Casandra is a key-value store and does not do the things you need like joins and sub-queries. Data stores such as these do have their application sweet spots though. If you don't need critical analysis of relational data these types of programs are pretty cool.
PostGres is a fine DB, but I would guess that your figure is derived from PG being in single user mode and nothing else making demands on it while it was doing that. They do however desperately need to fix the TXID rollover problem since in hi volume data inserts the vacuum process really kills performance. PG has other issues like portability which machine and disk configuration dependencies really put a hurt on, but everything is a trade-off.
I use a smaller drive array that has much faster disks for the things I can fit on them and the performance really goes way way up, but cramming in about 11 million rows a day 24/7/365 just gobbles up storage, so I have to keep it on the big, and much slower disk array.
Keep in mind that while doing all of this the system is servicing the request of many other processes and users, and like most databases we are IO-bound in many instances as the primary storage array spends lots and lots of time running at or very close to 100% utilization. all in all we rarely dip below around 2000 TPS and are often peaking in the area of 6000 to 8000 and we do all of this on a single dell 12 core box w/48G of ram.