The idea with the WindowManager is to act as a middle-man between the aggregation stage (which receives documents and outputs documents) and the window functions (which accumulate values, potentially remove, and return a computed result).

Some concepts that would be worth hashing out:

• How will the stage logic interact with the WindowManager? Or vice versa?
  • Likewise, who is responsible for driving the pulling of the next document in the pipeline?
• Can we assume that all docs in a partition will be buffered?
• What sort of state is shared across the different WindowManagers and what is unique?
• When the stage detects a new partition, what is the exact sequence of events to populate the remaining documents? This piece will likely require logic in both the stage and the window manager