Search papers, labs, and topics across Lattice.
The paper introduces Spark Policy Toolkit, a system designed to enable scalable policy learning in Spark by addressing limitations in row-wise Python execution and driver-side candidate materialization. It introduces two Spark-native primitives: partition-initialized vectorized inference (mapInPandas/Arrow) and collect-less split search, both governed by a fixed-input semantic contract to ensure consistent results. Experiments on a 40-worker cluster demonstrate significant throughput improvements and the preservation of policy semantics at scale, highlighting the importance of throughput and collect-less execution.
Spark Policy Toolkit unlocks scalable policy learning in Spark by guaranteeing consistent results even with distributed execution, finally making it possible to apply complex policy learning techniques to large datasets.
Custom policy-learning pipelines in Spark fail for two coupled systems reasons: rowwise Python execution makes inference impractical, and driver-side candidate materialization makes split search fragile at feature scale. We present Spark Policy Toolkit, a semantics-governed systems toolkit for scalable policy learning in Spark. The toolkit provides two Spark-native primitives: partition-initialized vectorized inference through mapInPandas and mapInArrow, and collect-less split search that scores candidates on executors. Both primitives are governed by one fixed-input semantic contract: the same rows, feature order, treatment vocabulary, preprocessing manifest, and split boundaries must preserve per-row score vectors, best-split decisions, and end-to-end learned policy outputs. The evaluation combines practical baseline ladders, backend parity checks, measured split-search scale results, synthetic and Hillstrom end-to-end policy preservation, missingness stress, partition and order perturbation tests, quantile-boundary sensitivity, and a concrete adversarial failure catalog. On a 40-worker Databricks cluster, mapInArrow reaches 4.72M rows/s at 10M matched rows and 7.23M rows/s at 50M rows, while collect-less split search remains valid from F = 10 through F = 1000 with 124000 candidate rows, where the driver-collect baseline is intentionally skipped. Across 24 backend-ablation settings, mapInArrow wins 18 while mapInPandas wins 6, so the paper treats backend choice as workload-dependent rather than universal. Once the fixed-input lock is enforced, all six tested repartition/coalesce/shuffle perturbations preserve identical signatures; before lock, all six drift. The central result is not speed alone: throughput and collect-less execution are the mechanisms that let policy semantics survive at Spark scale.
