Self–Optimizing AI data systems: a systematic literature review of CQLE-Aware query optimization, Edge–cloud offloading, Cloud/HPC orchestration, Data–Centric MLOps, and RAG

Abstract

Self-optimising data systems for AI are reviewed across five strands: learned query optimisation and approximate analytics, edge and IoT offloading, cloud and HPC orchestration, data-centric governance and MLOps, and retrieval-augmented generation. A search of six sources over 2020 to 2026 identified 1,261 records, of which 97 were retained against a four-item documentation checklist; 89 are extracted as studies and 8 are background references. Each extracted study was checked against all four cost, quality, latency, and energy (CQLE) axes and recorded where a value was reported. Evidence across learned approximate query processing and concurrency-aware plan selection (DBEst++, Bao, Lemo), content-adaptive video operators (FiGO), and updatable learned indexes shows accuracy and throughput gains when planning is tied to tail-latency and cost constraints. Edge work on split inference, quantisation control, and CPU and NPU co-scheduling, with cloud and HPC results on cache-affinity placement, serverless pre-warming and orchestration, multi-cloud spot hedging, and SmartNIC-offloaded collectives, reports predictable speed-ups and cost reductions under contention. Data-centric workflows and provenance-aware MLOps improve utility per dollar. The reporting record is the principal finding: latency is recorded for 27% of extracted studies, quality for 21%, cost for 17%, and energy for 7%, and none reports all four together. A two-timescale controller governing fast per-query actions and slower reconfiguration is outlined as a design agenda in Section 5, with a target envelope whose provenance is stated per value. The review consolidates the metrics and evaluation protocols needed to certify the four axes under concurrency, drift, and multi-tenant load.