pipistrelloandClaude Opus 5
537c2afc46
Users report desktop Outlook cannot be configured from outside the office, while OWA and mobile keep working — the connection-bound Windows auth signature. NPM (OpenResty) terminates HTTP, so NTLM/Negotiate legs land on different pooled upstream connections and the handshake never completes. SNI pass-through puts TLS termination back on Exchange and preserves the auth end-to-end. Also removes a second, quantified defect: ~82,000 ActiveSync Cmd=Ping 504s in the retained log window (~3,400/day, 100% of all 504s). NPM has no custom snippet on this host, so the default proxy_read_timeout 60s truncates every push heartbeat. Under pass-through IIS owns the long-poll and there is no proxy timeout to tune. Layout follows the proven 19-nutrilent stack, with two deliberate differences: - ACME uses TLS-ALPN-01, not HTTP-01. httpChallenge installs Traefik's own handler on :80 for /.well-known/acme-challenge/, which would compete with forwarding that same path to Exchange for ITS renewal. TLS-ALPN runs entirely on :443 and leaves :80 free; the Exchange SNIs are pass-through and never request a Traefik cert, so there is no overlap. - The http->https redirect is a normal priority:1 catch-all router rather than an entrypoint-level redirection, which would be installed at near-max priority and swallow the ACME path before exchange-acme (priority:1000) could match it. That exchange-acme router is the point: it forwards the challenge path for the 24 Exchange names to 192.168.0.6:80, so win-acme on SERVERMAIL renews itself without the manual port-80 NAT repoint every past renewal has needed. It requires the edge :80 NAT to point at the docker host (192.168.0.9) — it is currently on Exchange and must be moved back as part of the cutover. Called out in the README. Pass-through covers 24 SNIs — exactly the SAN list of Exchange's current cert. That is the 21 names NPM published plus three that had no NPM vhost at all and currently fail the TLS handshake outright (mail/autodiscover.alisailina.com, autodiscover.neviol.ru); they are on the cert already, so including them fixes them at no cost. Not migrated: nginx.shcnw.ru (NPM's own admin UI) and nginxlogs.shcnw.ru (goaccess parses NPM's logs, and Exchange — 99.1% of traffic — moves to L4 where there is no HTTP access log; retire it). Staged on the host and pre-flighted with a real Traefik container on reverseproxy-nw with NO published ports, so it could not touch NPM's :443: config parses, every router/service/middleware resolves, ChallengeTLSALPN provider starts. Config checksums verified against this repo. The only errors are ACME issuance failures, expected while NPM still holds :443. Two things the pre-flight caught: - traefik.shcnw.ru is NXDOMAIN, so the dashboard router is shipped commented out. ACME validates from the internet even for a LAN-only router. Create the A record and uncomment (the file hot-reloads); the htpasswd is already staged on the host. - All other L7 names and all 24 pass-through names resolve to 217.15.22.194. Deliberately unproven and flagged in the README: that TLS-ALPN issuance works while a TCP pass-through router shares the entrypoint. The SNI sets are disjoint so it should, but nothing here demonstrates it — hence the recommendation to make the first start use the Let's Encrypt staging CA. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
212 lines
11 KiB
Markdown
212 lines
11 KiB
Markdown
# traefik — edge reverse proxy for 43-meditsina
|
||
|
||
Single **Traefik v3** instance replacing **nginx-proxy-manager** on the docker host
|
||
(`192.168.0.9`). It does three jobs on the one public `:443` / `:80`:
|
||
|
||
| Job | Hosts | How |
|
||
|-----|-------|-----|
|
||
| **SNI pass-through (L4)** | 24 `mail.*` / `autodiscover.*` names → Exchange `192.168.0.6:443` | TLS terminates **on Exchange** → native NTLM/Negotiate survives → **desktop Outlook can be set up from outside the office** |
|
||
| **TLS-terminate + Let's Encrypt (L7)** | `crm.*`, `1c.exprinta.ru`, `start`, `portainer`, `cockpit` | Traefik terminates, auto-issues/renews, and preserves the real client IP so `ipAllowList` still works |
|
||
| **ACME forward (L7 on `:80`)** | the same 24 Exchange names, `/.well-known/acme-challenge/` only | forwarded to `192.168.0.6:80` so **win-acme on SERVERMAIL renews itself without the manual NAT repoint** |
|
||
|
||
**Why this migration:** an HTTP-terminating proxy (NPM = OpenResty) breaks connection-bound Windows
|
||
auth — NTLM/Negotiate must complete on one TCP connection, and NPM spreads the legs across pooled
|
||
upstream connections. Symptom reported by users: *Outlook cannot be configured from outside the
|
||
office*, while OWA and mobile (stateless auth) keep working. Pass-through fixes it end-to-end.
|
||
|
||
It also removes a second, quantified defect: **~82,000 ActiveSync `Cmd=Ping` 504s** in the retained
|
||
log window (≈3,400/day, 100% of all 504s). NPM's default `proxy_read_timeout 60s` truncates every
|
||
ActiveSync push heartbeat; under pass-through IIS owns the long-poll and there is no proxy timeout.
|
||
|
||
Full rationale and evidence:
|
||
`nextcloud/diagnostics/2026-08-26-43-meditsina-docker-npm-to-traefik-migration-plan.md`.
|
||
Modelled on the proven 19-nutrilent and 05-osk deployments.
|
||
|
||
---
|
||
|
||
## Files
|
||
|
||
| Repo file | Purpose | Goes on host at |
|
||
|-----------|---------|-----------------|
|
||
| `docker-compose.yaml` | the stack — **the only file Portainer pulls** | pulled by Portainer GitOps |
|
||
| `traefik.yml` | **static** config (entrypoints, ACME, providers) | `/mnt/containers/traefik/container-data/traefik.yml` |
|
||
| `dynamic/services.yml` | **dynamic** config (routers/services/middlewares) — hot-reloads | `/mnt/containers/traefik/container-data/dynamic/services.yml` |
|
||
| `users/dashboard.htpasswd.example` | template only | real file created on host, gitignored |
|
||
|
||
> **Portainer pulls only `docker-compose.yaml`.** The config files must already be on the host at
|
||
> **absolute** paths — Portainer runs compose relative to its *own* container, so relative `./mounts`
|
||
> resolve to empty dirs and Traefik crash-loops. Keep host copies in sync with this repo on every
|
||
> change.
|
||
|
||
---
|
||
|
||
## Two design decisions that differ from the 19-nutrilent stack
|
||
|
||
**1. ACME uses TLS-ALPN-01, not HTTP-01.** `httpChallenge` makes Traefik install its own handler on
|
||
`:80` for `/.well-known/acme-challenge/`, which would compete with forwarding that same path to
|
||
Exchange for *its* renewal. TLS-ALPN runs entirely on `:443` and leaves `:80` free. The Exchange SNIs
|
||
are pass-through and never request a Traefik cert, so there is no overlap.
|
||
|
||
**2. The http→https redirect is a normal router, not an entrypoint redirection.** An
|
||
entrypoint-level redirect is installed as an internal router at near-maximum priority and would
|
||
swallow the ACME path before `exchange-acme` could match. It is therefore a `priority: 1` catch-all
|
||
router in `dynamic/services.yml`, with `exchange-acme` at `priority: 1000`.
|
||
|
||
---
|
||
|
||
## Host preparation — **already done 2026-08-26**
|
||
|
||
Recorded here for rebuilds. Staged and pre-flighted on `docker` (43); config checksums verified
|
||
against this repo.
|
||
|
||
```bash
|
||
ssh -p 43009 -i ~/.ssh/id_ed25519 -o IdentitiesOnly=yes root@localhost
|
||
|
||
H=/mnt/containers/traefik/container-data
|
||
mkdir -p $H/dynamic $H/users $H/acme $H/logs
|
||
# copy traefik.yml and dynamic/services.yml from a checkout — write to a temp name and
|
||
# `mv` into place (atomic): a plain scp truncates-then-writes and the file watcher can
|
||
# read mid-write, throwing a transient `yaml: mapping key already defined`.
|
||
touch $H/acme/acme.json && chmod 600 $H/acme/acme.json # 600 or Traefik refuses to start
|
||
# dashboard basic-auth — htpasswd is NOT installed on this host, use openssl:
|
||
PW="$(openssl rand -base64 15)"
|
||
printf 'admin:%s\n' "$(openssl passwd -apr1 "$PW")" > $H/users/dashboard.htpasswd
|
||
chmod 600 $H/users/dashboard.htpasswd; echo "$PW" # store in the password manager
|
||
```
|
||
|
||
`reverseproxy-nw` already exists. SELinux is **Enforcing** — every mount carries `:Z`.
|
||
|
||
### Pre-flight (safe: no published ports, cannot touch NPM's `:443`)
|
||
|
||
```bash
|
||
timeout 15 docker run --rm --name traefik-precheck --network reverseproxy-nw \
|
||
-v $H/traefik.yml:/etc/traefik/traefik.yml:ro,Z -v $H/dynamic:/etc/traefik/dynamic:ro,Z \
|
||
-v $H/users:/etc/traefik/users:ro,Z -v $H/acme:/etc/traefik/acme:Z \
|
||
-v $H/logs:/var/log/traefik:Z traefik:v3.3
|
||
: > $H/acme/acme.json && chmod 600 $H/acme/acme.json # precheck may register an ACME account
|
||
```
|
||
|
||
Result 2026-08-26: config parses, all routers/services/middlewares resolve, `ChallengeTLSALPN`
|
||
provider starts. The only errors are ACME issuance failures — **expected**, because NPM still holds
|
||
`:443` so the ALPN challenge cannot be answered yet.
|
||
|
||
---
|
||
|
||
## ⚠️ Before cutover — two prerequisites
|
||
|
||
**1. Port 80 must be NAT'd to the docker host (`192.168.0.9`).** It is currently pointed at
|
||
**Exchange** (`192.168.0.6`) — repointed by hand on 2026-08-26 so win-acme could renew. Traefik's
|
||
`exchange-acme` router is what replaces that manual step permanently, but it can only do so if `:80`
|
||
arrives at Traefik. **Move it back to `192.168.0.9` as part of the cutover.**
|
||
|
||
**2. `traefik.shcnw.ru` has no DNS record.** The dashboard router is shipped **commented out** for
|
||
that reason — ACME validates from the internet even for a LAN-only router. Create
|
||
`traefik.shcnw.ru A 217.15.22.194`, then uncomment the router (hot-reloads, no restart). The
|
||
htpasswd is already staged.
|
||
|
||
---
|
||
|
||
## Deploy (Portainer GitOps)
|
||
|
||
1. Push this `traefik/` folder.
|
||
2. Portainer → **Stacks → Add stack → Repository** → this repo, branch `main`, compose path
|
||
`traefik/docker-compose.yaml`, env `TZ=Europe/Moscow`. **Do not start it while NPM holds `:443`** —
|
||
Traefik will crash-loop on the port conflict.
|
||
|
||
### Recommended: first start against Let's Encrypt STAGING
|
||
|
||
The one thing the pre-flight could **not** prove is that TLS-ALPN issuance works while a TCP
|
||
pass-through router shares the same entrypoint (the SNI sets are disjoint, so it should — but it is
|
||
unproven here). Validate it without burning production rate limits:
|
||
|
||
```bash
|
||
# in traefik.yml, uncomment:
|
||
# caServer: https://acme-staging-v02.api.letsencrypt.org/directory
|
||
# cut over, confirm all 5 L7 routers get a (staging) cert and Exchange pass-through works, then:
|
||
# re-comment caServer, truncate acme.json, restart the container -> production certs issue
|
||
```
|
||
|
||
## Cutover
|
||
|
||
```bash
|
||
docker stop nginx-proxy-manager nginx-proxy-manager-db goaccess # keep DEFINED for rollback
|
||
# deploy/start the traefik stack in Portainer
|
||
```
|
||
|
||
Then repoint the edge `:80` NAT to `192.168.0.9`.
|
||
|
||
**Rollback is instant** — NPM's data and config are untouched:
|
||
|
||
```bash
|
||
docker stop traefik && docker start nginx-proxy-manager nginx-proxy-manager-db
|
||
```
|
||
|
||
Keep NPM stopped-but-defined for ~1 week, then decommission.
|
||
|
||
---
|
||
|
||
## Verification
|
||
|
||
```bash
|
||
# 1. Pass-through is real L4 — the SNI serves EXCHANGE's own cert, not a Traefik cert
|
||
echo | openssl s_client -connect 127.0.0.1:443 -servername mail.pda.ae 2>/dev/null \
|
||
| openssl x509 -noout -issuer -subject -ext subjectAltName
|
||
# -> issuer Let's Encrypt, subject CN=mail.alisailina.com, 24 SANs (NOT a Traefik-issued cert)
|
||
|
||
# 2. An L7 host gets a Traefik-issued cert
|
||
echo | openssl s_client -connect 127.0.0.1:443 -servername 1c.exprinta.ru 2>/dev/null \
|
||
| openssl x509 -noout -issuer -subject
|
||
|
||
# 3. ipAllowList works BOTH ways (proves Traefik sees the real client IP)
|
||
curl -sk -o /dev/null -w '%{http_code}\n' --resolve start.shcnw.ru:443:127.0.0.1 https://start.shcnw.ru/ # 403
|
||
curl -sk -o /dev/null -w '%{http_code}\n' --resolve start.shcnw.ru:443:192.168.0.9 https://start.shcnw.ru/ # 200/302
|
||
|
||
# 4. Self-signed backends must not 502 (proves serversTransport is wired)
|
||
curl -sk -o /dev/null -w 'crm=%{http_code}\n' --resolve crm.shcnw.ru:443:192.168.0.9 https://crm.shcnw.ru/
|
||
|
||
# 5. The ACME forward reaches Exchange rather than being redirected to https
|
||
curl -sI http://mail.shcnw.ru/.well-known/acme-challenge/probe # -> from IIS, NOT a 301
|
||
curl -sI http://mail.shcnw.ru/ # -> 301 to https (catch-all)
|
||
|
||
# 6. Container stable, no real errors
|
||
docker inspect traefik --format 'RestartCount={{.RestartCount}} Running={{.State.Running}}'
|
||
docker logs traefik --since 5m 2>&1 | grep -iE "level=err" | grep -vi "connection reset by peer"
|
||
```
|
||
|
||
**End-to-end — the actual test:** configure a **desktop Outlook profile off-network**. It must
|
||
autodiscover and authenticate. Then confirm OWA and mobile ActiveSync still work, that ActiveSync
|
||
`Ping` now holds well past 60 s, and that an external browser is refused on the LAN-only hosts.
|
||
|
||
> **Benign noise:** `Error while handling TCP connection … 192.168.0.6:443 … connection reset by peer`
|
||
> on the pass-through router is normal L4 keep-alive churn, **not** a fault — provided there are
|
||
> **zero** TLS/SNI/handshake errors and `RestartCount` stays 0.
|
||
|
||
---
|
||
|
||
## Backend map (what replaced what)
|
||
|
||
| Old NPM proxy host | Domain(s) | Backend | Path here |
|
||
|---|---|---|---|
|
||
| 1, 9, 11, 12, 13, 14 | 21 × `mail.*` / `autodiscover.*` | `192.168.0.6:443` | **TCP passthrough** `exchange` |
|
||
| — | +3 names that had **no** NPM vhost and failed the handshake outright (`mail`/`autodiscover.alisailina.com`, `autodiscover.neviol.ru`) | `192.168.0.6:443` | same router — **fixed for free**, they are on Exchange's cert |
|
||
| 2 | `crm.shcnw.ru`, `crm.inkam.navy` | `192.168.0.10:443` (Bitrix, self-signed) | `crm` (+`insecure` transport) |
|
||
| 10 | `1c.exprinta.ru` | `192.168.0.4:80` (Apache 2.4.23 Win64) | `onec` |
|
||
| 3 | `start.shcnw.ru` | `flame:5005` | `start` (LAN-only) |
|
||
| 5 | `portainer.shcnw.ru` | `192.168.0.9:9443` | `portainer` (LAN-only) |
|
||
| 6 | `cockpit.shcnw.ru` | `192.168.0.9:9090` | `cockpit` (LAN-only) |
|
||
| 4 | `nginx.shcnw.ru` (NPM admin UI) | — | **dropped** — NPM is gone |
|
||
| 8 | `nginxlogs.shcnw.ru` (goaccess) | — | **dropped** — see below |
|
||
|
||
## Post-cutover follow-ups
|
||
|
||
- **Retire goaccess.** It parses NPM's logs, which go stale once NPM stops, and Exchange — 99.1% of
|
||
traffic — moves to L4 where there is no HTTP access log. Remove the `goacess-for-nginx` stack and
|
||
drop the `nginxlogs.shcnw.ru` DNS record. Use Traefik's native Prometheus metrics instead.
|
||
- **Drop the `nginx.shcnw.ru` DNS record.**
|
||
- **NPM's admin UI was published on `0.0.0.0:81`**, bypassing its own `192.168.0.0/22` ACL. That
|
||
exposure disappears with NPM; confirm nothing at the edge still forwards `:81`.
|
||
- **Decide `rdp.inkam.navy`** — it resolves here and has a stale NPM cert but no vhost; publish or retire.
|
||
- **Exchange keeps renewing its own certificate.** With `exchange-acme` in place that becomes
|
||
automatic; verify at the next renewal (**due 2026-10-20**, cert expires 2026-11-24) that win-acme
|
||
succeeds *without* a NAT change.
|