Distributed coordination

[andrewdavidmackenzie]

Flow is partitioned and more than one coordinator is involved in it's execution

[andrewdavidmackenzie]

• each node can have a different runtime and hence a different set of available context functions.

• partition the function graph over multiple nodes, either by manually drawing the boundaries, or assigning functions to a node, constrained by the available context functions and those used by the flow.

• Break up the function graph into multiple sub-graphs, each assigned to a node.

• transparent sending/receiving of Values between nodes at the sub-graph boundaries

  • need a unique flow id when deployed to identify destination/source, node id, function and input id in the flow
  • stub on sending subgraph to send
  • driver on receiving subgraph to receive

• coordinator on each node to receive it's subgraph for execution, the same as if a full graph.

• initial deployer does the partitioning (discovering other nodes and their capabilities first) and then submits the subgraphs to each as needed

• when remote node disconnects, that's like an EOF/Null allowing each subgraph to end.