In principle, 2-3TB should be possible to do without resorting to a shared-nothing architecture, but Vanilla PostgreSQL still does not have good parallel query functionality. All you will achieve by replication is the ability to apportion queries out to individual nodes. I don't believe that PostgreSQL supports federated queries out of the box and I would be very surprised to find that Tableau has direct support for sharding at the client. I'd guess that PostgreSQL won't perform very well on a data set this big.
There is work in progress on a parallel query facility for PostgreSQL, but AFAIK it's not included in 9.1. My gut instinct is that some release in the next few years will include this feature but it's not quite there yet. I don't see evidence of much effort being put into a federated query facility.
Another option: SQL Server
Unless you're married to Postgres you might find that SQL Server provides a cost-effective option to Greenplum for a 2-3TB data set. It is licenced by the socket rather than the core, so a loaded 2-4 socket Xeon or Opteron box gives good bang for buck as a platform. For a smaller user base I believe you can still licence Enterprise Edition by CAL as well.
A couple of 24/25 disk arrays on a high-end RAID controller will do sequential reads fast enough to saturate a PCIe-x8 slot (2GB/sec). A simple table scan query with SQL Server will handle data at this rate without using much CPU (obviously depending on the actual computation), so you have some headroom to add controllers and arrays if you want faster I/O.
SQL Server also comes with a passably good set of B.I. tooling, including an OLAP server. Tableau is not cheap, around £1,800/seat last time I looked. Depending on the number of users you may find that the 'good-enough' tooling that comes with SQL Server could offset the costs of the DB licence anyway. Most third party reporting tools also play nicely with SQL Server.
Disclaimer: I'm not any sort of die-hard Microsoftie but I have done a lot of B.I. work with SQL Server and Oracle over the past decade or so. SQL Server is actually quite a good B.I. platform.
Best Answer
I fail to see why you are loadbalancing between pgbouncers on the same server, I assume you have a single DB server instance, if they run on the same node you're not really balancing anything.
Than being said, this sounds like timeout issues, since pgsql connections tend to live longer than the standard haproxy timeouts I suggest your raise them: (for example in defaults section add : )
This will keep any connection that the client hasn't closed open for an hour. queries that take longer to run will be cut.
see https://github.com/gplv2/haproxy-postgresql/blob/master/haproxy-example.cfg for more inspiration from a working haproxy config.