We tracked clicks on our own (physical) servers but for impressions, you're talking a system built to serve upwards of 30bn a month. While nothing compared to industry-leaders (Google, DoubleClick, etc), that's still around 10,000 impressions a second.'

We had no problem building a system on our inhouse servers to handle that 10k/s number: A very thin PHP instance running on port 80 that does nothing but push the impression into a Gearman queue. Easy enough, even at that scale.

The problem of course is that you're not getting 10k/s. Some seconds you may get 100k, others 1k.

So we built a simple system that automatically spawned S3 instances when we began to see high loads, distributed the latest code to them, and when we no longer needed them, we downloaded the DB they held locally and moved the data to HDFS.

What we did is pretty common, I'm sure. It's all pretty straightforward software.