Answers are grounded in this paper and cite their sources.
Distributed databases that store Industrial Internet of Things (IIoT) telemetry are usually sharded once, by device identifier or by time, and left in that state. The queries that reach them do not stay fixed: operator drill-downs, real-time monitoring and fleet-level analytics alternate over a working day, and each favours a different placement of the data. Re-sharding can recover lost performance, but moving millions of rows is expensive, and a system that re-shards on every change may spend more time migrating than it saves. This paper presents DriftShard, a framework that monitors the mix of query classes reaching a Citus cluster, detects regime changes with a total variation distance test, predicts the latency of every candidate layout with a gradient-boosted cost model, and migrates only when the predicted saving over a planning horizon exceeds the predicted migration time. DriftShard was implemented on PostgreSQL 16 with Citus 12.1 and evaluated with real migrations on a four-node cluster, using a synthetic sensor fleet (0.71 million rows) and the TON IoT telemetry collection (3.58 million rows). Across 40 randomised regime sequences, DriftShard was 42% and 30% faster than a static device-hash layout on the two datasets, and 31% and 53% faster than drifttriggered re-sharding that ignores migration cost. It did not match the best static layout chosen in hindsight, and its benefit depended on cost-model accuracy: with measured rather than predicted costs it improved by a further 6% and 13%. The results identify when adaptive sharding pays off and where learned cost models limit it.
No takes yet. Share an insight, caveat, or question.
Karunarathne et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: