Using SQL databases, it is easy to do statistical / aggregate functions like covariance, standard deviation, kurtosis, skewness, deviations, means and medians, summation and product etc, without taking the data out to an application server. http://www.xarg.org/2012/07/statistical-functions-in-mysql/
How are such computations done effectively (as close as possible to the store, assuming map/reduce "jobs" won't be realtime) on NoSql databases in general and dynamodb(cassandra) in particular, for large datasets.
AWS RDS (MySQL, PostgresSQL, ...) is, well, not NoSQL and Amazon Redshift (ParAccel) - a column store - has a SQL interface and may be an overkill ($6.85/hr). Redshift has limited aggregation functionality (http://docs.aws.amazon.com/redshift/latest/dg/c_Aggregate_Functions.html, http://docs.aws.amazon.com/redshift/latest/dg/c_Window_functions.html)