Broadcast

Allows elements of a stream to be broadcast to multiple workers.

As the elements arrive, they are broadcast to all workers that have started evaluation before the element was pulled.

Elements are pulled as chunks from the source and the next chunk is pulled when all workers are done with processing the current chunk. This behaviour may slow down processing of incoming chunks by faster workers. If this is not desired, consider using the prefetch and prefetchN combinators on workers to compensate for slower workers.

Often this combinator is used together with parJoin, such as:

 Stream(1,2,3,4).covary[IO].broadcast.zipWithIndex.map { case (worker, idx) =>
   worker.evalMap { o => IO(println(s"$idx: $o")) }
 }.take(3).parJoinUnbounded.compile.drain.unsafeRunSync()

Note that in the example, above the workers are not guaranteed to see all elements emitted. This is due to different subscription times of each worker and speed of emitting the elements by the source. If this is not desired, consider using broacastThrough and broadcastTo, which are built on top of Broadcast.through, as an alternative. They will hold on pulling from source if there are no workers ready.

When source terminates, then the inner streams (workers) are terminated once all elements pulled from source are processed by all workers. However, note that when that source terminates, resulting stream will not terminate until the inner streams terminate.

Like apply but instead of providing a stream of worker streams, it runs each inner stream through the supplied pipes.

Supplied pipes are run concurrently with each other. Hence, the number of pipes determines concurrency. Also, this guarantees that each pipe will view all O pulled from source stream, unlike broadcast.

Resulting values are collected and returned in a single stream of O2 values.