Pushing ML Predictions Into DBMSs.

Area of Science:

• Database Management Systems
• Machine Learning
• Data Science

Background:

• Existing research on in-database machine learning (ML) primarily focuses on model training.
• ML inference, crucial for deploying ML models in applications, has been largely overlooked within database management systems (DBMS).
• Integrating ML inference into DBMS offers potential benefits in efficiency, performance, and data governance.

Purpose of the Study:

• To investigate the viability of using DBMS for ML prediction serving.
• To develop and evaluate a method for translating trained ML pipelines into SQL queries for in-DBMS execution.

Main Methods:

• A novel technique was developed to translate ML pipelines, including featurizers and models (linear, tree-based), into SQL queries.
• In-database ML inference performance was benchmarked against popular ML frameworks like Sklearn and ml.net.

Main Results:

• In-database ML pipelines achieved performance comparable to external ML frameworks in several scenarios.
• Significant performance limitations were observed for text featurization and neural network models when executed within the DBMS.

Conclusions:

• DBMS can be a suitable platform for ML prediction serving for certain types of models and data.
• Further research is needed to optimize in-DBMS performance for complex ML tasks, particularly text featurization and neural networks.