ADR-V2-001: Hybrid Kafka + Lakehouse architecture¶
Status: Accepted (proposed v2 — proposes evolution of ADR-001; does NOT supersede ADR-008) Source: prd-v2.1.md §B.15
Context¶
v1 ADR-001 chose Postgres + Redis Docker Compose for hackathon speed and explicitly deferred the production architecture decision. v2 must commit to a production scale path.
Decision¶
Adopt a hybrid architecture:
- Kafka (or MSK) for event transport.
- Databricks (Q-v2-2 resolved) for analytical lakehouse and ML lifecycle.
- Postgres + Redis retained for OLTP and hot-path serving — not deprecated. Remains the system of record for transactional state.
Consequences¶
- Production scale path is no longer hand-waved.
- Adds significant operational surface (Kafka, Databricks, Schema Registry).
- Cost shape changes from "negligible Docker Compose" to ongoing managed-service spend (~$5K-$15K/month estimated; needs validation).
- Two systems can disagree; reconciliation strategy is now mandatory.
- The pitch line changes from "real-time event-driven" to "hybrid transactional + analytical, real-time on the hot path".
Cross-references¶
- ADR-001 — proposes evolution of (does not supersede in the v1 hackathon scope).
- ADR-PROTO-001 — Part C defers Kafka to architecture diagram only; this ADR is the production target Part C points at.
- ADR-PROTO-005 — per-metric routing is the prototype slice that proves the data-binding piece of this hybrid architecture.
- prd-v2.1.md §B — full v2 production specification.