Skip to content

Concurrent Actions

FaaSr supports the invocation of multiple concurrent actions in the workflow. This is useful in workflows where the workflow developer is able to determine that multiple functions can safely execute in parallel - for example, when different functions operate on independent files/datasets There are two approaches possible to invoke actions to execute concurrently:

  • Multiple distinct actions explicitly invoked with edges from the same predecessor in the workflow DAG, or
  • A single action that is invoked N times, where N is a configurable integer number, greater than 1; each invocation of the action is assigned a unique rank (an integer from 1 to N) that is available to the user function via the faasr_rank API

Multiple distinct actions

Consider the example below, where the action DownloadData invokes three concurrent actions that are distinct and explicitly added to the workflow graph: ComputeA, ComputeB, and ComputeC:

alt text

Each one of these actions can invoke a different user function concurrently. At the end, the action Merge executes the last user function, only after the three distinct actions ComputeA, ComputeB, and ComputeC have completed.

Ranked actions

Ranked actions allows you to express concurrency with multiple, parameterized invocations of the same user function. Consider the workflow below, where the action DownloadData now invokes the same action ComputeRanked three times, concurrently:

alt text

Each concurrent ComputeRanked action invokes the same user function. Similarly to the previous example, at the end, the action Merge executes the last user function, only after all three ComputeRanked actions have completed.

Inside the user function invoked by each ComputeRanked action, the faasr_rank API can be invoked to return a unique number (1, 2, or 3 in this example) that can be used to differentiate between invocations. As described in the Python API and R API documentation, faasr_rank allows the user function to determine the maximum rank (3 in this example) in addition to the per-action rank (1, 2, or 3).

This information can be used, for example, to select different file names, folders, or data subsets that each action operates on. It is the responsibility of the programmer to properly use the rank information to partition datasets for safe concurrent execution, depending on their workflow and application logic.