Scaling and explaining machine learning powered database applications

Abstract

For decades, database systems have been the backbone of applications in a wide range of domains, e.g., finance, web services, business intelligence, social analysis, healthcare. Meanwhile, the resurgence of machine learning in particular large-scale deep learning services provided by big companies in recent years has been expeditiously reshaping these applications, enabling them to easily take advantage of model prediction services and be significantly more adaptive, intelligent and capable. However, this movement causes database applications to be less transparent, rendering database vendors incapable of offering reliable insights and explanations to their application customers. In addition, the increasing popularity of machine learning features rapidly boosts the scale of applications while the underlying legacy database systems are struggling to scale out and keep up the pace, causing tension between the application load and database system.

This thesis aims to address these two challenges. The first part of the thesis presents a new concept, referred to as on-database contextual explanation, and a suit of associated techniques to empower database applications to explain their learning powered decisions to end customers, even if they are generated by third-party cloud-based prediction services that opt not to offer explainability. The key intuition is that the data exchange between the databases and remote machine learning models already gives the applications a dynamic context that contains vital information to deduce reliable explanations, independent of the explainability of the remote models. To elaborate on and exploit this, we develop algorithms and systems to efficiently compute contextual feature explanation and counterfactual explanation by examining and monitoring databases at runtime, faithfully conforming to the “right-to-explanation” policy requested by GDPR. We also evaluate its effectiveness via extensive experiments and real-world case studies.

The second part of thesis develops a means to scale out legacy databases without migrating them to the cloud or re-deploying with added hardware. Our method is to augment legacy database systems at runtime with external caches, allowing us to offload database load to a look-aside cache on-the-fly. However, a caveat of extending database systems with lightweight caches is that the augmented system as a whole loses correctness guarantees that a typical database system offers especially for transactional workloads, requiring the developers to re-design the applications. To this end, we present transactional caching, a scheme that maintains application invariant over the augmented system. It works with any key-value in-memory caches, e.g., Redis and Memcached, and empowers them to assure that applications always see a monotonically increasing snapshot of the databases. Critical to the performance of such cache-augmented databases is the design of transactional cache replacement policies, which we prove is intractable as opposed to linear time decidable for conventional caching. Nonetheless, we develop efficient learning-augmented transactional cache policies with provable guarantees. Over real-life traces and benchmarks, they have shown effective in improving transaction throughput while guaranteeing application invariant.

These together give us a suit of concepts and techniques for database applications to benefit from powerful machine learning models, without compromising their transparency, reliability and correctness that database systems have been offering for decades.