0004: ERDDAP CSV and coordinate-based subsets

Status: accepted. Date: 2026-09-05.

Context

CoastWatch's blended SST metadata exposes a 0.05-degree global grid and an irregular daily time axis beginning 2019-07-22 at noon UTC. A query must select coordinates inside the requested bounds without silently snapping to another day or a neighboring grid cell.

Live probes of the same tiny subset returned different NetCDF bytes because ERDDAP adds request timestamps to its history attribute. Repeated CSV requests returned identical bytes while retaining coordinates and a units row.

Decision

Download raw griddap CSV, without response transformation. Support the four advertised variables, defaulting to analysed_sst. Keep ERDDAP metadata, axis parsing, and URL grammar in a transport module; keep dataset dimensions, variables, and selection policy in the CoastWatch adapter.

Validate spatial metadata, calculate contained grid centers, and read the actual time axis before selecting inclusive timestamps. Use coordinate values in asset URLs rather than indices that could shift when historical coverage changes. Sort/deduplicate variables for stable IDs. Allow a positive spatial stride and cap requests at one million output grid rows. An empty intersection returns no assets and participates in the existing manifest empty-source policy.

Consequences

No scientific dependency is added. Consumers must handle ERDDAP's second CSV header row containing units. NetCDF output and optional readers remain future work. Time-axis discovery costs a small request on each resolution, while locked restoration uses pinned URLs directly. Historical observations can still change; checksums detect this but cannot recover old upstream bytes. Tiny live tests exercise repeated downloads and locked restoration.