The Hidden Bottlenecks of Scaling Out: Network, CPU, and Memory

Scaling out databases with numerous small instances can lead to network and other resource bottlenecks, particularly for stateful applications like databases and caching layers. These bottlenecks often arise from imbalances in CPU, memory, disk bandwidth, and network bandwidth. Smaller instances typically face limitations such as lower guaranteed network bandwidth, reliance on burstable performance, and increased internal traffic, leading to issues like latency spikes and throttling during high demand. Redis and Valkey, due to their single-threaded nature, force users into a scale-out model with small instances, which accentuates these resource constraints. In contrast, Dragonfly is a multi-threaded system that optimizes the use of large cloud instances, allowing for better resource allocation and performance by scaling vertically before horizontally. This architecture reduces operational complexity, minimizes noise from neighboring workloads, and handles uneven traffic distribution more effectively. Dragonfly’s design enables it to fully leverage the capabilities of large instances, providing consistent, high-throughput performance by consolidating workloads onto fewer, more powerful nodes, thus avoiding the inefficiencies and limitations of many small instances.