How to Load a Parquet File into PostgreSQL

Postgres has no Parquet reader

COPY in core PostgreSQL reads text, CSV and its own binary format, so COPY orders FROM 'orders.parquet' is not an option on a standard install. There are two reliable routes that need nothing installed on the server: DuckDB writing straight into Postgres, or a CSV export loaded with \copy. Both were checked against PostgreSQL with DuckDB 2.0.

Route 1: DuckDB's postgres extension

DuckDB can attach a Postgres database and write into it like a local schema. It reads the Parquet file and sends the rows in one step:

INSTALL postgres;
LOAD postgres;

ATTACH 'dbname=shop user=etl host=127.0.0.1' AS pg (TYPE postgres);

CREATE TABLE pg.public.orders AS
SELECT * FROM 'orders.parquet';

Keep the password out of the connection string; the extension uses libpq, so PGPASSWORD or a ~/.pgpass file works. To add more files to the same table later:

INSERT INTO pg.public.orders
SELECT * FROM 'orders_2026_11.parquet';

DuckDB creates the table with mapped types: BIGINT to bigint, VARCHAR to character varying, TIMESTAMP WITH TIME ZONE to timestamptz, DECIMAL(10,2) to numeric(10,2) and INTEGER[] to integer[]. Two types need attention.

Structs fail. A STRUCT column stops the load with Composite types in Postgres must be named. Convert it to JSON in the same statement:

CREATE TABLE pg.public.orders AS
SELECT * REPLACE (to_json(address) AS address)
FROM 'orders.parquet';

The column arrives as json. Use ALTER TABLE ... TYPE jsonb afterwards if you want to index it. To get separate columns instead, see flattening nested Parquet columns.

Unsigned integers grow. UBIGINT becomes numeric(20,0) because Postgres has no unsigned type. If the values fit, CAST(col AS BIGINT) gives a smaller column.

Route 2: CSV and psql's \copy

When DuckDB cannot reach the database directly, for example through a bastion host or a locked-down network, export CSV and load it with psql. First look at the column types:

DESCRIBE SELECT * FROM 'orders.parquet';

Create the table by hand from that output:

CREATE TABLE orders (
  order_id   bigint,
  customer   text,
  created_at timestamptz,
  amount     numeric(10,2)
);

Export with the Parquet to CSV converter in the browser, or with the CLI:

duckdb -c "COPY (SELECT order_id, customer, created_at, amount
  FROM 'orders.parquet') TO 'orders.csv' (HEADER)"

Then load it from the client side:

psql -d shop -c "\copy orders FROM 'orders.csv' WITH (FORMAT csv, HEADER true)"

\copy reads the file on your machine, so it works without file access on the server, unlike plain COPY. NULLs round-trip correctly: DuckDB writes NULL as an empty field and an empty string as "", and Postgres CSV mode reads them back the same way. Timestamps with time zones include their offset, so they land on the correct instant.

List and struct columns do not fit CSV well. Drop them with EXCLUDE, or convert them with to_json() and declare the target column as jsonb.

Verify the load

Compare row counts on both sides:

-- DuckDB
SELECT count(*) FROM 'orders.parquet';
-- Postgres
SELECT count(*) FROM orders;

Then spot-check a few null-heavy columns with count(col) on both sides. If you have not inspected the file yet, the Parquet Viewer shows the schema and row count without uploading anything. For a quick query that does not need a database, the SQL Workbench may answer the question without any import.