Two-level caching uses an in-process L1 and a shared Redis L2. What is the main consistency challenge?