Installation and inbound network access
This page covers three things:
- what the installers and upgrades download;
- which ports cluster nodes use between each other;
- which ports users, load balancers and certificate authorities connect to.
What the running product connects out to (CAs, DNS providers, targets, Kerberos and so on) is on Outbound network access.
Installation and upgrade downloads
Standalone installer
| Service | Host | Port | When |
|---|---|---|---|
Bootstrap scripts (get.sh, update.sh) | raw.githubusercontent.com | 443 / HTTPS | The curl … | bash step of install and update |
| Release package | ghcr.io (and pkg-containers.githubusercontent.com, where GHCR serves the files from) | 443 / HTTPS | Install and update |
| MongoDB packages | repo.mongodb.org, pgp.mongodb.com | 443 / HTTPS | --mongo=local on each node that runs MongoDB. Skipped when a matching mongod is already installed |
Operating-system packages (curl, openssl, tar, gzip, gettext, ca-certificates, nginx, gnupg) | Your distribution's repositories or mirror | 80 / 443 | Only for packages not already installed. On RHEL an unregistered host with no repository fails at the nginx step |
sshpass | Your distribution's repositories (Debian/Ubuntu); EPEL mirrors on the RHEL family — EPEL must already be enabled, the installer does not enable it | 80 / 443 | Multi-node installs that use password SSH or --ssh-bootstrap only |
In a multi-node install the first node copies only the installer script to the other nodes over SSH. Each node then downloads the release package from ghcr.io itself, and each MongoDB node installs its own MongoDB packages. Open these hosts on every node.
The installer does not contact an NTP server and does not check time synchronization. A cluster still needs it — see Clock synchronization.
Docker and Helm
| Service | Host | Port | When |
|---|---|---|---|
Application images (jhonbrownn/certautopilot-backend, jhonbrownn/certautopilot-frontend) | Docker Hub (registry-1.docker.io and its auth and CDN hosts) | 443 / HTTPS | Docker and Helm installs, and every upgrade |
MongoDB image (mongo, or bitnami/mongodb in the Helm chart) | Docker Hub | 443 / HTTPS | When MongoDB runs in the same stack |
| Helm chart | ghcr.io (oci://ghcr.io/cloudnativeworks/charts/certautopilot) | 443 / HTTPS | helm install and helm upgrade |
Image registries and repositories can be overridden in the Helm values and in your compose file, so a private registry mirror works for both.
Between cluster nodes
| Flow | Port | Notes |
|---|---|---|
| First node to every other node, SSH | 22 / TCP (or --ssh-port) | Installation, and the cap-cluster key-management commands |
| Every node to each MongoDB node | 27017 / TCP | The three MongoDB nodes replicate over it, and every backend (nodes 4+ included) connects to the replica set over it |
| Load balancer to each node | The HTTPS port (--port, default 443) | The load balancer is yours; the installer sets up no virtual IP |
Nodes coordinate only through MongoDB. There is no other traffic between CertAutoPilot nodes.
When firewalld or ufw is active, the installer opens 27017 on each MongoDB node only to the other cluster nodes. Without a managed firewall it prints a warning instead, and the rule is yours to add. On cloud VMs you add it in the security groups.
On replica-set members MongoDB listens on all interfaces, so without a host firewall 27017 is reachable from anything that can route to the node. MongoDB still requires its password, but restrict the port to the cluster nodes in your network. A single-node install keeps MongoDB on 127.0.0.1 only.
Older installers opened 27017 only between the three MongoDB nodes. With firewalld or ufw active, the backends on nodes 4 and up could not reach the database. An upgrade does not re-run the firewall step. On each MongoDB node, allow 27017/tcp from nodes 4 and up yourself, for example:
sudo firewall-cmd --permanent --add-rich-rule='rule family=ipv4 source address=<node-4-ip> port port=27017 protocol=tcp accept'
sudo firewall-cmd --reload
# or, with ufw:
sudo ufw allow from <node-4-ip> to any port 27017 proto tcp
Inbound to CertAutoPilot
| Deployment | Listens on | Serves |
|---|---|---|
| Standalone | The HTTPS port (--port, default 443), TLS only | UI, API, health endpoints and the HTTP-01 challenge path (/metrics is not published here — scrape it on the host from the backend's loopback port). The backend itself listens on 127.0.0.1:18181 and is not reachable from outside. No plain-HTTP port is opened — only the --port you choose |
| Docker (reference compose) | 80 (frontend, plain HTTP); the backend's 8181 stays on the compose network unless you publish it | Put your TLS-terminating proxy or load balancer in front |
| Helm | NodePort 30051 (frontend, HTTPS) and 30050 (backend, HTTPS) by default | There is no Ingress in the chart. Add your own for a real host name |
Backend on its own (serve) | 0.0.0.0:8181 by default, HTTP or TLS 1.3 when server.tls is set | Only the api and all modes listen. Worker-only and scheduler-only processes open no port |
Unauthenticated endpoints
These paths answer without a login. On a standalone install /healthz, /readyz and the ACME path are served on the same port as the UI, and /metrics only on the backend's loopback port; the Docker and Helm frontends proxy only /api/, /healthz and the ACME challenge path, so /readyz and /metrics answer on the backend's own port there:
| Path | Purpose |
|---|---|
/healthz, /readyz | Liveness and readiness for your load balancer or orchestrator |
/metrics | Prometheus metrics (needs Authorization: Bearer <token> when telemetry.metrics.token is set) |
/.well-known/acme-challenge/<token> | HTTP-01 challenge answers |
The metrics endpoint describes your certificates and jobs. Set telemetry.metrics.token, or restrict it at the proxy or firewall to your monitoring system.
HTTP-01 challenge traffic
For certificates validated over HTTP-01, the CA connects in. It requests http://<domain>/.well-known/acme-challenge/<token> on port 80 of every domain on the certificate. That request must reach CertAutoPilot through an edge you run:
- A reverse-proxy route. The edge on :80 forwards the challenge path to CertAutoPilot, which can stay internal on any port.
- A 301 redirect. CAs follow challenge redirects only to ports 80 and 443. Redirect to CertAutoPilot's HTTPS port only when that port is 443.
CertAutoPilot's own listener speaks TLS only in a standalone install, even if --port=80 is used. It therefore cannot answer the CA's plain-HTTP request by itself. See HTTP-01 validation for setup.
Air-gapped installation
There is no offline installation bundle today.
- Standalone upgrade can run offline. Download the release package on a connected machine and run the upgrade from it; see Upgrade.
- A single-node standalone install can run from an extracted release package when the operating-system packages, nginx and MongoDB are already installed, or when you use
--mongo=external. - A multi-node standalone install needs
ghcr.ioon every node, because each node downloads the release package itself. - Docker and Helm installs work from a private registry mirror; override the image registry and repository.