Summary
Since #1951 the Parquet reader in contrib/datalake_fdw matches a table's columns to a file's by Iceberg field id (PARQUET:field_id in the Parquet schema). A file whose columns carry no field id can never be matched: every projected column reads as NULL. That is correct per spec for a file the table never wrote, and wrong for the one case the spec covers: tables created over existing Parquet data (add_files, migrated Hive tables), whose files predate the ids.
What the spec says
Iceberg resolves such files through the table property schema.name-mapping.default: a JSON mapping from field ids to the column names to look for in the file, including nested names and multiple names per id (for renamed columns). A reader applies it only to columns that have no field id.
What has to happen
- The metadata engine has to surface the property to the access method.
ProjectionSet (format/format.h) needs a way to hand the reader a name mapping alongside the field ids -- a second array of names per id, or a pointer to a parsed mapping.parquet_read.cpp's parquet_project() matches by id first and falls back to the mapping for id-less columns. Nothing changes in the decoder.- A file with neither ids nor a mapping match stays NULL, as now.
Deferred from #1951 on purpose: the framework there is settled, and the property needs the metadata engine, which is a later PR.
Summary
Since #1951 the Parquet reader in
contrib/datalake_fdwmatches a table's columns to a file's by Iceberg field id (PARQUET:field_idin the Parquet schema). A file whose columns carry no field id can never be matched: every projected column reads as NULL. That is correct per spec for a file the table never wrote, and wrong for the one case the spec covers: tables created over existing Parquet data (add_files, migrated Hive tables), whose files predate the ids.What the spec says
Iceberg resolves such files through the table property
schema.name-mapping.default: a JSON mapping from field ids to the column names to look for in the file, including nested names and multiple names per id (for renamed columns). A reader applies it only to columns that have no field id.What has to happen
ProjectionSet(format/format.h) needs a way to hand the reader a name mapping alongside the field ids -- a second array of names per id, or a pointer to a parsed mapping.parquet_read.cpp'sparquet_project()matches by id first and falls back to the mapping for id-less columns. Nothing changes in the decoder.Deferred from #1951 on purpose: the framework there is settled, and the property needs the metadata engine, which is a later PR.