LVS and Varnish

LVS and Varnish infrastructure

Diagram of "text" traffic flow through Wikimedia front edge LVS/nginx/Varnish infrastructure.

Notes

esams is representative of all cache-only sites; ulsfo is similar.

"upload" and "mobile" traffic operate similarly to "text" shown here.

"bits" does not; it lacks the bottom :3128 layer of varnish. The :80 layer jumps directly down to the appservers in the primary datacenter, and the cross-datacenter traffic is also between the :80 instances in this case.

When LVS balances traffic to the :80 varnishes directly for HTTP, it choses randomly. For :443 to the nginxes, it uses a hash of the client IP to help with SSL session persistence.

Within the caching layer (cp*/amssq* machines), the jump from nginx to varnish:80 is direct on the local host only.

However, the jump from varnish:80 (frontend) to varnish:3128 (backend) is different: for that jump, we hash on the URL (etc) when balancing to the backends to divide the cache space among all machines, and thus the request typically moves from one machine to another within the cache layer here.