Probe construction

A tunnel health check probe contains an ICMP (Internet Control Message Protocol) Open external link reply packet that originates from an IP address on the origin side of the tunnel and whose destination address is a public Cloudflare IP.

Cloudflare encapsulates the ICMP reply packet and sends the probe across the tunnel to the origin. When the probe reaches the origin router, the router decapsulates the ICMP reply and forwards it to the specified destination IP. The probe is successful when Cloudflare receives the reply.

Every Cloudflare data center configured to process your traffic sends tunnel health check probes. The rate at which these health check probes are sent varies based on tunnel and location. This rate can also be tuned up or down on a per tunnel basis by modifying the health_check rate of a tunnel with the API.

When a probe attempt fails for a healthy tunnel, each server detecting the failure quickly probes up to two more times to obtain an accurate result. We also do the same if a tunnel has been down and probes start returning success. Because Cloudflare global network servers send probes up to every second, you can expect your network to receive several hundred health check packets per second — each Cloudflare data center will only send one health check packet as part of a probe. This represents a relatively trivial amount of traffic.

Wireshark example of health check packets

​​ Health state and prioritization

There are three tunnel health states: healthy, degraded, and down.

Healthy tunnels are preferred to degraded tunnels, and degraded tunnels are preferred to those that are down.

Magic Transit steers traffic to tunnels based on priorities you set when you assign tunnel route priorities during onboarding. Tunnel routes with lower values have priority over those with higher values.

​​ Tunnel state determination

​​ Degraded

• When at least 0.1% or more of tunnel health checks fail in the previous five minutes (with at least two failures), Magic Transit considers the link lossy and sets the tunnel state to degraded (assuming the tunnel is not down).
• Magic Transit requires two failures so that a single lost packet does not trigger a penalty.
• Magic Transit then immediately sets the tunnel status to degraded and applies a priority penalty.

​​ Down

• When all health checks of at least three samples in the last one second fail, Magic Transit immediately transitions the tunnel from healthy or degraded to down, and applies a priority penalty to routes through that tunnel.
• A down state determination takes precedence over a degraded state determination. This means that a tunnel can only be one of the following: down, degraded, or healthy.

When Magic Transit identifies a route that is not healthy, it applies these penalties:

• Degraded: Add 500,000 to priority.
• Down: Add 1,000,000 to priority.

The values for failure penalties are intentionally extreme so that they always exceed the priority values assigned during routing configuration.

Applying a penalty instead of removing the route altogether preserves redundancy and maintains options for customers with only one tunnel. Penalties also support the case when multiple tunnels are unhealthy.

​​ Cloudflare data centers and tunnels

In the event a Cloudflare data center is down, Cloudflare’s global network does not advertise your prefixes, and your packets are routed to the next closest data center. To check the system status for Cloudflare’s global network and dashboard, refer to Cloudflare System Status Open external link .

​​ Recovery

Once a tunnel is in the down state, global network servers continue to emit probes according to the cadence described above. When a probe returns healthy, the global network server that received the healthy packet immediately sends two more probes. If the two probes return healthy, Magic Transit sets the tunnel status to degraded (as three consecutive successful probes no longer satisfy the condition for a down state).

Tunnels in a degraded state transition to healthy when the failure rate for the previous 30 probes is less than 0.1%. This transition may take up to 30 minutes.

Magic Transit’s tunnel health check system allows a tunnel to quickly transition from healthy to degraded or down, but tunnel transition occurs slowly from degraded or down to healthy. This scenario is referred to as hysteresis — which is when a system’s output depends on its history of past inputs — and dampens changes to tunnel routing caused by flapping and other intermittent network failures.

​​ Example

Consider two tunnels and their associated routing priorities. Remember that lower route values have priority.

• Tunnel 1, route priority 100
• Tunnel 2, route priority 200

When both tunnels are in a healthy state, routing priority directs traffic exclusively to Tunnel 1 because its route priority of 100 beats that of Tunnel 2. Tunnel 2 does not receive any traffic, except for tunnel health check probes. Endpoint health checks only flow over Tunnel 1 to their destination inside the origin network.

​​ Failure response

If the link between Tunnel 1 and Cloudflare becomes unusable, Cloudflare global network servers discover the failure on their next health check probe, and immediately issue two more probes (assuming the tunnel was initially healthy).

When a global network server does not receive the proper ICMP reply packets from these two additional probes, the global network server labels Tunnel 1 as down, and downgrades Tunnel 1 priority to 1,000,100. The priority then shifts to Tunnel 2, and Magic Transit immediately steers packets arriving at that global network server to Tunnel 2.

​​ Recovery response

Suppose the connectivity issue that set Tunnel 1 health to down becomes resolved. At the next health check interval, the issuing global network server receives a successful probe and immediately sends two more probes to validate tunnel health.

When all three probes return successfully, Magic Transit transitions the tunnel from down to degraded. As part of this transition, Cloudflare reduces the priority penalty for that route so that its priority becomes 500,100. Because Tunnel 2 has a priority of 200, traffic continues to flow over Tunnel 2.

Global network servers will continue probing Tunnel 1. When the health check failure rate drops below 0.1% for a five minute period, Magic Transit sets tunnel status to healthy. Tunnel 1’s routing priority is fully restored to 100, and traffic steering returns the data flow to Tunnel 1.

​​ Types of health checks

​​ Endpoint health checks

Endpoint health checks evaluate connectivity from Cloudflare distributed data centers to your origin network. Designed to provide a broad picture of Internet health, endpoint probes flow over available tunnels and do not inform tunnel selection or steering logic.

Cloudflare global network servers issue endpoint health checks outside of customer network namespaces and typically target endpoints beyond the tunnel-terminating border router.

During onboarding, you specify IP addresses to configure endpoint health checks.

​​ ​Tunnel health checks

Tunnel health checks monitor the status of the Generic Routing Encapsulation (GRE) and IPsec tunnels that route traffic from Cloudflare to your origin network. Magic Transit and Magic WAN rely on health checks to steer traffic to the best available routes.

During onboarding, you specify the tunnel endpoints the tunnel probes originating from Cloudflare’s global network will target.

Tunnel health check results are exposed via API. These results are aggregated from individual health check results done on Cloudflare servers.