PLB database load balancer keeps Premium QoS intact during replica failures

When a database replica fails, most systems default to spinning up replacements—but that takes time and extra resources. In constrained environments, immediate capacity restoration isn't an option, and service quality must degrade gracefully without flattening the tier hierarchy. A new paper from SPIRALS/CRIStAL researchers introduces PLB, a PostgreSQL JDBC middleware load balancer that automates differentiated QoS degradation by reassigning healthy replicas into Premium, Mixed, and Freemium roles after a fail-stop failure. The pool remains shared, but session routing respects each customer's service class, so Premium users don't get lumped in with Freemium traffic even at reduced capacity.

Evaluated under single and cascading failures, PLB significantly outperforms both isolated per-class pools and shared round-robin routing. The numbers are striking: median Premium goodput retention improves by 26-28 percentage points under a Premium-side fault, Premium goodput more than doubles during the worst cascading-failure phase, and Premium p95 latency drops by 18.2% versus round-robin. The repair-to-target policy runs entirely in middleware, meaning no changes to PostgreSQL itself. This offers a practical blueprint for operators who want predictable degradation with fixed budgets—preserving premium performance where it matters while keeping the cluster alive.