Advanced future and promises usage

The problem with future()+promise()

In an ideal situation, the number of available future workers (future::nbrOfFreeWorkers()) is always more than the number of future::future() jobs. However, if a future job is attempted when the number of free workers is 0, then future will block the current R session until one becomes available.

For a concrete example, let’s imagine a scenario, where seven plumber requests are received at the same time with only two future workers available. Also, let’s assume the plumber route(s) serving the first 6 requests use future::future() and take ~10s to compute slow_calc():

#* @get /slow/<k>
function() {
  future::future({
    slow_calc()
  })
}

Let’s also assume the plumber route serving the last request does not use any form of future or promises and takes almost no time to compute.

#* @get /fast/<k>
function() {
  fast_calc()
}

The figure below depicts the overall timeline of execution of these 7 requests under the conditions we’ve outlined above. Note that the y-axis is ordered from first request coming in (/slow/1) to the last request (/fast/7).

Note how R has to wait 20s before processing the 7th request (shown in green). This is a big improvement over not using future+promises at all (in that case, R would have to wait 60s before processing). However, since there are only two future workers available R still has to wait longer than necessary to process that last request because the main R session must wait for a future worker to become available. The video below animates this behavior:

The solution: future_promise()

The advantage of using future_promise() over future::future() is that even if there aren’t future workers available, the future is scheduled to be done when workers become available via promises. In other words, future_promise() ensures the main R thread isn’t blocked when a future job is requested and can’t immediately perform the work (i.e., the number of jobs exceeds the number of workers).

Continuing with the example above, we can swap out the calls to future::future() with future_promise().

#* @get /slow/<k>
function() {
  promises::future_promise({
    slow_calc()
  })
}

With this change to future_promise(), note how the /fast/7 route now does not have to wait on future work to finish processing. Therefore, plumber can complete the last requests almost immediately:

The vertical gray bars in the figure above represent timepoints where the main R session is actually busy. Outside of these gray areas, the R session is free to do other things, for example, executing other promises or, more generally, non-future work. The video below animates this behavior: