Machine Learning Model Deployment with Text-Based Versioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current version control solutions for machine learning applications are difficult to scale, require large storage, and necessitate manual input, failing to efficiently manage both models and associated data.
Innovation Solution
A machine learning deployment platform that stores machine learned models as text versions and outputs as append-only data to an analytics database, enabling efficient collaboration and scalability by utilizing code versioning and immutable data sources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If current version control solutions are used for machine learning applications, then model and data management is possible, but the system is difficult to scale and requires large storage
Solution Approach 1:
The patent creates text-based version representations of machine learning models and data instead of storing complete binary model files and raw datasets. This text copying approach allows multiple versions to be stored with minimal space, as each version is represented by a compact text description rather than the full model artifacts, directly resolving the contradiction between scalability and storage requirements
Solution Approach 2:
The system extracts only the essential version identification information into text form, separating the version metadata from the actual model and data storage. This extraction allows the version control system to track and manage multiple versions without requiring proportional storage for each version, enabling scalability while minimizing storage consumption
2Extent of automation
If current version control solutions are used for machine learning applications, then model versioning is possible, but manual input from developers is required
Solution Approach 1:
The system automatically generates text versions of machine learning models and data by processing the model artifacts and training data through automated conversion routines. This self-service approach eliminates the need for developers to manually create version descriptions, as the system performs the versioning operation autonomously, thereby increasing automation extent while simplifying ease of operation
Solution Approach 2:
The system performs preliminary automated conversion of model and data into text version formats during the model training and saving process. By preparing the version representations in advance through automated processes rather than requiring post-hoc manual intervention, the system achieves both high automation and operational simplicity
3Reliability
If complete model files and datasets are stored for each version, then full version control is achieved, but storage requirements increase significantly
Solution Approach 1:
Instead of storing complete model files and datasets for each version, the system creates compact text-based copies that capture the essential version information. This text copying method maintains version control reliability by preserving version identification and metadata while reducing storage requirements from gigabytes per version to kilobytes per version
Solution Approach 2:
The system applies different storage strategies to different components: text-based lightweight storage for version metadata and identification, and reference-based storage for the actual model artifacts. This local quality differentiation ensures version control completeness for tracking and management while minimizing overall storage space by only fully storing what is necessary
Data Source
AI summary
A method includes receiving a trained machine learned model, where the trained machine learned model is trained using a dataset stored at an analytics database. The method further includes storing an object version of the machine learned model at an object storage, storing a text version of the machine learned model at a machine learning deployment platform, and deploying, by the machine learning deployment platform, the machine learned model. The method further includes storing output of the machine learned model at the analytics database by appending the output data to the analytics database.