How SQLFlow Turns Simple SQL Queries into Powerful AI Models

SQLFlow leverages the SQL language to construct machine‑learning and deep‑learning workflows, pursuing the vision of “Make AI as simple as SQL” to democratize artificial intelligence and enable anyone who understands business logic to apply AI techniques.

Core Elements

SQLFlow combines three key aspects: describing business logic with data, empowering deep data analysis through AI, and providing an easy‑to‑use interface.

Supported Models

The platform offers a rich model library, including DNN classifiers, Shap + XGBoost interpretable models, auto‑encoder based unsupervised clustering, LSTM time‑series models, and more. Detailed model information is available in the official SQLFlow model repository.

Positioning and Goals

Unlike Google BigQueryML, TeraData SQL for DL, or Microsoft’s AI extensions for SQL Server, SQLFlow aims to connect a broader range of data engines and AI frameworks without being tied to a single vendor’s ecosystem. It is an open‑source project inviting global developers to contribute and grow the community.

Supported Data and AI Engines

SQLFlow currently integrates with data engines such as MySQL, Hive, and MaxCompute, and supports AI engines including TensorFlow, XGBoost, and Scikit‑Learn, as illustrated in the following tables.

How SQLFlow Works

To illustrate the workflow, the Docker‑based Iris example is used. The Iris dataset contains 150 samples with four feature columns (sepal length, sepal width, petal length, petal width) and a label column indicating one of three flower species.

Data is stored in an iris.train table, and a DNN classifier (default two hidden layers with 10 units each, three output classes, Adagrad optimizer, learning rate 0.1, and sparse categorical cross‑entropy loss) is trained using the following SQL statement:

SELECT * FROM iris.train
TO TRAIN DNNClassifier
WITH hidden_units = [10, 10], n_classes = 3, EPOCHS = 10
COLUMN sepal_length, sepal_width, petal_length, petal_width
LABEL class
INTO sqlflow_models.my_dnn_model;

SQLFlow parses the statement: the SELECT part fetches data from the specified engine, while TRAIN and WITH define the model type, architecture, and hyper‑parameters; COLUMN and LABEL specify feature and target columns.

The platform then translates the TRAIN and WITH clauses into a Python program that orchestrates data loading, model construction, and training. The overall process is visualized below.

By simplifying the end‑to‑end pipeline, SQLFlow empowers business experts to experiment with AI models directly from familiar SQL queries, accelerating the adoption of intelligent applications across various domains.