Data Product SDK

Turn an ODCS data contract plus a set of pure transforms into a running, quality-gated, multi-target data product.

Install

Requires Python 3.12+.

pip install dataproduct-sdk                 # DuckDB target (default)
pip install "dataproduct-sdk[databricks]"   # + the Databricks/Spark cloud target
pip install "dataproduct-sdk[api]"          # + the `api` port shape (`dataproduct serve` is not implemented for the DuckDB folder layout yet)

While the newest release is a release candidate, ask for it explicitly: pip install --pre dataproduct-sdk, or with uv, uv tool install dataproduct-sdk --prerelease=allow (which puts dataproduct on your PATH globally).

Start here

Build your first data product on your laptop, on DuckDB — no account, no credentials, no cost.

uv tool install dataproduct-sdk --prerelease=allow   # puts `dataproduct` on your PATH
dataproduct doctor                                   # is this machine ready? it names the fix if not
dataproduct init signups                             # scaffold a product
cd signups && uv sync
dataproduct run sample                               # publish sample rows through the quality gate

Then follow the path. The guides ship inside the package — dataproduct learn <topic> renders one in your terminal — and each ends where the next begins:

Going deeper, when you want it: quality-gate (shaping the gate), troubleshooting (the failures people hit), and from-scratch (how it works underneath, finished in examples/duckdb-bikeshare). Another target: databricks (the same code on a free cloud workspace).

--prerelease=allow is needed only while the newest release is a release candidate. To try the SDK without installing it: uvx --prerelease=allow --from dataproduct-sdk dataproduct learn. Inside a product, task ... or uv run dataproduct ... runs the SDK version that product pins; plain dataproduct ... runs the one on your PATH.

The commands you will actually use

Every command runs from your product's root. On DuckDB a command works on one environment — which data each input reads, and where the run publishes, both declared in the contract. It is dev unless you say otherwise: the scaffold writes DATAPRODUCT_ENVIRONMENT=dev to the product's .env, and --environment overrides it. DuckDB reads no config file; --config is for Databricks only.

Every day

When something looks wrong

Authoring

Serving and deploying

dataproduct --help lists everything; dataproduct <command> --help details one. explain and run --dry-run take --format json for machine-readable output. python -m dataproduct.checks is an equivalent entry point for check when the console script is not on PATH.

If you scaffolded with dataproduct init, every command above also has a task wrapper — task --list shows them, plus task tour and task demo:break.

What a product looks like

# product.py
from dataproduct import Product
import transforms as T

product = Product(name="orders", domain="sales")


@product.table("orders")
def orders(src):
    return T.build_orders(src.read("orders_raw"))


@product.view("public_orders", reads=["orders"])
def public_orders(orders):
    return T.build_public_orders(orders)


product.enable_observability()

You write three things: transformation logic (pure ibis expression-to-expression functions), declared constraints (an ODCS contract), and topology (which output ports exist and what feeds them). The SDK supplies everything else.

src.read("orders_raw") names an input. On DuckDB the contract declares it in its openlineage custom property — one OpenLineage dataset per environment, optionally bound to the upstream contract it must conform to — so a product's inputs and outputs live in one reviewed file.

What the SDK does for you

Extending it

Targets are resolved through the dataproduct.adapters entry-point group, so a new backend is a separate distribution — no fork and no change to this package. The core depends only on the Protocols in dataproduct.ports.

License and support

AGPL-3.0-only — copyright (c) 2026 Leandro Kellermann de Oliveira. See LICENSE for the additional permission covering scaffolded/generated files. For a proprietary license, OEM bundling, or support: lkellermann@leandroasaservice.com