feat(atlas): Move Atlas management traffic to private WireGuard - #311
Merged
Merged
Conversation
Atlas is a wg0 peer on each host with a tenant-0 address. The VM hook passes packets to that address to Linux, which routes them through wg0. Other tenants stay blocked by the tenant check that runs first. configure accepts --controller and stores the address in a map. Without the flag, the stored address stays.
metald.toml names the Atlas mesh address as wg_mesh.controller_address. Metal passes it to atlas-wg-mesh configure on every host setup.
Atlas gets one WireGuard identity for each region, with the last tenant-0 VM address. Atlas Settings stores the key, so a site backup keeps it. atlas-wireguard creates it once and writes atlas0.conf with one peer for each host and its tenant-0 VMs. A scheduler job and the new wireguard-link setup step write the file again. install-atlas-wireguard.sh adds a root timer that applies the file every 10 seconds and an input filter on atlas0. configure-wireguard.sh adds the Atlas peer to the host wg0, and metald.toml names the controller address. Host SSH uses the wg0 address after the WireGuard setup.
metald binds its control API to the host wg0 address, and Atlas calls it there. The use_public_ip_for_metald setting and the provider listen address are gone. A new host input firewall admits SSH and port 9000 only from the Atlas address on wg0, ports 9001 and 9002 only from hosts, and WireGuard only from the private network. Metal drops guest packets to the private network CIDR. Provider SSH calls use the server ssh_host.
A VM SSH connection now runs through its current host. The proxy command SSHes to the host wg0 address and runs nc to the guest inside the Metal namespace, so a guest needs no public address for Atlas SSH. Atlas reads the host for each connection, so a migrated VM stays reachable. SSH Task, the Cargo, IPv6 router and proxy setup, and Cargo admin commands use the proxy. install-metald.sh installs nc.
Cargo and the IPv6 router no longer take a public IPv4 allocation, and the router setup no longer waits for one. Only the HTTP proxy keeps a public IPv4 address. Its guest firewall admits TCP 80, 443 and ICMP from any address and all traffic from tenant-0 mesh addresses. Proxy peers replicate over the mesh. The apply command writes each peer name with its mesh address into /etc/hosts.
The site config key atlas_internal_url names an Atlas listener on its mesh address. Hosts and tenant-0 VMs use it for Atlas files, Cargo for the Atlas API, and Cargo and the proxies for the JWKS. Without it, they use atlas_base_url, then the site URL.
Import Server on the Metal Server list, and the import-metal-server command, add a Scaleway or AWS server by its provider ID. Atlas matches the size and image to the catalog, adds a missing Scaleway private network option, and runs the normal setup. An AWS instance outside the Atlas subnet is refused. Importing the same server again continues its setup. An import can name its own storage pool device. atlas-vm can import its own host after setup.
The Firecracker CI kernel has no TUN device, so wireguard-go cannot run in the guest. atlas-vm now downloads the Ubuntu 24.04 vmlinuz, checks its SHA-256, and extracts the ELF kernel with zstd. The unit removes a stale API socket before Firecracker starts, so a restart no longer fails, and the unit is enabled for boot. The FORWARD rules name the filter table explicitly.
atlas-vm setup has a new stage: it runs atlas-wireguard and installs the root timer for atlas0.conf, so a new Atlas VM reaches hosts through wg0 with no manual step. It installs wireguard-tools and wireguard-go, because the guest kernel has no WireGuard module.
A development Atlas runs outside the provider network. atlas-vm gateway mode runs a WireGuard relay VM on one host, and atlas-dev-gateway deploys it over SSH and writes atlas-gateway.conf on the developer machine. The atlas0 sessions pass through it end to end, so hosts expose no WireGuard port publicly.
The Atlas WireGuard address and keys now have their own Wireguard Config section on the Networking tab, next to the private network settings.
tanmoysrt
changed the title
atlas-wireguard is now configure-atlas-wireguard, and atlas-dev-gateway is now deploy-dev-gateway, like the other Atlas commands. The import rollback carries a bare nosemgrep, and the host commands drop the set_user call that frappe.connect already does.
tanmoysrt
force-pushed
the
feat/private-network
branch
from
October 1, 2026 20:45
3bedbb5 to
13a4f08
Compare
The Ubuntu kernel that atlas-vm boots has no modules in the guest root file system, so nft could not load nf_tables and the atlas0 filter failed to install. Setup and gateway mode now install the modules package for the running kernel. It also gives the guest kernel WireGuard, so wireguard-go is no longer needed.
The image catalog keeps only the newest AMI of each version, so an instance launched from an older AMI could not be imported. The import now reads the instance AMI and matches its version, such as Ubuntu_24.04.
A new provider host has no root key pair, so atlas-vm create stopped and asked the operator to run ssh-keygen. atlas-vm now creates the root key, and the gateway deploy no longer needs to.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Atlas now manages Metal hosts and service VMs only through WireGuard on the provider private network. Hosts accept no management traffic on their public address, and fewer public IPv4 addresses are needed.
Why
Service VMs used a public IPv4 address only so that Atlas could SSH to them, and Metal and host SSH were open on the public address. A private path closes those ports and saves public IPv4 addresses.
What changed
use_public_ip_for_metaldsetting is removed.Validation