Rethinking Database Architectures for Memory-Centric Computing
TL;DR
A memory-centric design model in which a database expresses what it needs from memory as declarative properties and a runtime resolves them against whichever fabric is present, and three architectural principles from recent work that instantiate this approach at different layers of the stack are consolidated.
Abstract
Datacenter architectures have steadily moved toward resource disaggregation, storage first, and now memory itself. This shift is arriving through several competing fabrics (RDMA, CXL pooling) that show no sign of converging. Their shared features lead to treating them as one class: memory has stopped being local, uniform, or CPU-owned. But their differences in latency, coherence, ownership, and failure behavior are large enough that a database engine wired to any one of them is costly to retarget when the next one arrives. In this position paper, we present a memory-centric design model in which a database expresses what it needs from memory as declarative properties and a runtime resolves them against whichever fabric is present. We consolidate three architectural principles from recent work that instantiate this approach at different layers of the stack: logical memory regions, declarative memory services, and state-centric task DAGs.