Mitigating Tail Catastrophe in Steered Database Query Optimization with Risk-Averse Contextual Bandits

Contextual bandits with average-case statistical guarantees are inadequate in risk-averse situations because they might trade off degraded worst-case behavior for better average performance. Designing a risk-averse contextual bandit is challenging because exploration is necessary but risk-aversion is sensitive to the entire distribution of rewards; nonetheless we exhibit the first risk-averse contextual bandit algorithm with an online regret guarantee. We apply the technique to a self-tuning software scenario in a production exa-scale data processing system, where worst-case outcomes should be avoided.