Every release is tested end to end on real machines with test/e2e. These are the environments Armada has been run in.
| Environment | Status |
|---|---|
| Cloudflare DNS-01 with Let's Encrypt staging | Tested on depa.id (amd64, Ubuntu 24.04): certificates for the root domain and a wildcard, domain verification, challenge records removed afterwards, the token kept out of logs and events, and requests through the Cloudflare proxy with trusted proxies (the client's address kept, forged headers from elsewhere dropped) |
| Local lab: Linux VMs (arm64), Pebble ACME, 1–2 nodes | Tested end to end: deploys, rolling update, secret files, internal ingress and DNS, crash recovery, containerd restart with containers kept running, reboot recovery, rolling upgrade with a state migration, rebalancing after a drained node resumed under HTTPS load without errors, replicas across nodes |
| Oracle Cloud (arm64, Oracle Linux 9, SELinux enforcing), one server, Let's Encrypt staging | Tested end to end: install.sh with systemd, doctor, wildcard certificate over Cloudflare DNS-01, rolling update under load without errors, crash recovery, containerd restart, reboot recovery, SELinux-confined apps and HAProxy, HAProxy restarted after a crash |
| depa.id (amd64, Ubuntu 24.04), three control plane servers and one node, Let's Encrypt staging | Tested end to end with test/e2e: server and node joins, only HAProxy public, certificates on every node, secrets, rolling update, rollback, and agent restart under load without errors, leader stop and SIGKILL under load, node loss and rebalancing under load, backup, all machines rebooted at once |
| Mixed architectures: three depa.id servers (amd64, Ubuntu 24.04) and an Oracle Cloud node (arm64, Oracle Linux 9, SELinux enforcing), Let's Encrypt staging | Tested end to end with test/e2e: every scenario above except the reboot, plus placement under load and images built for one architecture running only on nodes of that architecture |
depa.id (amd64, Ubuntu 24.04, kernel 6.8), two machines on a private network, --network wireguard |
Tested: the kernel WireGuard device on both nodes, handshakes in armadad doctor, requests between replicas on different nodes, only encrypted traffic on the private network, the tunnel staying up while an agent is stopped, the same key after an agent restart, and recovery after a node reboot |
depa.id (amd64, Ubuntu 24.04), machines with only public addresses, --network mesh |
Tested: init on one machine with the endpoint found on its own; a node joining with join-token --from, and two servers joining as control plane servers, each getting a mesh address; the join port closed without a token, open while one is, closed again once it is used, and the API port never public; doctor's kernel WireGuard and handshakes; replicas on different nodes reaching each other through the mesh; the leader stopped and another taking over; one machine rebooted, then all three at once, with the cluster and its app back on their own |
| depa.id (amd64, Ubuntu 24.04), two machines on a private network, switching modes | Tested: a plain cluster switched to wireguard (30 s) and back to plain (26 s) with armada cluster update --network, while a container on one node called an app on the other five times a second through the internal frontend: no request failed in either switch; armada cluster status showed each step and the peers a node did not reach yet; afterwards armada-net was the new tunnel on both nodes, armada-next gone, agent.json and doctor on the new mode; a switch given up while preparing left both nodes on plain |
| depa.id (amd64, Ubuntu 24.04), sites: two machines on a private network, one from another site with a private network of its own, and a control plane server from a third site with no private network | Tested: a plain cluster on the private network; a node joining from site far over the leader's public address with its own private IP, the join port open only while the token was and closed again once it was used; a control plane server joining from site cloud with a mesh address and serving the API; apps pinned to each site with armada.site answering a load from the first site through the internal frontend (the first request of the load's container failed, at its start and again after the reboot: a container's first call before its node knew it, fixed since; no other request failed); doctor's handshakes on armada-wg, its check that no other site's address lies in a local network, and its warning once one did; all four machines rebooted at once, with the cluster, the second server, and the apps back on their own |
| Containers of Ubuntu 24.04, Debian 12, Rocky Linux 9, AlmaLinux 9, Oracle Linux 9, Fedora 41, openSUSE Leap 15.6, and Alpine 3.20 | Tested: init and server join add the user who ran them through sudo to the armada group with the distribution's own tool (usermod, or BusyBox addgroup on Alpine), also when sudo's PATH lacks the sbin directories, and only once |
Reports from other providers are welcome.