fix gdal

Signed-off-by: Aykut Bozkurt <aykut.bozkurt@snowflake.com>

Author: sfc-gh-abozkurt

pg_lake_engine/include/pg_lake/pgduck/gdal.h

@@ -20,4 +20,5 @@
 #include "pg_lake/copy/copy_format.h"
 #include "nodes/pg_list.h"
 
-extern PGDLLEXPORT char *GDALReadFunctionCall(char *path, CopyDataCompression compression, List *options);
+extern PGDLLEXPORT char *GDALReadFunctionCall(char *path, CopyDataCompression compression, List *options,
+											  bool forceKeepWKB);

pg_lake_engine/src/pgduck/gdal.c

@@ -32,7 +32,8 @@
  * See https://gdal.org/user/virtual_file_systems.html for compression syntax.
  */
 char *
-GDALReadFunctionCall(char *path, CopyDataCompression compression, List *options)
+GDALReadFunctionCall(char *path, CopyDataCompression compression, List *options,
+					 bool forceKeepWKB)
 {
 	DefElem    *filenameOption = GetOption(options, "filename");
 	bool		emitFilename =
@@ -92,7 +93,7 @@ GDALReadFunctionCall(char *path, CopyDataCompression compression, List *options)
 	/* whether to preserve WKB as a raw byte array */
 	DefElem    *keepWKBOption = GetOption(options, "keep_wkb");
 
-	if (keepWKBOption != NULL && defGetBoolean(keepWKBOption))
+	if (forceKeepWKB || (keepWKBOption != NULL && defGetBoolean(keepWKBOption)))
 		appendStringInfoString(&stRead, ", keep_wkb=true");
 
 	/* select a specific layer */

pg_lake_engine/src/pgduck/read_data.c

@@ -471,8 +471,14 @@ ReadDataSourceFunction(List *sourcePaths,
 				}
 
 				char	   *path = (char *) linitial(sourcePaths);
+
+				/*
+				 * Force keep_wkb so geometry stays as raw WKB BLOB. This
+				 * avoids DuckDB's strict Arrow-to-GEOMETRY conversion that
+				 * throws on types like MULTICURVE.
+				 */
 				char	   *stReadCall =
-					GDALReadFunctionCall(path, sourceCompression, formatOptions);
+					GDALReadFunctionCall(path, sourceCompression, formatOptions, true);
 
 				appendStringInfoString(&command, stReadCall);
 				break;
@@ -1482,9 +1488,10 @@ BuildColumnProjection(char *columnName,
 
 
 		/*
-		 * GDAL format can store geometry as BLOB or GEOMETRY depending on the
-		 * driver. We need to cast to BLOB and then to hex to get the WKB
-		 * representation.
+		 * GDAL data scans always use keep_wkb=true so geometry arrives as raw
+		 * WKB BLOB. The hex(TRY_CAST(... AS BLOB)) path converts it directly.
+		 * The ST_AsHexWKB fallback exists for safety but should not be
+		 * reached.
 		 */
 		if (sourceFormat == DATA_FORMAT_GDAL)
 			return psprintf("COALESCE(hex(TRY_CAST(%s AS BLOB)), ST_AsHexWKB(TRY_CAST(%s AS GEOMETRY)))%s",

pg_lake_spatial/tests/pytests/test_gdal.py

@@ -200,14 +200,8 @@ def test_gdal_zip_gml(
         user_conn,
     )
 
-    # Querying is currently not supported because WKB is converted by DuckDB
-    # Error messages reflect different DuckDB spatial versions
-    error = run_query("SELECT shape FROM test_gdal.fdw", user_conn, raise_error=False)
-    assert (
-        "Geometry type 10 not supported" in error
-        or "'MULTICURVE' is not supported" in error
-        or "Unsupported geometry type in WKB" in error
-    )
+    result = run_query("SELECT shape FROM test_gdal.fdw", user_conn)
+    assert len(result) == 5
 
     user_conn.rollback()
 

pg_lake_table/src/describe/describe.c

@@ -375,8 +375,12 @@ DescribeColumnsQueryForURL(char *url,
 	}
 	else if (format == DATA_FORMAT_GDAL)
 	{
+		/*
+		 * DESCRIBE needs the native GEOMETRY type so foreign table columns
+		 * are created as PostGIS geometry, not bytea.
+		 */
 		appendStringInfo(&command, "FROM %s",
-						 GDALReadFunctionCall(url, compression, copyOptions));
+						 GDALReadFunctionCall(url, compression, copyOptions, false));
 	}
 	else
 	{