Protect access to the cluster, store application secrets, and keep recoverable backups.

The controller serves its administrator API on two listeners:

• A local socket, /var/lib/orch-controller/admin.sock, that only root and the controller account can open. orch on the controller host uses it when no target is selected, and orch target add NAME ssh://root@HOST reaches it over SSH. Whoever can log in to the controller as root can administer the cluster, which they could already do by reading its state.
• The HTTPS port, which requires an mTLS administrator certificate. orch init leaves one in /etc/orch/admin on the controller for break-glass use; orch cluster install --admin-dir ... writes it where you choose.

Socket targets store no credentials on your workstation; access follows your SSH keys. For a certificate context, transfer the whole directory through a protected channel, restrict access to it, remove unnecessary copies, and use named targets instead of copying certificate paths into scripts:

orch target import production ./orch-production-admin/context.json --use
orch target test

Anyone with this context can administer the cluster.

GitHub Actions workflows need none of these. orch github trust lets a repository's runs exchange their GitHub OIDC token for a 15-minute deployer certificate that can deploy, roll back, and inspect one application and nothing else. The trust is bound to GitHub's repository and owner IDs on the first deploy, so a repository that later takes over the name cannot use it. See deploy from GitHub.

Agents and scripts elsewhere use the same limited credential through an access token: orch token create --app APP [--read-only] prints a value for ORCH_TOKEN that the controller exchanges for a 15-minute certificate limited to that application. Read-only tokens can see status, logs, and history but cannot deploy or roll back. The controller stores only a hash of each token; orch token revoke stops it within 15 minutes, and tokens expire after their --ttl (30 days by default).

orch token create --admin makes a token for tools that manage the whole cluster, such as a web panel. It is exchanged for a 15-minute administrator certificate, so whoever holds it can do everything an administrator can, including creating more tokens. Keep it in an orch secret or a password manager, and revoke it when the tool is gone.

The controller and administrator certificates are valid for one year; the cluster CA is valid for ten. Worker certificates renew automatically, but the controller and administrator certificates do not. orch status shows both expiry times, and orch cluster doctor warns 30 days before the controller certificate expires.

Renew both on the controller host. This works before or after the certificates expire, as long as the cluster CA is still valid. The CA does not change, so workers and existing backups are unaffected:

sudo orch cluster certificates renew \
  --data-dir /var/lib/orch-controller \
  --admin-dir /root/orch-production-admin-2027
sudo systemctl restart orch-controller

The controller loads its new certificate only when it restarts. --admin-dir must be a new directory. It receives a fresh administrator context that you handle like the one from installation. Replace the credentials of an existing target and keep its transport settings:

orch target import production ./orch-production-admin-2027/context.json --replace

SSH targets reach the controller's local socket and need no change after a renewal. Once the controller restarts, it refuses every administrator certificate issued before the renewal, so renewing is also how you lock out a lost laptop or a leaked context file. It does not lock out SSH targets: the socket needs no certificate, so remove the laptop's SSH key from the controller host as well. Import the new context everywhere you still use the old one. Access tokens keep working; revoke a leaked one with orch token revoke. After renewing, refresh the pki copy in your off-host backups. If a restore finds an expired controller certificate, it reports that and still restores the data. Run the renewal against the restored data directory before starting the controller.

Create one join token per worker:

orch join-token worker-1

A token admits one node with that name, once, until it expires (one hour by default). It also pins the cluster CA fingerprint, so orch join only enrolls with your controller. A used or expired token is worthless, but an unused one admits a node: send it through a channel you trust. While orch join runs, the token is visible in the worker's process list; pass it on stdin with orch join - if other users share that host. Revoke a token you no longer need with orch node join revoke.

Join files from orch node join create carry the same credential. Transfer them privately, keep them mode 0600 and delete copies after the worker has joined.

Send values through stdin:

printf '%s' "$DATABASE_URL" | orch secret set hello database-url \
  --from-stdin --target production

Application files reference secret names and can remain in source control. Secret values are encrypted by the controller and delivered only to assigned workers. Files are projected below /run/secrets; environment values are provided only to the workload that references them.

Rotate a secret by setting the same name again. The controller reconciles right away and replaces the services and cron jobs that reference it; no deploy is needed. Remove unused values with orch secret delete. See Secrets for details.

The controller writes a complete backup once a day into /var/lib/orch-controller/backups, including the secret-encryption key and the pki directory, and keeps the newest 7; see backups. Copy that directory off the host on a schedule, or set a backup destination so the controller copies every backup to S3 itself, encrypted to a recovery key that only you keep. orch init --recovery-dir DIR and orch cluster install can also write a day-zero recovery directory. For an extra copy at any time:

sudo orch cluster backup \
  --data-dir /var/lib/orch-controller \
  --output /secure/orch-backups/production.db \
  --secret-key-output /secure/orch-backups/production.key

Copy each completed backup off the controller. Protect it like administrator credentials because it includes:

• the controller database backup;
• the separately written secret-encryption key.

Also preserve the controller's pki directory with an access-controlled system backup. The database backup does not contain private keys, administrator contexts, worker identities, workload logs, or local-volume contents.

A backup stored only on the controller does not protect against host loss.

Use a fresh supported host and an off-host database backup, secret key, and PKI directory. Inspect the backup before restoring it:

orch cluster backup inspect \
  --backup /secure/orch-backups/production.db \
  --pki-dir /secure/orch-backups/pki \
  --secret-key /secure/orch-backups/production.key \
  --json

sudo orch cluster restore \
  --backup /secure/orch-backups/production.db \
  --pki-dir /secure/orch-backups/pki \
  --secret-key /secure/orch-backups/production.key \
  --data-dir /var/lib/orch-recovered

sudo orch controller serve \
  --data-dir /var/lib/orch-recovered \
  --listen 0.0.0.0:9443

Restore the same canonical controller HTTPS address. Existing workers keep their assigned workloads running during the outage and reconnect when the controller returns.

After recovery:

orch target test production
orch nodes --target production
orch status --target production

Confirm node identity, application state, secrets, ingress, and a new backup before resuming changes.

Controller, worker, and operator traffic uses mutually authenticated TLS. Installation creates separate credentials for each role.

• The controller is a single administrative and scheduling authority.
• Workers run the orch binary the controller offers after a controller upgrade, checked against the SHA-256 the controller names. Whoever controls the controller host can therefore run code as root on every worker; protect it like the workers themselves.
• Workers keep existing workloads running during a controller outage.
• New scheduling and administrative changes require the controller.
• orch does not provide automatic fencing or a highly available control plane.
• Local volumes remain on one worker and need an application-level backup plan.
• Workloads default to no Linux capabilities or device access. A service can request specific capabilities while keeping seccomp, no-new-privileges and blocked host devices. Capability grants require administrator credentials on deploy and rollback; choose the smallest set because some grants can weaken node isolation. A service can also set privileged. A privileged replica has root on its worker, so only administrator credentials may deploy one; its image tag, and autoUpdate on it, is as trusted as the administrator who deployed it.
• Every workload, privileged or not, can run at most 16384 processes and threads, so a fork bomb cannot exhaust the PID space of its worker.
• Workloads cannot reach the host of their own worker, or of any other worker through the mesh, except for cluster DNS, HTTP ingress on ports 80 and 443, and the ports of raw TCP routes. SSH, the controller API, and any other service on the host are refused, even when the provider firewall only blocks them from the internet. Cloud metadata endpoints (169.254.0.0/16), which hand out instance credentials, are refused too. So are hosts on private networks (10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16, 100.64.0.0/10) outside the cluster's workload network, unless an administrator deploys the application with allowPrivateNetwork = true. Workloads of different applications can still reach each other. See networking.
• Deployer credentials (access tokens and GitHub trust) may route any domain unless they were created with --domain, which limits them to those domains and their subdomains and refuses the public listener. The limit covers HTTP routes; raw TCP routes and node ports still need administrator review.
• Pkl manifests may import modules over HTTPS. Treat a remote module like any other dependency and pin it to a version you trust.
• Host firewalls, operating-system updates, DNS, and off-host backup retention remain operator responsibilities.

See limits and expectations before using orch for a workload with strong availability or storage requirements.