dataflow

Dataflow is a concept in computer science describing computation modeled as the flow of data tokens through a graph of operations. Execution is triggered by availability of input data rather than a predefined sequence of statements. Nodes perform computations; edges carry data between nodes; graphs can be acyclic or cyclic with feedback.

In dataflow programming, programs are expressed as networks of operators; languages and systems implement dataflow execution. Data becomes available, and the corresponding operators fire; this enables implicit parallelism and can simplify concurrency control. Notable environments include LabVIEW, Pure Data, and Max/MSP, and dataflow graphs have been used in machine learning frameworks such as TensorFlow, as well as in streaming data systems that process data continuously.

Dataflow architectures are hardware designs where processing elements react to incoming data tokens, communicating through channels; they aim to alleviate bottlenecks of control-flow machines by enabling fine-grained parallelism. Early research explored data-driven processors and dataflow machines; modern variants appear in FPGA-based and streaming hardware designs.

Dataflow analysis is a static analysis technique used in compilers to infer properties about programs by propagating information along the control-flow graph. It includes dataflow equations for liveness, reaching definitions, available expressions, and constant propagation; it supports optimization and verification tasks.

Applications include data processing pipelines, extract-transform-load processes, and stream processing, with frameworks such as Apache Beam, Google Dataflow, Apache NiFi, and Spark streaming enabling scalable, parallel processing of data as it arrives. In practice, the term dataflow is used in different contexts, from programming models to compiler analysis to hardware design.