# ADR-0002: One storage format on every tier — Parquet + DuckDB, Hive layout

## Status

Accepted (2026-07-08)

## Context

Data lives across [storage floors](../00-glossary.md#t0--hot) T0–T4. Primary persistent
stores are [T1](../00-glossary.md#t1--warm) (drone NVMe in flight) and
[T3](../00-glossary.md#t3--warehouse) (ground warehouse after landing);
[T0](../00-glossary.md#t0--hot) is the in-memory hot window,
[T2](../00-glossary.md#t2--shared) optional on-board derived replication,
[T4](../00-glossary.md#t4--dev) dev/sim slices. If each floor used a different
format or layout, offload would need transformation code,
schemas would drift, and a query proven on the bench would not run unchanged on
flight data.

## Decision

Use **one storage format on every floor**: columnar **Parquet** files in an
identical **Hive-partitioned** layout (`dataset=…/flight=…/drone=…/…`), queried
with **DuckDB**. Flight offload T1 → T3 is therefore a plain **mirror**
(copy/rsync of partition directories), not an ETL step.

## Consequences

- Offload is a file operation, testable with a checksum; no format converters
  to maintain or version.
- The same SQL runs against T1, T3, and simulated data — the simulator
  ([`../../simulator/`](../../simulator/)) generates the *real* layout, so tests
  exercise the production format.
- Partitioning choices are load-bearing; changing them is a schema migration.
  Detailed in [`../03-data-platform.md`](../03-data-platform.md).
- DuckDB's single-file, no-server model fits an air-gapped drone with no room
  for a database daemon.