Blackboard

The blackboard pattern is an architectural pattern for problems that have no known deterministic solution strategy. A set of independent, specialized components called knowledge sources cooperate to build a solution incrementally, communicating only through a shared data structure known as the blackboard.

The blackboard holds the partial, evolving state of the problem. Each knowledge source watches that state and contributes its own piece of expertise whenever it can make progress. There is no fixed order of execution. A separate control component decides which knowledge source acts next, based on what is currently on the blackboard. The solution emerges opportunistically, piece by piece, until it is complete.

The pattern originated in artificial intelligence. Its canonical example is the HEARSAY-II speech recognition system, developed at Carnegie Mellon University in the 1970s, in which acoustic, phonetic, lexical, and semantic specialists each refined a shared hypothesis about a spoken utterance. The pattern was later codified as an architectural pattern in Pattern-Oriented Software Architecture.

The blackboard pattern suits problems that combine incomplete knowledge, uncertain data, and several distinct areas of expertise, eg. image understanding, fault diagnosis, and complex planning. When a problem can be decomposed into a fixed pipeline of stages, pipe and filter is the simpler choice.

Trade-offs

  • Opportunistic control. Knowledge sources need not know about each other, only about the blackboard. New specialists can be added without disturbing the rest of the system.
  • Loose coupling. Because all communication flows through the shared space, knowledge sources are decoupled from one another and independently testable.
  • Shared state as a bottleneck. The blackboard is a single point of failure and a serialization point. Scaling it across many nodes is hard; keeping replicas consistent generally requires scatter-and-gather style coordination.
  • Non-deterministic execution. With no fixed control flow, debugging and reproducing failures is difficult.

Implementations

The tuple-space model introduced by the Linda language is the best-known mechanism for implementing a blackboard. Workers read, write, and take typed tuples from a shared space. JavaSpaces, part of the Jini framework, and the commercial GigaSpaces product bring tuple spaces to the JVM.

See also

References