Hidden Technical Debt in Machine Learning Systems
A foundational paper authored by Sculley et al. and published in 2015 that systematically identifies engineering debt in machine learning systems. The paper argues that model code constitutes only a small fraction of an ML system, with surrounding data pipelines, feature engineering, and monitoring infrastructure representing the true sources of complexity.
Core Facts
In the paper "Hidden Technical Debt in Machine Learning Systems," Google pointed out that in real ML systems, the actual ML code constitutes only a small fraction of the entire system, surrounded by large amounts of data collection, feature extraction, configuration management, serving infrastructure, and monitoring code.
Google's 2015 paper "Hidden Technical Debt in Machine Learning Systems" revealed that the actual machine learning code in ML systems accounts for only a very small proportion.
According to a Google research report, the actual model training code in a production-level ML system often accounts for only 5% of the overall engineering effort, with the remaining 95% being supporting engineering infrastructure.
Timeline (last 90 days)
Google's paper "Hidden Technical Debt in Machine Learning Systems" points out that machine learning code accounts for only a small portion of real-world ML systems.
According to estimates in Google's research paper 'Hidden Technical Debt in Machine Learning Systems,' glue code and configuration code may account for 40%–60% of total development time in early-stage projects.
In the paper "Hidden Technical Debt in Machine Learning Systems," Google pointed out that in real ML systems, the actual ML code constitutes only a small fraction of the entire system, surrounded by large amounts of data collection, feature extraction, configuration management, serving infrastructure, and monitoring code.
All Facts (5)
In the paper "Hidden Technical Debt in Machine Learning Systems," Google pointed out that in real ML systems, the actual ML code constitutes only a small fraction of the entire system, surrounded by large amounts of data collection, feature extraction, configuration management, serving infrastructure, and monitoring code.
80%VerifiedGoogle's 2015 paper "Hidden Technical Debt in Machine Learning Systems" revealed that the actual machine learning code in ML systems accounts for only a very small proportion.
80%VerifiedAccording to a Google research report, the actual model training code in a production-level ML system often accounts for only 5% of the overall engineering effort, with the remaining 95% being supporting engineering infrastructure.
70%UnverifiedGoogle's paper "Hidden Technical Debt in Machine Learning Systems" points out that machine learning code accounts for only a small portion of real-world ML systems.
50%UnverifiedAccording to estimates in Google's research paper 'Hidden Technical Debt in Machine Learning Systems,' glue code and configuration code may account for 40%–60% of total development time in early-stage projects.
50%