Implements a process-based worker pool using ZMQ for inter-process communication. Provides start, ventilate (distribute work items), get_results (collect results), and stop lifecycle. The ventilator pattern decouples work distribution from result collection, enabling deterministic ordering and throttling. This module is self-contained with minimal internal dependencies (imports from workers_pool.ventilator and workers_pool.worker_base) and could be extracted as a standalone parallel task executor.

Provides an abstract shuffling buffer base class (ShufflingBufferBase) and concrete implementations: RandomShufflingBuffer and NoopShufflingBuffer. The pattern enables decorrelation of row groups during training data loading, with configurable buffer size and support for both row-level and batch-level shuffling (BatchedRandomShufflingBuffer). The abstraction is clean, isolated, and can be reused in any streaming data pipeline requiring controlled shuffling.

Implements FilesystemResolver, a factory that creates a PyArrow filesystem from a dataset URL, supporting local, HDFS, S3, GCS, and ADLS via fsspec. Handles path resolution, file globbing, and serialization for distributed workers. The resolver pattern abstracts storage backend details behind a simple interface, making it reusable in any project that needs to read from multiple filesystems with a consistent API.