Gdal Knowledge Patch

GDAL Knowledge Patch

Use this skill when changing GDAL applications, bindings, out-of-tree drivers, builds, command lines, pipelines, or data-access configuration. Inspect the project's manifest and linked GDAL at runtime before applying version-sensitive advice. Prefer the project's headers, generated API documentation, tests, and observed driver metadata when they disagree with assumptions.

Reference index

Upgrade checks first

Update out-of-tree driver signatures

For the 3.11 API transition:

• Include gcore/gdal_fwd.h instead of redeclaring public opaque types.
• Override protected IGetExtent()/IGetExtent3D() hooks rather than the checked public extent methods.
• Override ISetSpatialFilter() and return/check OGRErr from spatial-filter calls.
• Handle GDT_Float16 and GDT_CFloat16; request GDT_Float32 conversion when native half precision is unsuitable.
• Treat partial coordinate-transform failure as aggregate failure and inspect the per-point success or error-code arrays.

For the 3.12 API transition:

• Make dataset and layer inspection overrides const-correct and store returned layer definitions and spatial references through const pointers.
• Handle GFT_Boolean, GFT_DateTime, and GFT_WKBGeometry; check the CPLErr returned by raster-attribute-table mutation.
• Change geotransform overrides from six-double pointers to GDALGeoTransform references.
• Do not assume raw-file VRT bands have unrestricted access; honor the runtime policy and the build-time raw-band gate.

For the 3.13 API transition:

• Check OGRErr from C point-setting and point-adding functions.
• Replace MIN, MAX, and ABS from GDAL headers with CPL_MIN, CPL_MAX, and CPL_ABS; arrange your own M_PI declaration.
• Update GDALDataset::Close() overrides for progress arguments and option lists to CSLConstList where required.
• Treat metadata lists as const and update custom VSI Read()/Write() overrides to the single-count form.
• Account for RasterIO resampling operating in the output buffer type unless bOperateInBufType is false.

Audit removals, restorations, and binary compatibility

The 3.11 line removed many legacy raster and vector drivers, several writers, the OpenCL warper, and unofficial utilities. It also redirected FileGDB writes to OpenFileGDB, deprecated OGR Memory in favor of MEM, removed PDF GEO_ENCODING=OGC_BP, and bumped the shared-library major version. GSBG, GSAG, and BT were restored in later 3.11 maintenance releases; Tiger and UK .NTF returned later as future-removal candidates. Never infer availability from the initial removal list alone. Rebuild binary dependents whenever the shared library major changes.

Migrate unified command paths

The unified gdal front end covers raster, vector, multidimensional, dataset, driver, and VSI work. In 3.12, several gdal vector geom ... operations moved directly beneath gdal vector, and geom set-type became set-geom-type. Update scripts to the direct paths. CLI info commands default to text while API calls retain JSON defaults, and progress goes to standard output unless quiet.

The 3.13 command argument spelling favors --input and --output; older --src/--dst aliases remain accepted. Avoid parsing incidental standard output from tiling pipelines, and request JSON explicitly when automation depends on it.

Treat failure status as part of the API

• Check algorithm argument validation, cancellation, close-time progress, field-domain operations, raster-attribute mutations, and geometry mutations.
• ogr2ogr now fails by default when destination field creation fails; use -skip only when dropping fields is intentional.
• ogr2ogr also returns a nonzero status for VRT-processing errors.
• INIT_DEST=NO_DATA without a nodata value warns and zero-fills in 3.11.5, but fails in 3.13.0. Gate behavior when supporting both.
• GTI reads fail when a source is unreadable, missing Kerchunk and ADBC targets are errors, and contouring an all-nodata raster succeeds with empty output.

High-value current workflows

Compose algorithms and pipelines

Use the algorithm registry rather than shelling out when an in-process API is needed. Python exposes dynamically generated gdal.alg.*() calls and accepts a progress callback. Pipelines can mix raster and vector stages, nest other pipelines, branch with tee, invoke external, override saved parameters, and use _ to select a non-first dataset emitted by a preceding stage.

Materialize intermediate data when a following stage requires a concrete dataset. Named outputs infer their format, while an anonymous COG can flow directly into tile. Keep filters before gdal vector limit; current behavior applies the filtered stream.

Use raster algebra and VRT expressions

Raster bands support arithmetic, comparisons, AsType(), common math and aggregation functions, and IfThenElse(). VRT derived bands provide richer aggregates, expressions, reclassification, constants, coordinate variables, transpose and block controls. Ensure muparser is present for C++ VRT expressions; exprtk is optional and increases library size.

For calculations, choose --dialect=muparser|builtin deliberately, account for nodata, and use --flatten where required. The built-in dialect can reduce all bands of one input to one output band. Inputs without geotransforms and nested pipeline inputs are supported.

Select correct warp and resampling semantics

Check per-point transform results, output-buffer data type, exact integer nodata comparisons, and NaN nodata handling. Lanczos no longer has its old half-valid-pixel threshold, RMS overview normalization changed, and sum resampling no longer carries the earlier chunking artifacts. Use TransformBounds() semantics for -te with -te_srs.

When --target-aligned-pixels is used for a raster mosaic, also provide --resolution. Reprojection can use all CPUs by default, so set thread counts explicitly when resource limits matter. For global or TPS warps, retain the corrected longitude and SOURCE_EXTRA=5 behavior.

Handle modern raster types and COG creation

Treat GDT_UInt8 as the canonical unsigned eight-bit type; GDT_Byte is an alias. Support Float16/CFloat16 in dispatch, prediction, statistics, and nodata paths. COG supports band/tile interleave, complex data, random-write creation, and BigTIFF temporary files; multiband overview creation works through its multithreaded path.

When translating a selected mask to COG, the mask becomes a regular alpha-tagged band. Creation options for lossy JPEG XL require JXL_LOSSLESS=NO; otherwise GTiff and COG warn. Do not assume a structurally valid COG needs GDAL's ghost area for LAYOUT=COG reporting.

Work safely with cloud and curl-backed paths

Authentication changes invalidate relevant cloud caches. Keep redirect authorization scoped, especially for S3-like redirects, and use path-specific controls for query strings, headers, connection limits, cache behavior, and verbatim path handling. /vsis3/ supports IAM Identity Center, credential_process, directory buckets, and endpoint URLs with schemes.

Use permitted header_file names, expect /vsicurl/ to fall back to a bounded GET when size is absent from HEAD, and remember that multithreaded cloud sync includes empty files. For Azure/ADLS, token and SAS settings are exposed in handler option metadata.

Preserve vector schemas and geometry fidelity

Unified vector algorithms propagate field domains, relationships, and metadata. Conversion warns when curves, Z, or M cannot be preserved. Parquet supports editable layers, geometry logical types, list variants, Timestamp With Offset, Hive filtering, covering bounding boxes, partition metadata, and ignored-field name collisions.

Expect geometry-producing operations to return richer types: edge-built polygons may be multipolygons, polar reprojection closes polygon rings, and GeoJSON export has explicit curve, measure, and coordinate-order controls. PostGIS can use full-geometry intersection; select local or server-side spatial filtering deliberately.

Verification checklist

Before shipping a change:

1. Confirm the loaded GDAL version and shared-library ABI.
2. Query driver metadata for create, update, append, upsert, subdataset, reopen, close-visibility, and directory-oriented capabilities.
3. Exercise failure paths, progress interruption, and explicit Close() for buffered Arrow/Parquet or other long-running output.
4. Test representative nodata, mask, alpha, Float16, complex, and very large raster cases.
5. Test CRS axis mapping, partial failures, polar/global bounds, and geotransform consistency.
6. Run command automation with explicit output format and inspect exit status.
7. Test cloud paths with the intended credentials, redirects, cache policy, empty objects, and path normalization.
8. For bindings, test ownership, repeated closure, no-GIL execution where applicable, NumPy dtypes/strides, and option-list parsing.