SSB Parquedit

A Python package for manually editing tabular data stored as Parquet files on DaplaLab — Statistics Norway's cloud data platform. Built on top of DuckDB and the DuckLake catalog, it provides a clean Python interface for creating tables, inserting data, querying results and editing rows directly from Google Cloud Storage (GCS). Intended for single-table editing. Does not support primary- and foreign keys.

Table of Contents

• Features
• Requirements
• Installation
• Usage
  • Basic setup
  • Creating a table
  • Inserting data
  • Editing a row
  • Deleting rows
  • Querying data
  • Counting rows
  • Checking table existence
  • List all tables
  • List edits
  • Drop table
• Maintenance
  • Flush inlined data
  • Merge adjacent files
  • Export catalog
  • Import catalog
• Advanced
  • Accessing the raw DuckDB connection
  • Setting up local connection
  • Restoring a local catalog backup with GCS data
• Project structure
• Contributing
• License

Features

• Auto-configuration — reads Dapla environment variables to build connection config automatically
• DuckLake catalog integration — metadata stored in PostgreSQL, data stored in GCS
• Create tables from a pandas or polars DataFrame, a JSON Schema dict, or an existing GCS Parquet file
• Insert data from a pandas or polars DataFrame or a gs:// Parquet path — rows are automatically assigned a unique rowid within a table
• Edit data - Update value(s) in a single row by its rowid.
• Delete rows - Delete one or more rows matching a where-condition, logged individually to the changelog.
• Query tables with where-conditions, column selection, sorting, pagination, and multiple output formats (pandas, polars, pyarrow)
• Find edits Retrieve historical column-level edits for a specified table
• Count rows
• Check table existence safely
• Partition tables by one or more columns
• Export/import catalog back up the DuckLake metadata catalog to a DuckDB file on GCS and restore it later

Requirements

• Python >=3.12
• Access to a DaplaLab environment
• A PostgreSQL instance reachable at localhost for DuckLake metadata storage
• A GCS bucket following the naming convention ssb-{team-name}-data-produkt-{environment}

Python dependencies

Installation

poetry add ssb-parquedit

Usage

Basic setup

ParquEdit reads its connection configuration automatically from Dapla-environment variables.

from ssb_parquedit import ParquEdit

# Auto-configure from environment
con = ParquEdit()

Creating a table

Tables can be created from a pandas or polars DataFrame schema, a JSON Schema dict, or an existing Parquet file.

import pandas as pd

df = pd.DataFrame({"name": ["Alice", "Bob"], "age": [30, 25]})

# Option 1: Create from DataFrame (empty — schema only)
con.create_table(table_name="my_table_1",
                 source=df,
                 product_name="my-product",
                 user_defined_id=["name"])

# Option 2: Create and immediately populate with data
con.create_table(table_name="my_table_2",
                 source=df,
                 product_name="my-product",
                 user_defined_id=["name"],
                 fill=True)

# Option 3: Create from a JSON Schema
schema = {
    "properties": {
        "name": {"type": "string"},
        "age":  {"type": "integer"},
    }
}
con.create_table(table_name="my_table_3",
                 source=schema,
                 product_name="my-product",
                 user_defined_id=["name"])

# Option 4: Create from an existing GCS Parquet file (schema inferred from file)
con.create_table(table_name="my_table_4",
                 source="gs://my-bucket/path/to/file.parquet",
                 product_name="my-product",
                 user_defined_id=["id", "year"])

# Option 5: Create with partitioning and immediately populate with data
con.create_table(table_name="my_table_5",
                 source=df,
                 product_name="my-product",
                 part_columns=["age"],
                 user_defined_id=["name"],
                 fill=True)

# Option 6: Create from a polars DataFrame
import polars as pl

df_polars = pl.DataFrame({"name": ["Alice", "Bob"], "age": [30, 25]})
con.create_table(table_name="my_table_6",
                 source=df_polars,
                 product_name="my-product",
                 user_defined_id=["name"],
                 fill=True)

Notes:

• product_name is required and is stored as a comment on the table.
• table_name must be lowercase, start with a letter or underscore, contain only lowercase letters, numbers, and underscores, and be at most 20 characters.
• user_defined_id — a list of columns that together uniquely identify a row in a table, used to mimic a primary key.
• Column names must not exceed 63 bytes when UTF-8 encoded (PostgreSQL's identifier limit). Longer names — easy to hit with non-ASCII characters like æ/ø/å, which take 2 bytes each — raise a ValueError at table creation instead of silently corrupting the table later.

Inserting data in an existing table

# Insert from a pandas DataFrame
con.insert_data(table_name="my_table_1",
                 source=df)

# Insert from a polars DataFrame
con.insert_data(table_name="my_table_6",
                 source=df_polars)

# Insert from a GCS Parquet file
con.insert_data(table_name="my_table_4",
                 source="gs://my-bucket/path/to/file.parquet")

Both pandas and polars DataFrames are supported as source for create_table() and insert_data().

Each inserted row is automatically assigned a unique rowid within the table

Editing a row

edit() updates exactly one row — identified by its rowid — and logs the change reason and comment to the DuckLake snapshot.

# First look up the rowid of the row you want to edit
result = con.view(table_name="my_table_1",
                  where="name = 'Alice'")
rowid = result["rowid"].iloc[0]

# Then edit it
con.edit(
    table_name="my_table_1",
    rowid=rowid,
    changes={"name":"Alice B", "age": 33},
    change_event_reason="REVIEW",
    change_comment="Corrected name and age after data review",
)

changes is a dict of {column_name: new_value} pairs.

change_event_reason must be one of: OTHER_SOURCE, REVIEW, OWNER, MARGINAL_UNIT, DUPLICATE, OTHER

Deleting rows

delete_row() selects rows with a where clause — the same syntax as view() — and deletes all matching rows. Deletions of multiple rows are logged as one entry in the changelog get_edits()- The where-clause used and number of affected rows are logged.

# Delete a single row by its rowid
con.delete_row(
    table_name="my_table_1",
    where="rowid = 1",
    change_event_reason="REVIEW",
    change_comment="Removed duplicate entry",
)

# Delete multiple rows at once
con.delete_row(
    table_name="my_table_1",
    where="age < 18",
    change_event_reason="OTHER",
    change_comment="Removed underage entries",
)

change_event_reason must be one of: OTHER_SOURCE, REVIEW, OWNER, MARGINAL_UNIT, DUPLICATE, OTHER

Querying data

# View all rows (returns pandas DataFrame by default)
result = con.view(table_name="my_table_1")

# Filter with a WHERE clause
result = con.view(table_name="my_table_1", where="age > 25")
result = con.view(table_name="my_table_1", where="name = 'Alice' AND age >= 30")

# Limit and offset (pagination)
result = con.view(table_name="my_table_1",
                  limit=10,
                  offset=2)

# Select specific columns
result = con.view(table_name="my_table_1",
                  columns=["name", "age"])

# Sort results
result = con.view(table_name="my_table_1",
                   order_by="age DESC")

# Return as polars or pyarrow
result = con.view(table_name="my_table_1",
                   output_format="polars")

result = con.view(table_name="my_table_1",
                   output_format="pyarrow")

Counting rows

total = con.count(table_name="my_table_1",
                   where="name='Alice'")

Checking table existence

if con.exists(table_name="my_table_1"):
    print("Table found")

List all tables

con.list_tables()

List edits

get_edits() - Retrieves the full changelog for a table by reading DuckLake snapshot metadata. Each row represents a single edit, with columns for who made the change, when, the reason, which row was affected (identified by its unique key), and the old and new values for all modified columns.

Optionally filter by table name, or omit it to get the changelog for all tables.

# All edits for a specific table
df = con.get_edits(table_name="my_table")

# All edits across all tables
df = con.get_edits()

The returned DataFrame includes these changelog columns:

Drop table

drop_table() - Drops a table from the DuckLake catalog. By default, only removes the table from the catalog. DuckLake preserves data files and snapshot history, so edit history remains accessible via get_edits() after a normal drop. When cleanup=True, additionally expires snapshots and deletes GCS data files. This permanently destroys all history and cannot be undone.

# Removes the table from the catalog
con.drop_table(table_name="my_table")

# Removes the table from the catalog, expires snapshots and deletes data files
con.drop_table(table_name="my_table", cleanup=True)

Maintenance

Flush inlined data

Flush inlined data materializes inlined rows into Parquet files for a table. This is a maintenance operation for workloads with frequent small writes, helping keep storage layout efficient and query performance stable. It does not change table values, only how data is physically stored. The operation is safe to run repeatedly: Running it when nothing is pending has no effect.

# Flushes inlined data for table 'my_table'
con.flush_inlined_table(table_name="my_table")

Merge adjacent files

Merge adjacent files compacts a table’s small Parquet files into fewer, larger files. This is a maintenance step for tables that receive many small writes, improving scan efficiency and reducing file-management overhead. It preserves table data and history semantics, changing only physical file layout. The operation is safe to run repeatedly: Running it when nothing is mergeable has no effect.

# Merge adjacent files for table 'my_table'
con.merge_adjacent_files(table_name="my_table")

Export catalog

export_catalog() backs up the DuckLake metadata catalog. It flushes and merges inlined data for every table, then copies all tables from the PostgreSQL-backed catalog schema into a DuckDB file, which is uploaded to GCS. Returns the full GCS path (including filename) of the exported backup file.

# Export using the default path ({data_path}/catalog-export)
backup_path = con.export_catalog()

# Export to a custom GCS path
backup_path = con.export_catalog(export_path="gs://bucket/backups")

Import catalog

import_catalog() restores the DuckLake metadata catalog from a backup file produced by export_catalog(). For every table found in the backup, existing rows in the catalog are deleted and replaced with the backed-up rows.

# Restore the catalog from a backup file
con.import_catalog(backup_file_path="gs://bucket/backups/20250101_120000_my_schema.duckdb")

Advanced

Accessing the raw DuckDB connection

ParquEdit wraps a DuckDBConnection, which exposes the underlying duckdb.DuckDBPyConnection via its .raw property. This is useful when integrating with libraries that require a native DuckDB connection, such as Ibis.

import ibis
from ssb_parquedit import ParquEdit

con = ParquEdit()
raw = con._get_connection().raw  # duckdb.DuckDBPyConnection

ibis_conn = ibis.duckdb.connect(conn=raw)
table = ibis_conn.table("my_table_1")

Notes:

• _get_connection() is an internal method. The raw connection shares state with ParquEdit — closing either will affect both. Do not close the raw connection manually while ParquEdit is still in use.
• When using the raw connection, the user is resposible to provide the required information that ParquEdit-methods gives. E.g when creating and editing tables.

Setting up local connection

Create a ParquEdit instance backed by a persistent local SQLite catalog. Useful for local development and testing without GCS or PostgreSQL access. The catalog and data files are stored at path and persist across sessions. The directory is created if it does not already exist.

con = ParquEdit().local(path="/home/onyxia/work/")

Restoring a local catalog backup with GCS data

ParquEdit.local_with_gcs_data() attaches a local DuckDB catalog file (e.g. one produced by export_catalog()) while the actual Parquet data still lives on GCS. Useful for inspecting or restoring from a DuckLake catalog backup without needing a live PostgreSQL connection. Must be used in DaplaLab to get access to GCS-buckets.

con = ParquEdit.local_with_gcs_data(catalog_path="localcopy.duckdb")

Project structure

src/ssb_parquedit/
├── parquedit.py            # ParquEdit facade — main public API
├── connection.py           # DuckDB + DuckLake catalog connection management
├── ddl.py                  # DDL operations (CREATE TABLE, partitioning)
├── dml.py                  # DML operations (INSERT, EDIT, DELETE)
├── query.py                # Query operations (SELECT, COUNT, EXISTS)
├── maintenance.py          # Maintenance operations (flush inlined data, merge adjacent files)
├── catalogexportimport.py  # Catalog backup/restore (export/import to/from GCS)
├── functions.py            # Environment helpers (Dapla config auto-detection)
├── local.py                # Local DuckDB connection backed by SQLite (dev/testing)
├── local_backup.py         # Local DuckDB catalog file + GCS-hosted data connection
└── utils.py                # Schema utilities and SQL sanitization

Contributing

Contributions are very welcome. To learn more, see the Contributor Guide.

License

Distributed under the terms of the MIT license. SSB Parquedit is free and open source software.

Issues

If you encounter any problems, please file an issue along with a detailed description.

Credits

This project was generated from Statistics Norway's SSB PyPI Template. Maintained by Team Fellesfunksjoner at Statistics Norway (Data Enablement Department 724).