In some OptaPlanner examples, we have seen major performance degradation when switching to Constraint Streams. (As much as 95 % of DRL performance can be lost.) This has been shown to come from the way that the groupBy() building block is implemented using Drools.
The problem gets progressively worse with bigger data sets, clearly pointing to a scalability issue.
The solution is to redesign the CS-D implementation to handle groupBy() more efficiently. More specifically, the use of more than one from() DRL statement needs to be avoided.
Some more information can be found in the following working document: