9 min read

Learning Objectives

By the end of this module, you will be able to:

  • Distinguish the edge components whose high availability the platform delivers from those whose high availability you build
  • Design an active-passive VM pair with an IP failover group, within its limits on LAN type, address type and MAC handling
  • Explain why every stable edge endpoint rests on a reserved public IPv4 address
  • Compose edge HA with zone placement inside a site and with DNS steering across sites

Unit 3.5: High Availability at the Network Edge

Introduction

FinCorp's edge now has a public ALB, a private NLB, and soon a NAT Gateway and a VPN Gateway. Some of these are highly available by design; others are VMs that FinCorp runs itself, such as a vendor-supplied partner gateway appliance. The question for every edge component is the same: if one instance fails, who moves the traffic, how fast, and to which address?

This design unit answers it in three layers: what the managed services do on their own, what IP failover does for your VMs and what it leaves to you, and how both compose with zones and DNS.

1. Two Kinds of Edge HA

1.1 Managed Services: HA Is Part of the Product

The managed edge services carry their own redundancy. The Managed Application Load Balancer, the Managed Network Load Balancer and the Managed NAT Gateway each operate in high-availability mode, with service recovery executed within seconds. For the NAT Gateway, this is handled with multiple gateways deployed across multiple availability zones in a region. Each still has a weekly maintenance window, Monday between 02:00 and 04:00 local time, with an interruption of up to 5 seconds, which your clients' retry behaviour must absorb.

The VPN Gateway makes HA a choice. Each tier can be combined with High Availability, in which the gateway instances operate in active-passive mode to minimise downtime during a failover, at a higher price (Unit 3.6). Cloud DNS is redundant at every level of its infrastructure.

For these services you design the inputs, not the failover: a reserved public address, a correct LAN placement and a maintenance-tolerant client.

1.2 Your VMs: IP Failover Groups

Some edge functions run on VMs you operate: a firewall appliance, a reverse proxy cluster, a vendor gateway. For these, IONOS CLOUD offers IP failover groups. A group assigns one shared reserved public IP address to all member NICs, so that the address can move to a standby VM when the active one fails. One member is the master, which corresponds to the master of the HA setup on your VMs and can change without updating the group in the DCD. A server can belong to several groups, and several groups can serve different services.

The division of labour is explicit. The feature only provisions the same IP to multiple NICs of different servers on the same LAN. It does not monitor the availability of the service behind the address. Monitoring, and the gratuitous ARP (GARP) announcements to the gateways, must be done by you on each VM that is a member. Creating the group in the DCD alone is not a failover: the HA software, such as keepalived or the appliance vendor's clustering, must be fully installed and configured on the VMs.

1.3 The Limits That Shape the Design

The following limits decide where IP failover can be used:

Limit Consequence for the design
Public LANs only, and only public LANs that contain no load balancer The failover pair needs its own public LAN, separate from the ALB's edge LAN
Reserved public IPv4 addresses only; DHCP-generated addresses cannot be used Reserve the shared address in the VDC's region before you create the group
IPv4 only No IPv6 failover address
Virtual MAC addresses are not supported Protocols that depend on a virtual MAC, such as CARP, do not work
A NIC can hold the primary IP of only one group per LAN Plan one primary group per NIC; use additional IPs for further groups

Within those limits, the supported mechanism is clear. Gratuitous ARP is supported, and the Virtual Router Redundancy Protocol (VRRP) is supported based on gratuitous ARP; for VRRP to work, IP failover groups must be configured. Configure the VRRP implementation to announce the failover address with gratuitous ARP from the NIC's own MAC address. You can set a custom unicast MAC address when a NIC is provisioned, but you cannot change it afterwards, so a design that moves MAC addresses between VMs is not available.

Two further details prevent surprises. Use the failover IP as the primary IP of the member NICs where possible, because the primary IP is used to calculate the gateway IP; this replaces any previously provisioned primary IP. And allow VRRP between the pair in the Network Security Group of the pair (VRRP is one of the supported rule protocols), or the members cannot see each other's advertisements.

2. Stable Endpoints and Zones

2.1 Every Stable Endpoint Is a Reserved Address

A public address that DHCP assigns to a NIC can change when the VM is deallocated or the NIC is removed. Every address that clients, partners, DNS records or firewall rules elsewhere depend on must therefore be reserved: the ALB's public listener, the NAT Gateway's public IP that partners allow list, the VPN Gateway's endpoint, and every IP failover address. Reserved addresses are region-bound, and a returned address cannot be reserved again (Unit 2.1). Treat the reserved block as a long-lived asset with an owner, and never return it during a rebuild.

2.2 Zone Placement

A failover pair protects only against failures that do not hit both members. Compute Engine offers Zone 1, Zone 2 and Auto (the default) as availability zones, and IP failover members should ideally be distributed across different availability zones. Set Zone 1 and Zone 2 explicitly for the two members, so that the separation is a recorded decision rather than an assumption about Auto. Unit 7.1 returns to the auto-zone trap for every redundant pair.

2.3 From Site HA to Cross-Site Failover

Everything in this unit works inside one VDC. An IP failover address cannot leave its LAN, and a reserved address cannot leave its region. When the whole site is lost, traffic must be steered to another VDC, and that is the job of DNS: your monitoring updates a Cloud DNS record through the API, and the record's TTL bounds how quickly clients follow (Unit 3.7). Unit 7.1 combines both layers into a recovery design with explicit RTO and RPO.

3. FinCorp's Edge HA Design

FinCorp's edge in fc-prod-fra-core uses three patterns:

  • Customer portal: the public ALB on fc-edge. Its HA is the product's; FinCorp's part is the reserved listener address and portal servers in Zone 1 and Zone 2.
  • Partner file gateway: a vendor appliance pair in Zone 1 and Zone 2 using VRRP. It sits on its own public LAN fc-partner-edge, because IP failover is not available on a LAN that contains a load balancer. The pair shares a reserved address from a separate block; fc-nsg-partner allows VRRP between the two members and the partner ports from the partners' ranges. The vendor HA configuration is part of the build runbook and is tested by powering off the master.
  • Egress and hybrid links: the NAT Gateway (HA by design) and a VPN Gateway with High Availability enabled (Unit 3.6).

Cross-site failover to the Berlin VDC fc-prod-txl-dr is designed in Unit 3.7.

Decision Summary

Edge function HA mechanism Who detects failure Hard constraints Choose when
HTTP(S) entry point Managed ALB in HA mode The platform; health checks for targets Reserved public IP; 5-second maintenance interruptions Default public entry for web and API traffic
TCP entry point or private balancing Managed NLB in HA mode The platform; TCP health checks for targets TCP only Non-HTTP or end-to-end TLS traffic
Outbound internet Managed NAT Gateway in HA mode The platform Private LANs only; reserved public IP All private-tier egress
Site-to-site tunnel VPN Gateway with High Availability The platform (active-passive) Higher price than without HA Any production hybrid link
Your own appliance or proxy pair IP failover group plus VRRP or vendor HA Your HA software on the VMs Public LAN without load balancers; reserved IPv4; no virtual MAC Functions no managed service covers
Whole site Cloud DNS record updated by your monitoring Your monitoring TTL bounds client switchover Cross-VDC failover (Units 3.7, 7.1)

Summary

At the edge, high availability is either a property of a managed service or a pattern you build. The ALB, NLB and NAT Gateway handle their own failover; the VPN Gateway does when you enable High Availability. For your own appliances, IP failover groups move a reserved public address between NICs on a public LAN without load balancers, while detection and gratuitous ARP remain your HA software's job. Reserved addresses make every endpoint stable, explicit zones make every pair independent, and DNS takes over where a site ends.

Key Points:

  • The ALB, NLB and NAT Gateway run in HA mode with recovery within seconds; the VPN Gateway offers active-passive HA per tier.
  • IP failover provisions a shared address; it does not monitor services. Your HA software announces the move with gratuitous ARP.
  • IP failover works only on public LANs without load balancers, with reserved IPv4 addresses, and without virtual MAC addresses.
  • VRRP is supported based on gratuitous ARP and requires IP failover groups; CARP is not supported.
  • Place the members of every pair explicitly in Zone 1 and Zone 2.

Important Terminology:

  • IP failover group: a reserved public IPv4 address provisioned to the NICs of several VMs on one public LAN, with one designated master.
  • Gratuitous ARP (GARP): an unsolicited ARP announcement that tells the network which MAC address now answers for an IP address.