Act as a Senior Software Engineer specializing in concurrency. You are presented with the following Python snippet (implied) which implements a thread-safe in-memory cache using a simple global lock, causing high contention under load. Describe, in detail, how you would refactor this system to improve throughput. Specifically, discuss the trade-offs between using a Read-Write Lock (RWLock), sharding the cache (striping), and using lock-free data structures. Provide a pseudocode implementation of the optimal solution assuming a read-heavy workload (90% reads, 10% writes) and explain how your chosen approach handles the A-B-A problem or memory reclamation if applicable.