The MPP store makes months of behavior queryable in seconds.
The MPP (massively parallel processing) database is the analytical half of the raw store. It holds the same raw records as the transactional SQL database, but laid out column-oriented and spread across many nodes, so large scans — cohorts, funnels, lifetime aggregates — run in parallel instead of straining the operational database. Change data capture keeps it aligned with the SQL store, so analytics reflect the same truth as operations.
The column-oriented, massively parallel store that answers large analytical queries over the same raw records the SQL store holds.
The value is operational, not just technical.
When the AI reasons over scale — every event for a segment, a metric computed across months — it hits the MPP store rather than the transactional database. Heavy analytical work runs in parallel and never competes with operational reads, so both stay fast and consistent.
Runs large analytical scans in parallel across nodes, so cohort and funnel queries return quickly.
Keeps heavy analytics off the transactional SQL store, protecting operational latency.
Holds the same raw records as SQL, kept aligned by CDC — analytics never drift from operations.
Serves the metrics layer's large aggregations from a store built for exactly that shape of query.
What this layer is responsible for
Stores raw records column-oriented for scan-heavy analytical access.
Distributes data and query execution across many nodes for massively parallel processing.
Stays synchronized with the SQL store through change data capture rather than batch copies.
Serves the abstraction layer's metrics and cohort queries without touching the operational database.
What it makes usable
This is how the layer improves AI decisions.
Without a dedicated analytical store, large queries run against the transactional database — slowing operational reads, timing out on big scans, or forcing a separate warehouse that drifts hours behind the operational truth.
A cohort scan over a year of events can run on the MPP store without slowing operational lookups.
A lifetime-value metric across the whole member base is computed in the store built for large aggregation.
A committed change in the SQL store appears in the MPP store through CDC, not a nightly load.