Stop Hand-Tuning MySQL: Teaching a Kubernetes Operator to Do It for You

Ask ten DBAs how to size innodb_buffer_pool_size and you’ll get ten formulas and none of them survive Kubernetes, where resources change with a one-line YAML edit and yesterday’s carefully tuned config is today’s OOMKill. Meanwhile, most MySQL clusters on Kubernetes still run near-default configurations, wasting the memory and CPU they’re allocated. In this talk we introduce Autopilot Mode in the Percona Operator for MySQL: the operator derives a full production-grade MySQL configuration. InnoDB buffer pool and redo log, session buffers, connection limits, Group Replication settings from each cluster’s actual CPU, memory, and a declared workload profile. It’s on by default, and it re-tunes automatically every time resources change: scale the pod, and the config follows. We’ll cover how tuning formulas that assumed dedicated hardware had to change for containers (sidecars, limits vs requests, headroom for OOM safety), how we embedded an open source configuration calculator into the operator’s reconcile loop without causing restart storms, and where the boundary lies between what automation should decide and what stays in the DBA’s hands.