Post

Most data centre setups use a Top of Rack router or switch: every server is connected up to a box at the top which handles networking for the whole rack. This gives centralised control & full rack connectivity without each server having to worry much about routing. That's not how Keisson does it.

Instead, all of our servers are connected directly to one another, forming a direct mesh. We pull this off by using our own networking code, which turns each and every server into its own router for itself and the containerised workloads running on it. This brings us several advantages...

Latency between servers is as low as physically possible. By reducing the number of machines that network traffic has to transit through to get from its source to its destination, our infrastructure can be lightning fast.

Massive resiliency against equipment failure. Our mesh model gives every machine many routes to get traffic where it needs to go. Node failure? No problem. And route fall-over and recalculation happens automatically and in real time.

In a Kubernetes environment, every machine has to act as a pseudo router and firewall for its containers anyway, as each has its own network identity. Why duplicate the networking effort twice? This allows us to wring as much usage out of our racks as possible.