Query Engine Concurrency Scaling for In-Database ML Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data warehouse systems lack integrated machine learning capabilities and efficient methods for training and testing machine learning models, particularly in handling diverse hardware configurations and resource allocation for data preprocessing and prediction tasks.
Innovation Solution
A database system with integrated machine learning capabilities that includes computing resources for training and testing models, supports hardware-agnostic model compilation, and employs workload managers for dynamic resource allocation and on-demand provisioning of secondary query engines to handle diverse hardware configurations and data preprocessing operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a data warehouse system integrates machine learning capabilities with diverse hardware configurations, then the system's adaptability and functionality are improved, but the device complexity increases
Solution Approach 1:
The system employs a universal query engine architecture that can execute machine learning workloads across diverse hardware configurations (GPUs, FPGAs, ASICs, CPUs) through a unified interface. The workload manager and compilation system enable the same query engine to adaptively utilize different hardware types, making the system multi-functional without requiring separate specialized systems for each hardware type.
Solution Approach 2:
The patent introduces a workload manager and hardware-agnostic compilation system as intermediary layers between the query engine and diverse hardware devices. This intermediary layer translates high-level machine learning queries into hardware-specific instructions, shielding the user from hardware complexity while enabling broad hardware support through a standardized interface.
2Productivity
If the system dynamically provisions secondary query engines for machine learning workloads, then productivity is improved, but the device complexity increases
Solution Approach 1:
The system implements dynamic provisioning of secondary query engines based on workload characteristics. The workload manager automatically determines when to spin up additional query engine instances for machine learning tasks versus using the primary query engine, and dynamically allocates computing resources (GPUs, FPGAs, ASICs, CPUs) to these secondary engines. This dynamic behavior enables efficient resource utilization without requiring permanent dedicated infrastructure for each workload type.
3Loss of time
If the system performs data preprocessing operations within the query engine, then the loss of time is reduced, but the device complexity increases
Solution Approach 1:
The patent merges data preprocessing operations with the machine learning model training and testing operations within the query engine. Instead of performing preprocessing in separate external systems before passing data to the machine learning workload, the system integrates preprocessing steps (such as data transformation, filtering, and feature engineering) directly into the query execution plan, allowing preprocessing and model operations to execute together in a unified manner.
Data Source
AI summary
A database system may include a machine learning model which may be used to perform various data analytics for data stored in the database system. In response to a request to invoke the machine learning model to generate a prediction from data stored in the database system, the database system may evaluate the request to determine whether to perform the request at a primary query engine or assign the request to another computing resource distinct from the primary query engine. The database system may provision the other computing resource as needed to create a secondary query engine to use the machine learning model to generate the prediction as requested.