Personally, if I'm going to shard manually I'll stick with postgresql. One of the primary reasons to use something like Cassandra is that it solves that for you.
[1] http://www.slideshare.net/brizzzdotcom/facebook-messages-hba...
[1] https://ccp.cloudera.com/display/CDH4DOC/Software+Configurat... [2] http://blog.cloudera.com/blog/2012/10/quorum-based-journalin...
I ran a HBase cluster with 1PB storage, it became very unwieldy at this scale, thousands of regions and lots of tricks to keep it happy. As for SPOF, the name node now has HA and it works very well.
Because of the I/O the WAL had to be turned off, this introduced problems when Region Servers occasionally died. Implementation of large regions 10GB, and fairly large HBlocks 512MB, increasing flush sizes to reduce minor compactions. Use of MSLAB to virtually eliminate GC all together, use of large heap 12GB on RS.
Worst problems we experienced was META corruption, that really , really sucked.
0.94+ has MSLAB by default, with HFileV2 (0.92+) we can support much larger regions (20G or bigger). Curious about the 512M HBlocks, did you have scan-heavy read-load?
14k TX peak per regionserver? x 2k that's 28M/s (56M with WAL). Should be doable now even with WAL (definitely with deferred flush). Well, maybe not with concurrent very heavy read load, depending on disk configuration.
Probably on top of Hadoop 0.20-append? Hadoop-2.x.x should be far better too.