The really cool use case in my mind is combining OLTP+OLAP in real time. Hence the interest in GPU acceleration. We have an OLTP platform that integrates a CPU version of the GAS engine used in MapGraph:
We are big fans of GraphLab. We used the version that was available when we collected the performance data on GraphLab - last summer. The asynchronous engine was seg-faulting which is why we used the synchronous engine. Lots of details about this in the SIGMOD paper:
http://bigdata.com/bigdata
http://bigdata.com/mapgraph
Eventually we'd like to see the two platforms converge.