In-Memory Databases: Evolution, Advantages, Challenges, and Future Directions

In‑memory databases, also called main‑memory databases, are DBMSs that store data primarily in RAM, offering dramatically faster read/write speeds compared to traditional disk‑based databases.

Traditional databases improve performance by adding memory buffers, but in‑memory databases place the entire database in RAM, achieving orders‑of‑magnitude speed gains for latency‑sensitive workloads.

1. Maturity of memory technology

Memory capacity has grown rapidly: early 64 KB chips, 256 KB SIMM after 80286, 512 KB‑2 MB SIMM in the late 80s, 4‑16 MB EDO‑DRAM in the early 90s, 64 MB SDRAM in 1995, and DDR3 reaching 16 GB by 2019. Prices have fallen nearly nine orders of magnitude, with 1 GB costing only $3‑5 in 2019, making large‑scale in‑memory storage feasible.

These trends enable massive data to be stored and processed directly in memory.

2. Bottlenecks and breakthroughs

Traditional storage hierarchies place hot data near the CPU; in‑memory databases keep all data in DRAM, which is volatile and costly at large scales, requiring persistence solutions.

Persistent memory (PM), also known as Storage Class Memory (SCM), sits between DRAM and SSD, offering load/store access with data durability. While slower than DRAM, PM provides larger capacity and lower cost; compared to NAND SSD, it offers better latency but lower capacity.

The evolution of in‑memory databases includes four phases: prototype, theoretical maturity, market growth, and rapid expansion.

Advantages: High‑performance read/write

Eliminating disk I/O allows microsecond‑level latency; single‑node QPS can exceed 100 k, and with user‑space networking and huge pages, tens of millions of QPS are achievable—far beyond traditional relational databases.

Challenges: Data volatility

DRAM’s volatility necessitates persistence mechanisms; current key‑value in‑memory databases provide limited durability, and persisting every operation degrades performance. Strategies include full persistence per operation or periodic checkpointing, each with trade‑offs.

Emerging non‑volatile memory technologies promise to resolve volatility, potentially expanding in‑memory database use cases.

Source: China Academy of Information and Communications Technology