The Raw Data Store keeps one raw substrate for both transactions and analytics.
Everything collected by the ingestion layer lands here as raw data, stored in two databases at once: a SQL store optimized for transactional consistency and an MPP store optimized for analytical scale. The two are kept aligned directly through change data capture, so the same raw records are available for operational reads and for large analytical queries without maintaining two separate copies of the truth by hand.
Ingested data lands as raw records in both a SQL (transactional) database and an MPP (analytical) database, kept in sync directly by CDC.
The value is operational, not just technical.
The AI and everything above it can rely on a single raw substrate. Transactional lookups hit the SQL store while heavy aggregation hits the MPP store, but both reflect the same raw records because CDC keeps them synchronized.
Preserves ingested data as raw records before any modeling or abstraction happens.
Serves transactional consistency from the SQL store and analytical scale from the MPP store.
Uses CDC to keep the two stores aligned directly, instead of nightly copy jobs.
Gives every higher layer one consistent raw source to build on.
What this layer is responsible for
Ingested data is written as raw records rather than pre-aggregated tables.
A SQL database holds the transactional, row-oriented view for operational consistency.
An MPP database holds the analytical, column-oriented view for large-scale queries.
Change data capture propagates committed changes between the two stores to keep them in sync.
What it makes usable
This is how the layer improves AI decisions.
Without one CDC-synced raw substrate, teams end up with drifting copies: a transactional database and a separate analytics warehouse that disagree, plus batch jobs that leave analytical data hours behind operational truth.
An operational lookup for a member's current state can hit the SQL store directly.
A large cohort scan over months of events can run on the MPP store without loading the SQL store.
A committed change in one store is reflected in the other through CDC rather than a nightly batch.