A Final Review

We now know how to construct a static content distribution network capable of servicing a peak load of 47,000 queries per second (assuming full utilization on each cluster). More importantly, we have accomplished the following goals:

• Build a static content acceleration server to make our news site faster.

• Reduce the latency and potentially increase the throughput for most users' requests to the site.

• Build something from scratch that is easy to maintain, inexpensive to build, and inexpensive to operate in the long run.

Our final architecture has four identical cluster installs that are globally and strategically placed.

• Each cluster is completely functional and standalone. Six inexpensive commodity machines running Wackamole for IP failover provide high availability and high performance.

• The architecture runs Squid and acts as a cache-on-demand system that needs no out-of-band content synchronization.

• DNS servers using Anycast ensure that clients are delivered to the most network-appropriate cluster.

Was It Worth It?

Earlier a statement was made: "It will become clear that as the architecture scales horizontally, it is a cost-effective and appropriate technology." Let's look at this a bit more to better understand the drastic savings resulting from this architecture choice over a classic "white-paper" approach.

Figure 6.5 shows what the architecture would look like if the image cluster was fronted by the same web switch that drives the load-balancing for the dynamic content cluster. The "web switch" in that diagram is really two routers, four wiring-closet switches, and two web switches. When we presented the architecture depicted in Figure 6.7 in context, it was reasonable because that infrastructure was already there to support the dynamic application, so there was no investment to be made. However, by placing the image clusters at three other locations in the world, we would now need that same "web switch" in front of each.

In all fairness, we can't just ignore all that hardware. We will need portions of it in the new architecturenamely two routers and two switches at each location for high availability. But immediately we see that we can avoid purchasing six wiring-closet switches and six web switches. Let's put some dollars on that.

Web switches capable of actually load-balancing more than 200Mbs of traffic carry a pretty price tagoften more than $50,000 each. Wiring-closet switches often run around $5,000 each. That adds up to $625,000. Well worth it if you ask me.