Abstract
A service spanning cloud, edge, and serverless nodes must balance latency, data exposure, carbon, and cost, which operators may rank strictly, e.g., meeting a latency SLO first. Weighted sums approximate such a ranking only with deployment-specific weights, and their decisions change when an objective’s numerical scale is misestimated. We present Lexi , a continuum orchestrator applying a thresholded lexicographic rule to predictions of an off-policy counterfactual outcome model, which assumes objectives decompose across stages. We instantiate the rule as SLO ≻ privacy ≻ carbon ≻ cost and use a simple topological privacy proxy only to separate placements already tied on the SLO. In a 15-seed simulation, Lexi needs no weights and reaches 1.46% SLO violations with PII exposure 0.05, comparable to hand-calibrated and SLO-constrained weighted sums. An untuned sum reaches 23% SLO violations and 1.94 PII exposure, and only one of 84 tested weightings meets a modest target. Under-scaling privacy 10× raises the tuned and constrained baselines’ exposure to 1.85 and 1.84 of two stages, respectively. Lexi is unchanged. The evidence covers one simulated application under the additive no-interference assumption. Keywords: Service orchestration, Cloud-edge-serverless continuum, Lexicographic optimisation, Scale invariance