Coordinated Design of Cache Consistency and Invalidation Recovery in High Concurrency Backends

Abstract

Databases and caches jointly support typical read and write paths in high concurrency backends. When they are operated independently, successful writes, cache invalidation, and stale query fills can interleave and create races. This study examines single deletion, delayed double deletion, epoch token validation, and a combined strategy that integrates validation with invalidation replay in the cache-aside pattern. A finite-state model with one key, one reader, and one writer enumerates ten event schedules under specified partial-order constraints. Three delivery conditions are considered: normal delivery, loss of the first invalidation, and loss of all invalidations. Model execution shows that, under normal delivery, single deletion and double deletion leave stale cached data in two schedules and one schedule, respectively. Token validation eliminates these residual states under the stated assumptions but still fails when the first invalidation is lost. With successful replay, the combined strategy restores convergence within the model. These results distinguish two responsibilities: preventing stale requests from filling the cache and ensuring that invalidation eventually executes. Both require coordinated treatment. The contribution is a set of reproducible concurrency checks and a design framework that separates fill control, event recovery, and application read semantics. The experiment does not measure production-cluster throughput or establish global strong consistency.