Scale Out and Conquer: Architectural Decisions Behind Distributed In-Memory Systems