DuckDB to Apache Parquet round-trip type fidelity for TIMESTAMPTZ and INTERVAL

Answer

The installed DuckDB release does not automatically restore the original TIMESTAMP WITH TIME ZONE or INTERVAL types when a Parquet file is read with read_parquet() without an explicit schema; TIMESTAMP WITH TIME ZONE is inferred as plain TIMESTAMP (zone offset lost) and INTERVAL is inferred as a string or numeric type, discarding interval semantics.

Confirmed facts

• read_parquet samples file schema; if no TYPE hint is supplied, TIMESTAMP WITH TIME ZONE becomes TIMESTAMP and INTERVAL becomes a generic string or decimal.
• Underlying instant values are preserved, but the column type in DuckDB changes, so SQL type statements may need adjustment.

Steps to verify and enforce fidelity

1. Write the table to Parquet using COPY ... TO 'file.parquet' (FORMAT PARQUET).
2. Open the file with read_parquet('file.parquet') and inspect the inferred type via SELECT typeof(column) FROM read_parquet(...).
3. If the types are not as expected, rerun the read with an explicit schema, e.g., CREATE TABLE t AS SELECT * FROM read_parquet('file.parquet', TYPE 'timestamp with time zone', TYPE 'interval');
4. Alternatively, use COPY with the SCHEMA option to define column types at write time, then read without schema changes.

Note: behavior may differ between DuckDB versions prior to 1.2, where INTERVAL serialization is limited. Please provide the exact DuckDB version installed to confirm whether interval handling is fully supported.