Low Rank Approximation for Learned Query Optimization

We present LimeQO, a learned steering query optimizer based on linear methods, such as matrix completion, for repetitive workloads. LimeQO can forgo expensive neural networks by taking advantage of the low-rank structure of query workloads. Using offline execution, LimeQO can accelerate workloads by up to 2x with zero regressions in just a few hours, while using 100-1000x fewer computational resources than deep learning techniques.