Concurrency-Aware SLA Scheduling and Admission Control for Analytical Queries in a Data Lake

Lakehouse query engines increasingly serve mixed analytical workloads under per-query deadlines alongside continuous ingestion, commonly assumed the primary risk to deadline compliance. Four independent tests show otherwise: ingestion has little measurable effect on query latency here; query concurrency is the dominant factor – a single-node Trino deployment slows a point-lookup query 19× at sixteen concurrent queries versus one. We reframe scheduling accordingly. Conditioning runtime prediction on current concurrency materially changes scheduling decisions in simulation, and live evaluation confirms the broader contention-driven collapse in deadline adherence and the aggregate advantage of dynamic-programming (DP) based scheduling, though not a consistent live edge for adaptive over static prediction specifically. An objective-optimal DP scheduler closes much of the remaining gap. Beyond a certain concurrency level no ordering satisfies every deadline; the more defensible response is admission control, deferring excess load rather than letting it fail: admitted queries met deadlines more often than the blended average in every live condition tested, and under live moderate-concurrency contention the mechanism deferred 82.5% of arrivals once its feasibility test judged further admission unsafe. We report ingestion’s null effect as a rigorously tested finding, and state this work’s scope plainly: predictions are point estimates, concurrency is modeled at three coarse levels, and one live admission-control result falls short of its predicted target – discussed, not hidden.

This manuscript is a preprint and has not been peer reviewed.