Partitioning is a core technique in data engineering that lets engineers scope computations to specific subsets of data — by date, customer, region, or other dimensions — rather than processing everything at once. A partitioned data asset sits between a single monolithic asset and a set of distinct assets, enabling bulk operations like backfills while still allowing independent per-partition debugging and reruns. Effective data orchestration requires native partition awareness to answer questions like 'is my data up-to-date?' and 'which partitions need to be rerun?' Task-centric orchestrators like Airflow struggle here because partitions are a property of data, not tasks. Dagster models partitions natively via PartitionsDefinition types (Hourly, Static, Dynamic, Multi-dimensional), supports partition dependency mappings, partition-aware backfills, and auto-materialization. Complex scenarios covered include cross-frequency dependencies (daily depending on hourly), rolling window dependencies, multi-dimensional partitions, and dynamic partitions added at runtime.
