Skip to contents

Merge a freshly built asset registry with the one already published

Source: R/publish.R
registry_merge.Rd

native_asset is a registry of what is published, but publish_native.qmd assembles it from whatever a given run happened to build, then writes it with overwrite = TRUE. A run that skips a build chunk (NATIVE_SKIP_PMTILES, or an opt-in class such as PUBLISH_MERGED_COG left off) therefore rewrote the table from its own products alone, silently deleting every row of the classes it did not touch. That is exactly how all 2,234 vector-range PMTiles rows disappeared from the published v8 registry while the tiles themselves sat untouched on S3 and the file host: the species app stopped offering any range input, and the content-hash checkpoint recorded the loss as a clean run (n_pmtiles_native: 0).

Usage

registry_merge(
  new,
  prior,
  allow_shrink = FALSE,
  key = c("ds_key", "asset_type", "representation")
)

Arguments

new

registry rows built by this run (data frame; may be zero-row)

prior

registry rows currently published (data frame, NULL, or zero-row)

allow_shrink

write new verbatim — no carry-forward, no shrink check

key

columns defining a class

Value

new plus any carried-forward rows, with the carried class labels in the "carried" attribute and the per-class counts in "classes"

Details

A class is one (ds_key, asset_type, representation) triple — the unit the app actually reasons about, since that is what becomes an input layer. This function:

  • carries forward every prior class the new registry has no rows for at all (the skipped-chunk case: those assets are still published, the run just did not enumerate them); and

  • errors when a class the new registry does cover came back smaller than it was — a partial build masquerading as a complete one, which carrying forward would paper over with a mix of fresh and stale rows.

Deliberate removals stay possible but must be explicit: allow_shrink = TRUE (wired to NATIVE_REGISTRY_REBUILD=1) writes exactly what the run produced.