Why? Plenty of options to scale Postgres. What alternatives do you use?
I can't speak for Exuma, but I've built several ad networks and buying tools that needed to handle over as much as a trillion daily events, and that means lots of HTTP event collectors writing events down into a lot of otherwise independent streams that you need to merge-sort together into coherent reporting and deduplicate client retries.
In principle, anything with statement-replication (such as one of the multi-master plug-ins for Postgres) could be made to work here, but error recovery tends to be poorly considered (and users expect no downtime here) and so it's almost certainly easier to just write a log file (or let cloudfront or whatever do it), and write a custom log shipper. Dedup in a window is easy, but UPDATE on a tall skinny table even with an INDEX is slow.
If you want to materialise that in Postgres back at home-base, that can probably work. There are a lot of single-writer methods to scale Postgres. I've used Postgres for this and it's fine.
The other problem is configuration: Outside of naive impression counting, your HTTP "event collector" might also need to be smart enough to choose an ad. Naively relying on replication from your configuration database to all your nodes means it's easy to get split-brain, and you're still taking an (often unacceptable) performance hit asking the "local" Postgres server your questions -- Meeting hard-realtime guarantees consistently (usually 30-50msec) is hard for a bit of PHP and Postgres to accomplish. Usually I'll build the decision logic into the configuration payload -- actually pre-compile the decision tree (in one implementation, into C that was loaded by a openresty+lua driver every config-change) so that the code doesn't waste any time asking the database over and over for the same key that probably hasn't changed.
Erlang is also a good choice for adtech people, because it has this cluster-wide messaging, hot code reloading (good for config!), a built-in binary log (disk_log) with replication capabilities, and a in-memory database (mnesia) that is good for generating reports out of.