# Instances on your own hosts

How instances, networks, security groups and elastic IPs behave when your organization's machines run on hosts of its own behind NAT.

An operator can bind your organization to hosts of its own (an office rack, for example) instead
of the platform's cloud. The API, CLI and console stay the same; this page lists what differs.

## Launching

- Launching without a public IPv4 and **Keep OS patched** are refused with `400`: "Launching
  without a public address is not offered on this organization’s own hosts yet" and "“Keep OS
  patched” is not offered on this organization’s own hosts yet".
- Three launch options exist only there: `publicHostname` (a name in one of your DNS zones, and
  `*.name`, that reaches the machine through the site's front: HTTPS by SNI, SSH through the jump
  on port 2222; needs `dns:ChangeRecords` on that zone), `idleSuspendMinutes` (suspend after 10 to
  10 080 idle minutes; the next HTTPS request or SSH connection through the front wakes it) and
  `requiredHostTags` (capability tags the host must carry, such as `fsync:high`). Elsewhere they
  are refused with `400` "A public hostname and idle suspend are offered on own hosts only" or
  "Required host tags are offered on own hosts only".
- A zone you ask for (by name or through a subnet) must have an active host of the site, or the
  launch is refused with `409` "This organization’s own hosts have no host in …; they are in …". When no host has room, the instance's state is `waiting-for-capacity` and its page shows
  **Waiting for capacity.**; nothing is built or billed while it waits.
- An instance's address in its network (in the subnet's range when you chose one) is fixed when
  it is launched; a launch with no free address left is refused with `409` "No free address in
  …".

## Elastic IPs

An elastic IP is held on a public forwarder of the platform and every TCP and UDP port is forwarded
to the instance it is associated with:

- It is inbound only: the instance's own outbound traffic leaves through the site's connection.
- The instance sees connections coming from the forwarder's address, not the client's. A load
  balancer in front keeps the client address for HTTP.
- An organization holds at most three such addresses unless the site's operator set another
  limit; beyond it, allocating is refused with `409` "This organization holds 3 of at most 3
  public addresses on …".
- The address can only be associated with an instance that runs on that site; otherwise
  associating is refused with `409` "Instance does not run on …, where this address forwards to".
- One forwarder serves the site; if it fails, it is replaced within minutes with the same
  addresses.

## Networks and security groups

- Each of your networks is isolated on the site from every other network, yours and other
  organizations'.
- An instance's site-side connection (where its elastic IP and load balancers reach it) is on a
  network of your organization's own on that site, a `/24`. No other organization's machines are on
  it. Only the platform's forwarder and load balancer nodes join it, to deliver your traffic.
- A security group filters that site-side connection, not traffic inside your network. Because the
  forwarder hides the client address, a rule limited to one source range does not match traffic
  that arrives through an elastic IP; use `0.0.0.0/0` for ports you publish.

## Source filtering

Every network interface of an instance drops packets whose source address or MAC is not the
instance's own, so an instance cannot pose as another machine. The platform sets this on its own
and it cannot be turned off. The instance's **Networking** tab shows it under **Source filtering**,
and `GET /ec2/api/instances/{id}` returns it as `sourceFilter` (null outside your own hosts).

## Related

- [Use an elastic IP](use-an-elastic-ip.md)
- [Networks and addresses](networks-and-addresses.md)
