RDD

Resilient Distributed Dataset (RDD) is a core data abstraction in Apache Spark. It represents an immutable, partitioned collection of elements that can be processed in parallel across a cluster. RDDs are fault-tolerant because they retain a lineage graph describing how each dataset was derived, allowing lost partitions to be recomputed from the original data.

RDDs support two types of operations: transformations and actions. Transformations, such as map, filter, flatMap, union, intersection, and distinct, return a new RDD and are evaluated lazily. Actions, such as count, collect, reduce, first, and take, trigger computation and return a value to the driver or write results to storage. For key-value data, specialized operations like reduceByKey, groupByKey, and combineByKey enable aggregation and joins.

Dependencies between RDDs are declared as narrow or wide. Narrow dependencies (e.g., map, filter) allow produced partitions to be computed using a single parent partition, enabling pipelining. Wide dependencies (e.g., shuffle, join) involve data movement across executors and incur a shuffle cost.

RDDs can be created from external storage systems (HDFS, local file systems, cloud storage) or from existing RDDs via transformations. They can be persisted in memory or on disk using various storage levels, such as MEMORY_ONLY, MEMORY_AND_DISK, or serialized forms, to optimize iterative computations.

Runtimes and language support include Scala, Java, Python, and R. While Spark later introduced DataFrames and Datasets with optimizations via the Catalyst optimizer, RDDs remain a lower-level, flexible API useful for fine-grained control, custom data types, or algorithms not easily expressed in SQL-like operations.