Data Science Through the Looking Glass

Abstract

The recent success of machine learning (ML) has led to an explosive growth of systems and applications built by an ever-growing community of system builders and data science (DS) practitioners. This quickly shifting panorama, however, is challenging for system builders and practitioners alike to follow. In this paper, we set out to capture this panorama through a wide-angle lens, performing the largest analysis of DS projects to date, focusing on questions that can advance our understanding of the field and determine investments. Specifically, we download and analyze (a) over 8M notebooks publicly available on GITHUB and (b) over 2M enterprise ML pipelines developed within Microsoft. Our analysis includes coarse-grained statistical characterizations, finegrained analysis of libraries and pipelines, and comparative studies across datasets and time. We report a large number of measurements for our readers to interpret and draw actionable conclusions on (a) what system builders should focus on to better serve practitioners and (b) what technologies should practitioners rely on.

Pushing ML Predictions Into DBMSs

“…(i.e., no deep neural networks). Traditional methods are the state-of-the-art over structured data [32], and it is still the more widely-used type of ML [25], [26]. Nevertheless, we did test the performance of a shallow neural network in Section 5.8.2.…”

“…Manuscript received XXX (e.g., in [25] we found that pipelines can have up to hundreds of operators); 2) models are often trained once and served many times (e.g., rendering of web pages based on users' profiles, batch prediction of asset prices based on historical data), and this pattern appears quite amenable for in-DBMS execution; 3) applications where prediction serving will likely be used (e.g., websites, smart BI dashboards) are often backed by a DBMS; 4) the top used operators in practical data science over tabular data are not compute-heavy neural networks, but rather memory-intensive operations (such as one-hot encoding or tree ensemble methods [25], [26]) which should benefit from in-DBMS execution; 5) when data already resides in a database, execution of in-DBMS predictions is a natural choice, whereas a different solution will require pulling the data out of the database. This not only is a path not always practicable, for instance, if for security reasons data cannot be moved outside the database, but it also causes performance costs, while making it difficult to enforce the "Enterprise-grade" features without resorting to bespoken solutions (and likely increasing the technical debt).…”

End-to-end Optimization of Machine Learning Prediction Queries

“…Traditional ML is most widely used. According to the latest Kaggle survey [32] and an analysis of publicly available Python notebooks [69], traditional ML algorithms, such as linear/logistic regression and tree-based models (decision trees, random forests, gradient boosting) are the most popular by a large margin. 80% of the Kaggle responders use them, as opposed to 43% for neural networks.…”

“…Trained pipelines. We evaluate Raven over four popular traditional ML model types [32,69], namely, logistic regression (LR), decision tree (DT), gradient boosting (GB), and random forest (RF). Each trained pipeline includes featurizers for numerical and categorical inputs: we normalize the former using standard scaling, and encode the latter using one-hot encoding [79,80].…”

“…These queries often include additional data processing operators (e.g., filters or joins), implementing prediction-specific logic. As we discuss in §2, ML inference drives the majority of the cost (up to 90% [31]) associated with ML in the enterprise, whereas traditional ML models (i.e., nonneural networks, such as linear regression and tree-based models) are still the most widely used by a large margin [32,69]. Moreover, based on the analysis of 130 customer engagements at Microsoft, we found that batch inference is chosen over online inference in most enterprise scenarios.…”