Autotuning Geo-distributed systems

Abstract

Geo-distributed databases must serve transactions close to where their data is accessed to keep latency low, which is why systems like Detock assign each data item a ”home” region and route its transactions there. But Detock leaves the placement policy open: where should each key live, and when should it move? Reactive heuristics fill this gap poorly. Deciding one key at a time, they split co-accessed records across regions and turn otherwise single-home transactions into expensive multi-home ones. Migrations that fire too often or at the wrong time also force concurrent transactions reading a stale home to restart.

This paper presents an adaptive placement agent for Detock. It groups co-accessed keys into communities using the Leiden algorithm, then uses a contextual bandit to decide when and where to migrate each community. The bandit learns from the locality and restart signals it observes at runtime, with no prior knowledge of the workload. We evaluate the agent on a two-region, follow-the-sun deployment of the Product–Parts–Supplier benchmark. Under matched migration budgets, the agent tracks the shifting hotspot as well as a DynaMast-style baseline while issuing about 29% fewer home-movements and causing roughly 4.3× fewer transaction restarts. These results show that a lightweight, group-aware and cost-aware learned policy is a practical way to drive home-movement in a deterministic geo-distributed database.