Foundational machine learning model for learned database tasks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for database tasks are prohibitively expensive due to high training costs and inference times, making them unsuitable for performance-critical tasks like cardinality and cost estimation during query planning, and require extensive data for training which is not feasible.
Innovation Solution
A foundational model comprising task-independent expert sub-models, including graph neural network-based models, that are pre-trained to perform various database tasks with reduced training overhead, using a combination of base and derived experts to generate responses efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If task and instance specific machine learning models are trained for various database tasks, then model accuracy for specific tasks is improved, but training costs and deployment expenses increase prohibitively
Solution Approach 1:
The patent applies universality by creating a single foundational machine learning model that can perform multiple database tasks (cardinality estimation, query optimization, workload management, query rewrites) instead of training separate specialized models for each task. This multi-functional model reduces training costs and deployment expenses while maintaining the ability to accurately execute various database operations through a unified architecture.
2Adaptability or versatility
If generative models like large language models are used to solve different tasks with one model, then model versatility is improved, but inference costs increase to seconds which renders them unattractive for performance critical database tasks
Solution Approach 1:
The patent applies local quality by designing a foundational model with task-specific adapter components that are activated only when needed for particular database tasks. Instead of using a generic large language model that processes all tasks uniformly, the system employs specialized adapter modules (e.g., for cardinality estimation, query optimization) that are selectively engaged based on the specific task at hand, thereby reducing overall inference time while maintaining versatility.
3Adaptability or versatility
If generative models are used for database tasks, then task versatility is improved, but extensive amounts of training data are required which would not be feasible to construct for learning database tasks
Solution Approach 1:
The patent applies segmentation by dividing the training process into two distinct phases: (1) pre-training the foundational model on general-purpose data to learn broad patterns and representations, and (2) task-specific adaptation using smaller, targeted datasets for each database task. This segmented approach eliminates the need to construct extensive task-specific training corpora, as the model leverages knowledge acquired during pre-training and fine-tunes it using minimal task-specific data.
Data Source
AI summary
Aspects of the disclosure are directed to a foundational model for learned database tasks. The foundational model can be task-independent and used with low overhead to accomplish various learned database tasks. The foundational model can include a plurality of expert sub-models, each pre-trained to be utilized alone or in combination based on the database task. The foundational model can accomplish the various learned database tasks with reduced training overhead by relying on one or more particular expert sub-models based on the database task at hand.