Lambda architecture.

Running a batch layer (accurate, periodic) and a speed layer (real-time, approximate) side by side and merging their results at query time.

the two-minute refresher

Batch layer: recomputes from all history, the source of truth for accuracy.
Speed layer: covers the gap since the last batch, discarded when the batch catches up.
Serving layer: batch view + real-time view at query time.
The big problem: two codebases (Spark and Flink) that must produce the same answer.