Machine-learning pipelines increasingly span cloud providers, regions, Kubernetes clusters, storage systems, API gateways, accelerators, and governed enterprise data sources. Cross-cloud execution can improve cost, locality, resilience, and accelerator availability, but it also expands the surface for policy drift, inconsistent identity, data movement risk, and incomplete audit evidence. This paper surveys cross-cloud ML pipelines and multi-cluster governance from a 2025 systems perspective. We synthesize work on governed cloud data engineering, cross-cloud workload optimization, secure API gateway control, PMU stream processing, hybrid semantic-relational retrieval, historical legal profiles, agreement-gated learning, structured extraction, and GPU-parallel optimization. The survey frames multi-cluster intelligence as a joint control problem over placement, evidence validity, policy reconciliation, accelerator use, confidence gating, and rollback. A comparative coding of representative systems indicates that placement-aware optimization is most reliable when provider benchmarks are interpreted alongside data-governance constraints and cluster-local enforcement evidence. The resulting guidance is to separate optimization from authority: ML pipeline controllers should recommend placements and configuration changes, while policy verifiers, gateway controllers, and audit ledgers decide whether those changes are eligible to run.
Kodali et al. (Thu,) studied this question.