Skip to content

feat(raster-gdal): port indb-raster loader utilities to canonical N-D schema#818

Closed
james-willis wants to merge 6 commits into
apache:mainfrom
james-willis:jw/raster-gdal-utils-port
Closed

feat(raster-gdal): port indb-raster loader utilities to canonical N-D schema#818
james-willis wants to merge 6 commits into
apache:mainfrom
james-willis:jw/raster-gdal-utils-port

Conversation

@james-willis
Copy link
Copy Markdown
Contributor

Summary

Reintroduces append_as_indb_raster and dataset_to_indb_raster on top of the canonical N-D raster schema. These helpers were deleted on #749 (jw/nd-raster-type) because the version on main (added in #811) targets the legacy BandMetadata / StorageType / band_data_writer API that the N-D port removes. This PR ports the helpers and their full test suite to the new builder + reader API so the deletion in #749 isn't permanent.

  • Loader rewritten against start_raster_2d / start_band_2d and band_data_writer().append_value — bands are emitted as canonical ["y", "x"] 2-D bands with the identity view.
  • Test suite (single-band GeoTIFF, UInt64 / Int64 / UInt16 nodata round-trips, multi-band GeoTIFF, per-band MEM nodata, multi-raster append) ported to the new RasterRef / BandRef accessors and is_indb().

Stacking

⚠️ Blocked on #749 — this branch sits on top of jw/nd-raster-type. The first 4 commits in the diff are #749's; only the final commit (feat(raster-gdal): port indb-raster loader utilities to canonical N-D schema) is unique to this PR. Review will be cleanest after #749 merges and this PR rebases onto main.

Test plan

  • cargo build -p sedona-raster-gdal
  • cargo test -p sedona-raster-gdal (51 passed, including 7 new utils::tests::*)
  • cargo clippy -p sedona-raster-gdal --all-targets -- -D warnings
  • cargo fmt --all --check

Replaces apache#787's 2D-only band schema with the canonical N-D schema:
spatial_dims/spatial_shape at the raster level; bands carry dim_names,
source_shape, nullable view, outdb_uri, outdb_format, plus the
non-nullable data buffer. Removes nodata_value, storage_type,
outdb_url, and outdb_band_id - every one is encodable in the new
schema:

- storage_type ↔ outdb_uri.is_null() (null = InDb, set = OutDbRef).
- outdb_url ↔ outdb_uri (no rename, same string).
- outdb_band_id ↔ encoded inside outdb_uri (#band=N or GDAL native
  subdataset URI), parsed only inside the GDAL format driver.
- nodata_value ↔ typed nodata: Binary (a null row means "no nodata").

Top-level adds spatial_dims: List<Utf8View> and spatial_shape:
List<Int64>; nullable view is List<Struct<source_axis, start, step,
steps: Int64>> where a null row encodes the canonical identity view.

Note: intermediate commits in this PR are not expected to build; only
the PR tip is CI-green. The trait, reader/builder, RS_* migration,
and GDAL loader port land in subsequent commits.
RasterRef and BandRef accessors over the canonical N-D schema:
spatial_dims/spatial_shape, transform, crs, num_bands, band(i), and
band-level dim_names, source_shape, shape (visible, derived from view),
view, data_type, nodata, outdb_uri, outdb_format, nd_buffer,
contiguous_data returning Cow<[u8]>.

validate_view enforces all view rules including i64-overflow on
start + (steps-1)*step. NdBuffer exposes raw buffer + shape + byte
strides + offset for zero-copy access (numpy / Arrow C Data Interface
boundary); VIEW → byte strides happens inside nd_buffer().

Adds BandRef::is_2d() default method as the gate GDAL-backed paths
use to refuse N-D input cleanly: true iff dim_names == ["y","x"]
over the identity view.
… reader/builder + RS_* migration

View-aware Arrow reader (RasterStructArray, BandRefImpl) with corruption-
surgery (negative steps, bad source_axis, length mismatch) that
round-trips an ArrowError. Builder exposes start_raster / start_band
for full N-D plus start_raster_2d / start_band_2d for legacy 2D, with
identity-view default written as a null view row. finish_raster
validates each band's visible shape against the raster's spatial_shape
along the spatial dims.

All 33 RS_* functions migrated mechanically; outputs on 2D inputs are
byte-identical to apache#787. RS_BandPath keeps its existing inline
fragment-stripping (format-agnostic display, untouched by the GDAL
parser). Test helpers in sedona-testing rewritten on the N-D builder
API.
Reads outdb_uri + parse_outdb_source instead of apache#787's storage_type /
outdb_url / outdb_band_id triplet. Each GDAL-backed SQL function gates
on BandRef::is_2d() at entry and returns an Execution error on N-D
input. VSI normalization, the dataset cache, and RasterIO bodies are
byte-for-byte unchanged from apache#787 - only the schema-read sites move.

In-db reads use BandRef::contiguous_data() and require Cow::Borrowed
so MEM datasets can point at the StructArray's backing buffer without
copying; for is_2d identity views this always holds.

Tests rebuilt to use RasterBuilder directly. Adds an N-D rejection
test for raster_ref_to_gdal_mem and the VRT path, plus an end-to-end
`raster_ref_to_gdal_mem` previously returned a `Result<Dataset>` and
guarded against `BandRef::contiguous_data()` returning `Cow::Owned`
with a runtime tripwire ("Internal: contiguous_data must be borrowed
for is_2d bands; got owned"). The check was correct — handing GDAL a
pointer into a `Vec<u8>` that drops at the end of the iteration would
dangle — but it ties an internal invariant ("`is_2d` ⇒ Borrowed") to
incidental properties of today's reader. Any future copy path in the
reader (compression, BinaryView block-boundary stitching, alignment
fix-up, sliced/broadcast/transposed views from apache#813 / apache#750) would
detonate the tripwire on perfectly valid 2-D rasters.

Change: return `Result<(Dataset, Vec<Vec<u8>>)>`. On `Cow::Borrowed`
the GDAL band still points directly at the StructArray buffer
(zero-copy). On `Cow::Owned` we move the `Vec<u8>` out of the Cow
without copying — the reader's existing materialization is the only
allocation — and stash it in the returned vector. The caller (the
provider in `gdal_dataset_provider.rs`) parks it in a new
`RasterDataset::_owned_band_bytes` field that lives as long as the
MEM dataset that holds the pointers.

`raster_ref_to_gdal_empty` discards the always-empty vector.
… schema

Reintroduces `append_as_indb_raster` and `dataset_to_indb_raster`
(deleted on the N-D-schema branch because they used the legacy
`BandMetadata`/`StorageType`/`band_data_writer` API). Reads are
written against the canonical `["y", "x"]` 2-D schema using
`start_raster_2d` / `start_band_2d` and `band_data_writer().append_value`.
The full test suite (single-band GeoTIFF, uint64 / int64 / uint16
nodata round-trips, multi-band GeoTIFF, per-band MEM nodata, multi-raster
append) is ported with assertions rewritten against the new
`RasterRef` / `BandRef` accessors and `is_indb()`.
@james-willis james-willis force-pushed the jw/raster-gdal-utils-port branch from 21235c3 to fed97ea Compare May 6, 2026 19:58
@james-willis
Copy link
Copy Markdown
Contributor Author

Folded into #749. The loader () has been restored verbatim and the GDAL backend now uses the compatibility shim to stay close to main's API surface, so a separate port is no longer needed.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant