This document discusses combining real-time and batch processing by using a semi-aggregated strategy with snapshots. It presents a simple example of counting events in real-time using Storm and periodically aggregating the counts in Hadoop. This strategy allows for real-time processing with Storm while still being able to run batch jobs on historical data in Hadoop. It also discusses how this approach can be extended to other online algorithms like Bayesian bandits by representing distributions with sampling and updating counts in real-time.