build-and-release shower`, then update `images[].newTag` in `argocd/manifests/shower/kustomization.yaml` to the resulting `v1.0.0-<sha>-nix`
- [x] **DNS**: `mise run dns-up` after merge
- [x] **Fly cert**: `fly certs add shower.eblu.me -a blumeops-proxy`
- [ ] **Caddy push**: `mise run provision-indri -- --tags caddy`
- [ ] **Fly redeploy** to pick up the new nginx block + fail2ban jail: `mise run fly-deploy`
- [ ] **ArgoCD sync**: `argocd app set shower --revision shower-app-deploy && argocd app sync shower` to test from this branch before merging
## Test plan
- [ ] Container builds successfully on nix-container-builder runner
- [ ] Pod starts, migrations run, gunicorn answers on :8000
- [ ] `kubectl --context=k3s-ringtail -n shower logs deploy/shower` clean
- [ ] `curl -sf https://shower.ops.eblu.me/` returns the splash page (tailnet)
- [ ] `curl -I -H "Host: shower.eblu.me" https://blumeops-proxy.fly.dev/` returns 200 (pre-DNS verification)
- [ ] `curl -I -H "Host: shower.eblu.me" https://blumeops-proxy.fly.dev/admin/users/` returns 403 (edge block)
- [ ] `curl -I -H "Host: shower.eblu.me" https://blumeops-proxy.fly.dev/admin/login/` returns a Django login response
- [ ] After DNS is up: `curl -I https://shower.eblu.me/` returns 200 with `X-Clacks-Overhead`
- [ ] Grafana dashboard "Shower APM" appears and starts showing traffic
- [ ] `mise run services-check` passes
Reviewed-on: #349
The static asset cache block (css/js/png/etc) was missing
proxy_set_header Host, so Caddy received "forge.eblu.me" instead of
"forge.ops.eblu.me" and couldn't route the request. HTML loaded fine
because the main location / block had the header.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Tailscale Ingress pods in k8s can't establish direct WireGuard
connections (stuck behind pod-network NAT → DERP relay → 20s latency).
Indri's host-level Tailscale CAN peer directly with Fly.
Change all nginx upstreams to route through Caddy on indri instead of
per-service Tailscale Ingress endpoints. Tag indri as flyio-target in
the Tailscale ACL so the Fly proxy can reach it.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
## Summary
- Replace per-request DNS resolution (variable-based `proxy_pass`) with static `upstream` blocks and `keepalive` connection pools
- Reuses TLS connections through the Tailscale tunnel instead of handshaking per request
- Add `mise run fly-reload` for nginx config reload without full redeploy (re-resolves upstream DNS)
## Trade-off
DNS is resolved at config load, not per-request. If Tailscale Ingress pods get new IPs (restart, reschedule), `mise run fly-reload` is needed. A Grafana alert will be added to detect this.
## Still TODO on this branch
- [ ] Grafana alert for upstream unreachable (triggers fly-reload reminder)
- [ ] Docs pass
- [ ] Deploy from branch and verify latency improvement
- [ ] Changelog fragment
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Reviewed-on: #337
Was sending Connection: upgrade on every proxied request, which is
semantically wrong for normal HTTP traffic. Use a map to conditionally
send 'upgrade' only when the client requests a WebSocket switch,
'close' otherwise.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Crawlers follow auth redirects to /user/login which is pointless for them.
Saves round-trips for both sides.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Facebook has been scraping forge mirror repos at ~3-4 req/s, slowing
down the Forgejo instance. Serve robots.txt directly from nginx to
disallow /mirrors/ while leaving eblume/* accessible to crawlers.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The general rate limit zone used $binary_remote_addr (Fly's internal
proxy IP), causing all external clients to share one bucket. Switch to
$http_fly_client_ip to match forge_auth's correct behavior.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
## Summary
- Adds a branded 503 error page served when upstreams are unreachable (indri offline, Tailscale tunnel down, emergency shutoff, etc.)
- Stale cache is still served first when available (`proxy_cache_use_stale` takes priority)
- Test endpoint at `docs.eblu.me/_error` to preview the page without killing upstreams
- `proxy_intercept_errors on` also catches error responses returned by the upstream itself
## Files Changed
- `fly/error.html` — Self-contained error page (dark theme, links to BlumeOps repo)
- `fly/nginx.conf` — `error_page`, `internal` location, `/_error` test location, `proxy_intercept_errors`
- `fly/Dockerfile` — COPY error.html into image
## Test Plan
- [ ] Deploy to Fly.io
- [ ] Visit `docs.eblu.me/_error` to verify the page renders
- [ ] Optionally stop indri/Tailscale to confirm the page shows on real 502/503/504
Reviewed-on: https://forge.ops.eblu.me/eblume/blumeops/pulls/133
## Summary
- Start nginx after Tailscale connects (community best practice for Tailscale sidecars)
- Switch to `bluegreen` deploy strategy — old machine serves until new one is healthy
- Replace top-level `[checks]` with `[[http_service.checks]]` — only service-level checks gate traffic routing ([confirmed by Fly.io staff](https://community.fly.io/t/clarifying-the-types-of-health-checks/20379))
- Remove sentinel file and nginx if-check (no longer needed)
Supersedes the approach in #131 — that helped (502 window dropped from ~30s to ~3s) but couldn't fully eliminate it because top-level checks don't gate routing and Fly.io's proxy sends traffic as soon as the port is reachable.
## Deployment and Testing
- [ ] Merge and `fly deploy` from `fly/` directory
- [ ] Verify deploy completes with zero 502s (watch `fly logs` and Grafana docs-apm)
- [ ] Confirm `fly checks list` shows the new service-level check passing
Reviewed-on: https://forge.ops.eblu.me/eblume/blumeops/pulls/132
## Summary
- Health check (`/healthz`) now returns 503 until Tailscale is connected
- `start.sh` creates `/tmp/tailscale-ready` sentinel after `tailscale up` succeeds
- Fly.io keeps the old machine serving traffic during the ~7s startup window
Previously, nginx passed the health check immediately, Fly.io routed traffic to the new machine, but MagicDNS wasn't available yet — causing upstream DNS timeouts and 502s on every request until Tailscale connected.
## Deployment and Testing
- [ ] Merge and `fly deploy` from `fly/` directory
- [ ] Verify deploy completes with zero 502s (check Grafana docs-apm dashboard)
- [ ] Confirm health check transitions from 503 → 200 in `fly logs`
Reviewed-on: https://forge.ops.eblu.me/eblume/blumeops/pulls/131
## Summary
- Add `client_ip` field to the Fly.io nginx JSON log format, sourced from `Fly-Client-IP` header
- Extract `client_ip` in the Alloy pipeline so it's available as a parsed field in Loki
- Keeps `remote_addr` (the internal proxy IP) for debugging
Fixes: Grafana access logs for docs.eblu.me showing 172.16.11.178 for every request instead of real visitor IPs.
## Deployment and Testing
- [ ] Deploy updated fly.io proxy: `fly deploy` from `fly/` directory
- [ ] Verify in Grafana that new log lines include `client_ip` with real IPs
- [ ] Confirm `remote_addr` still shows the proxy IP (preserved for debugging)
Reviewed-on: https://forge.ops.eblu.me/eblume/blumeops/pulls/130
## Summary
- Start nginx before Tailscale in `start.sh` so port 8080 is bound immediately, eliminating the "app is not listening on the expected address" WARNING during `fly deploy`
- Switch `proxy_pass` to use a variable with `resolver 100.100.100.100 valid=30s` so nginx can start without resolving MagicDNS names at config load time
- DNS results cached 30s per worker — no per-request lookup overhead
## Context
The WARNING was a race condition: Fly checks for listeners right after the machine starts, but `start.sh` ran ~5-10s of Tailscale setup before starting nginx. The health check always passed later, but the warning was noisy.
## Test plan
- [ ] Merge and let the deploy-fly workflow trigger
- [ ] Check runner logs for absence of the WARNING
- [ ] Verify `docs.eblu.me` still serves correctly
- [ ] Verify `/healthz` still passes
Reviewed-on: https://forge.ops.eblu.me/eblume/blumeops/pulls/128
## Summary
- Adds a Fly.io reverse proxy (`blumeops-proxy`) that tunnels public traffic to homelab services over Tailscale
- First service exposed: `docs.eblu.me` — the Quartz static docs site
- Includes Pulumi IaC for Tailscale auth key/ACLs and Gandi DNS CNAME
- Adds mise tasks (`fly-deploy`, `fly-setup`, `fly-shutoff`) and Forgejo CI workflow
## Key details
- Fly.io Firecracker VMs support TUN devices natively — no userspace networking needed
- Tailscale auth key is `preauthorized=True` to avoid device approval hangs on container restarts
- nginx caches aggressively for the static site; health check is on the default_server block
- ACLs restrict `tag:flyio-proxy` to `tag:k8s` on port 443 only
- DNS CNAME deployed and verified: `docs.eblu.me` → `blumeops-proxy.fly.dev`
## Test plan
- [x] `curl -sf https://blumeops-proxy.fly.dev/healthz` returns `ok`
- [x] `curl -I -H "Host: docs.eblu.me" https://blumeops-proxy.fly.dev/` returns 200 with `X-Cache-Status`
- [x] `curl -I https://docs.eblu.me/` returns 200 with valid Let's Encrypt cert
- [x] `dig forge.ops.eblu.me` still resolves to 100.98.163.89 (private services unaffected)
- [x] Set `FLY_DEPLOY_TOKEN` Forgejo Actions secret for CI auto-deploy
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Reviewed-on: https://forge.ops.eblu.me/eblume/blumeops/pulls/120
