Experiment-driven System Management

Опубликовано 17 августа 2016, 21:52

Experiment-driven management is an important and overlooked part of system administration that aims to automate the trial-and-error practices used to perform certain administrative tasks today. I will first present how experiment-driven management simplifies tuning, benchmarking, testing, troubleshooting, and capacity planning for a variety of systems like databases, storage servers, and MapReduce frameworks. I will then present Xplus, a new query optimizer that employs principles of experiment-driven management to tune poorly-performing SQL queries. A user or higher-level tuning tool can tell Xplus: 'I am not satisfied with the performance of the plan being used for the query Q that runs repeatedly. Can you generate a 50 better plan or a guarantee that no such plan exists for the current database configuration and optimizer cost model. Xplus uses a new abstraction of the physical plan space and an extensible architecture comprising multiple SQL-tuning experts. Finally, I will discuss why it is important to make experiment-driven management accessible to a wide range of users beyond system administrators; and present ongoing work on a language and run-time system being developed to achieve this goal on cloud computing platforms.