delta2ducklake 0.2.0

Register Delta Lake tables into a DuckLake catalog without copying the underlying Parquet data

delta2ducklake

Register a Delta Lake table into a DuckLake catalog without copying or rewriting the underlying Parquet files. Reads the Delta transaction log directly (JSON commits + checkpoints), no dependency on the deltalake Python package.

Status: beta. Published on PyPI.

Supported

• Full (copy_table) and incremental (sync_table) conversion, including table/file statistics and Hive-style partitioning
• Delta column mapping (columnMapping.mode = name / id)
• Deletion vectors, converted to DuckLake positional delete files
• A standalone refresh_stats() utility to (re)compute stats for any DuckLake table's columns (useful since Delta's dataSkippingNumIndexedCols often leaves trailing columns with none)
• Catalog backends: a local DuckDB database file (DuckLake's own native format), SQLite, PostgreSQL (including Entra ID/Azure AD token auth for Azure Database for PostgreSQL)
• Storage backends: local filesystem, Azure Blob/ADLS Gen2 (delta2ducklake[azure])

See docs/IMPLEMENTATION.md for design notes, protocol details pinned down against real fixtures, and known limitations (e.g. DuckDB's own ducklake reader can't materialize partition column values for tables whose physical layout isn't Hive-style).

Ships an Agent Skill at delta2ducklake/skills/delta2ducklake/SKILL.md (included in the installed package) so AI coding assistants pick up correct usage automatically once this package is a project dependency.

Install

uv add delta2ducklake
uv add "delta2ducklake[azure]"      # if your Delta table or Postgres catalog lives on Azure

Quick start

A DuckLake catalog needs to be bootstrapped once (creates its metadata schema), then tables are registered into it and kept up to date with sync_table() — it creates the table on first call and updates it on every call after, so it's safe to call unconditionally (e.g. from a recurring job). The simplest catalog is a local DuckDB database file — DuckLake's own native format:

from delta2ducklake.ducklake.bootstrap import bootstrap_catalog
from delta2ducklake.ducklake.catalog import DuckDBCatalogConfig
from delta2ducklake.convert import sync_table

catalog = DuckDBCatalogConfig("/abs/path/to/catalog.ducklake")
bootstrap_catalog(catalog, data_path="/abs/path/to/ducklake_data/")  # one-time; safe to re-run

# Registers "my_table" if it doesn't exist yet, or brings it up to date if it does.
sync_table("/path/to/my_delta_table", catalog, "my_table")

Use copy_table() instead if you specifically want a hard failure when the table already exists, rather than an update.

DuckDBCatalogConfig also accepts an already-open duckdb.DuckDBPyConnection instead of a path (e.g. for in-memory testing, or when you already manage a connection to the catalog file) — connect() reuses it as-is and close() leaves it open, since the caller owns its lifecycle. bootstrap_catalog() still needs a real path (it ATTACHes from a separate connection), so bootstrap first, then wrap a connection to the same file if you want to reuse one.

Or use a SQLite file:

from delta2ducklake.ducklake.catalog import SQLiteCatalogConfig

catalog = SQLiteCatalogConfig("/abs/path/to/catalog.sqlite")
bootstrap_catalog(catalog, data_path="/abs/path/to/ducklake_data/")

Use absolute paths. DuckDB's own ducklake extension pins a catalog to the exact DATA_PATH string it was first bootstrapped with and rejects any later ATTACH whose DATA_PATH doesn't match that exact string — a relative path resolves differently depending on the current working directory, so bootstrap_catalog()/ATTACH calls made from different directories (or a cron job vs. a shell) will spuriously conflict unless every caller passes the identical absolute path.

For Postgres, swap in a DSN:

from delta2ducklake.ducklake.catalog import PostgresCatalogConfig

catalog = PostgresCatalogConfig("host=localhost dbname=mydb user=me password=...")
bootstrap_catalog(catalog, data_path="/abs/path/to/ducklake_data/")  # or an Azure URL, see below

For Azure Database for PostgreSQL, authenticate with an Entra ID (Azure AD) token instead of a static password by setting entra_user (needs delta2ducklake[azure]; mirrors the pattern used by bmsuisse/pgdevkit). A fresh token is fetched on every connect()/bootstrap, since Entra tokens expire:

catalog = PostgresCatalogConfig(
    "host=myserver.postgres.database.azure.com dbname=mydb",
    entra_user="app@mydb",       # omit user/password from the DSN — these are supplied for you
    managed_identity=False,      # True to use ManagedIdentityCredential instead of DefaultAzureCredential
)

There's also experimental support for Quack, DuckDB's client-server RPC protocol, letting a remote DuckDB instance act as the catalog:

from delta2ducklake.ducklake.catalog import QuackCatalogConfig

catalog = QuackCatalogConfig("localhost:9494", token="...")  # matches quack_serve(...) on the server
bootstrap_catalog(catalog, data_path="/abs/path/to/ducklake_data/")

Known limitation: as of DuckDB v1.5.5, tables reached via Quack's remote attach only support INSERT/SELECT — UPDATE and DELETE both fail with Binder Error: Can only update/delete from base table (confirmed against a real quack_serve() instance). Since copy_table() writes its bookkeeping via DELETE+INSERT and sync_table() retires removed files via UPDATE, neither currently works against a Quack-backed catalog — only bootstrap_catalog() and raw read-only queries do. This is a limitation of the experimental protocol itself, not something delta2ducklake can work around; it should start working with no code changes here once Quack's DML support matures.

Then query the result with DuckDB directly:

INSTALL ducklake;
ATTACH 'ducklake:sqlite:/abs/path/to/catalog.sqlite' AS dl (DATA_PATH '/abs/path/to/ducklake_data/');
SELECT * FROM dl.my_table;

Backfill or recompute stats for specific columns of any DuckLake table (not just ones this package created):

from delta2ducklake.ducklake.stats_refresh import refresh_stats

refresh_stats(catalog, "my_table", columns=["a_wide_table_column_delta_never_indexed"])

CLI

The same operations are available as delta2ducklake <command> (installed as a console script):

delta2ducklake bootstrap --catalog sqlite:catalog.sqlite --data-path ./ducklake_data/
delta2ducklake copy   /path/to/my_delta_table --catalog sqlite:catalog.sqlite --table my_table
delta2ducklake sync   /path/to/my_delta_table --catalog sqlite:catalog.sqlite --table my_table
delta2ducklake refresh-stats --catalog sqlite:catalog.sqlite --table my_table --columns a,b

# Postgres: --catalog "postgres:host=localhost dbname=mydb user=me password=..."

# Azure Postgres with an Entra ID token instead of a static password:
delta2ducklake copy /path/to/my_delta_table \
  --catalog "postgres:host=myserver.postgres.database.azure.com dbname=mydb" \
  --entra-user app@mydb --table my_table

Run delta2ducklake <command> --help for the full option list (e.g. --schema, --version for time travel).

Development

uv sync --all-extras --all-groups
uv run pytest
uv run ruff check src tests
uv run ty check src/delta2ducklake

Postgres-backed tests are skipped unless DELTA2DUCKLAKE_TEST_PG_DSN is set to a libpq keyword=value connection string for a database the test user can create/drop databases in (each test provisions and tears down its own isolated database):

export DELTA2DUCKLAKE_TEST_PG_DSN="host=localhost port=5432 user=postgres password=..."
uv run pytest

CI (.github/workflows/python-test.yml) runs lint, ty check, and the full test suite (including Postgres, against a postgres: service container) on every push/PR. Publishing (.github/workflows/python-publish.yml) runs on GitHub Release (or manual dispatch) and pushes to PyPI via Trusted Publishing (OIDC, no stored token) — set that up once in the PyPI project's settings before cutting a release.