Towards a hands-free query optimizer through deep learning

Towards a hands-free query optimizer through deep learning Marcus & Papaemmanouil, CIDR’19

Where the SageDB paper stopped— at the exploration of learned models to assist in query optimisation— today’s paper choice picks up, looking exclusively at the potential to apply learning (in this case deep reinforcement learning) to build a better optimiser.

Why reinforcement learning?

Query optimisers are traditionally composed of carefully tuned and complex heuristics based on years of experience. Feedback from the actual execution of query plans can be used to update cardinality estimates. Database cracking, adaptive indexing, and adaptive query processing all incorporate elements of feedback as well.

In this vision paper, we argue that recent advances in deep reinforcement learning (DRL) can be applied to query optimization, resulting in a “hands-free” optimizer that (1) can tune itself for a particular database automatically without requiring intervention from expert DBAs, and (2) tightly incorporates feedback from past query optimizations and executions in order to improve the performance of query execution plans generated in the future.

If we view query optimisation as a DRL problem, then in reinforcement learning terminology the optimiser is the agent, the current query plan is the state, and each available action Continue reading

1 … 1,374 1,375 1,376 1,377 1,378 … 3,872