We have also seen this. The huge memory capacities now offered in the cloud makes single node processing very capable with entire TB datasets fitting into RAM, and running quickly enough to offset the hourly cost. Spark clusters are more efficient though for longer-running or continuous background processes.