Kakao Reduces Global POP Costs and Ensures High Availability with Anycast Technology

Kakao has extended Anycast technology beyond DNS to various application services, including Contents Cache. In particular, by using the Quagga open source, it achieved cost reduction and high availability at global POPs within its own network.

Principles of Anycast and Background of Adoption

Anycast is a concept where different hosts use the same IP address, routing traffic to the server on the most optimal path via the BGP routing protocol. This technology was widely adopted to ensure the availability of DNS services following the 2002 Root DNS DDoS attack and the 2003 Internet outage. Like Google Public DNS, globally distributed servers use the same IP, with a structure that automatically bypasses failures.

Kakao's Actual Use Cases

Kakao utilizes Anycast in the following areas:

• Anycast DNS: Configures geographically distributed DNS servers to improve response times, with servers in other regions automatically taking over during a failure in one area.
• Contents Cache: Used in conjunction with Anycast to complement inaccurate targeting issues with GSLB. It enables accurate location-based routing even when mobile users use Google Public DNS or designated filtering DNS.
• Virtual IP: Also used for Virtual IP purposes, allowing IP adjustments through configuration alone without changing the server's physical IP, differing from its original purpose.

Implementation Method and Advantages

Kakao simplified the configuration by using only the bgpd daemon of Quagga. It sets the Anycast IP on the server via a Loopback interface and communicates with the router using BGP. The main advantages of this approach are as follows:

• Cost Reduction: Implementable with software alone, without expensive L4/L7 load balancer hardware.
• Stability: Stopping the BGP daemon automatically removes the server from the router, allowing safe service shutdown without traffic loss.
• Flexibility: Custom monitoring scripts control the BGP daemon upon detecting application failures, automating service removal.

However, since it selects the server closest along the network path, the physical location and logical path may not align, and separate monitoring tools are required to compensate for the lack of router-level health checks.