Tiresias: Enabling Predictive Autonomous Storage and Indexing

Download PDF

Abstract

Database systems that support hybrid transaction/analytical processing (HTAP) or mixed workloads typically fully replicate data in two forms: a row layout to execute OLTP transactions and a column layout for OLAP queries. However, maintaining replicated data in HTAP systems incurs storage and compute overheads. There has been growing interest in system designs that selectively and adaptively replicate data in different storage layouts to execute HTAP workloads efficiently. On-the-fly changes in storage layout requires such systems to make cost-driven storage adaptation decisions predictively. We present Tiresias, a predictor that learns and predicts the latency cost of data accesses and their likelihood under different storage layouts. Tiresias makes these predictions by collecting observed latencies and access histories to build predictive models in an online manner. Experimental evaluation shows the benefits of predictive adaptation and their trade-offs that Tiresias brings to autonomously adapt storage layouts.