future() and resolved() handle FutureError:s differently for different backends

[HenrikBengtsson]

Issue

Future orchestration errors (i.e. FutureError) occurring when calling future() and resolve() are handled differently depending on future backend. Below are a few examples.

multicore

library(future)
plan(multicore, workers = 2)

segfault <- function(ii) {
  if (ii == 2) tools::pskill(Sys.getpid()) else Sys.sleep(1)
  ii
}

fs <- lapply(1:4, FUN = function(ii) {
  message(sprintf("Launching future #%d", ii))
  future({ segfault(ii) })
})
#> Launching future #1
#> Launching future #2
#> Launching future #3
#> Launching future #4
#> Warning message:
#> In mccollect(jobs = jobs, wait = TRUE) :
#>   1 parallel job did not deliver a result
#> Calls: lapply ... value.Future -> result -> result.MulticoreFuture -> mccollect

resolved(fs)
#> [1] TRUE TRUE TRUE TRUE

fs <- resolve(fs)

rs <- lapply(fs, FUN = result)
#> Error: Failed to retrieve the result of MulticoreFuture (<none>) from the
#> forked worker (on localhost; PID 687068). Post-mortem diagnostic: No
#> process exists with this PID, i.e. the forked localhost worker is no longer
#> alive. The total size of the 2 globals exported is 6.52 KiB. There are two
#> globals: 'segfault' (6.47 KiB of class 'function') and 'ii' (56 bytes of
#> class 'numeric')

vs <- value(fs)
#> Error: Failed to retrieve the result of MulticoreFuture (<none>) from the
#> forked worker (on localhost; PID 687068). Post-mortem diagnostic: No
#> process exists with this PID, i.e. the forked localhost worker is no longer
#> alive. The total size of the 2 globals exported is 6.52 KiB. There are two
#> globals: 'segfault' (6.47 KiB of class 'function') and 'ii' (56 bytes of
#> class 'numeric')

multisession

library(future)
plan(multisession, workers = 2)

segfault <- function(ii) {
  if (ii == 2) tools::pskill(Sys.getpid()) else Sys.sleep(1)
  ii
}

fs <- lapply(1:4, FUN = function(ii) {
  message(sprintf("Launching future #%d", ii))
  future({ segfault(ii) })
})
#> Launching future #1
#> Launching future #2
#> Launching future #3
#> Error in unserialize(node$con) : 
#>   MultisessionFuture (<none>) failed to receive message results from
#> cluster RichSOCKnode #2 (PID 687611 on localhost 'localhost'). The reason
#> reported was 'error reading from connection'. Post-mortem diagnostic: No
#> process exists with this PID, i.e. the localhost worker is no longer alive.
#> The total size of the 2 globals exported is 6.52 KiB. There are two
#> globals: 'segfault' (6.47 KiB of class 'function') and 'ii' (56 bytes of
#> class 'numeric')

future.callr::callr

library(future)
plan(future.callr::callr, workers = 2)

segfault <- function(ii) {
  if (ii == 2) tools::pskill(Sys.getpid()) else Sys.sleep(1)
  ii
}

fs <- lapply(1:4, FUN = function(ii) {
  message(sprintf("Launching future #%d", ii))
  future({ segfault(ii) })
})
#> Launching future #1
#> Launching future #2
#> Launching future #3
#> Launching future #4

> resolved(fs)
#> Error: CallrFuture (<none>) failed. The reason reported was '! callr
#> subprocess failed: could not start R, exited with non-zero status, has
#> crashed or was killed'. Post-mortem diagnostic: The parallel worker
#> (PID 686807) started at 2023-08-08T09:08:46+0000 finished with exit
#> code -15. The total size of the 2 globals exported is 6.52 KiB. There
#> are two globals: 'segfault' (6.47 KiB of class 'function') and 'ii'
#> (56 bytes of class 'numeric')

fs <- resolve(fs)
#> Error: CallrFuture (<none>) failed. The reason reported was '! callr
#> subprocess failed: could not start R, exited with non-zero status, has
#> crashed or was killed'. Post-mortem diagnostic: The parallel worker
#> (PID 686807) started at 2023-08-08T09:08:46+0000 finished with exit
#> code -15. The total size of the 2 globals exported is 6.52 KiB. There
#> are two globals: 'segfault' (6.47 KiB of class 'function') and 'ii'
#> (56 bytes of class 'numeric')


> rs <- lapply(fs, FUN = result)
#> Error: CallrFuture (<none>) failed. The reason reported was '! callr
#> subprocess failed: could not start R, exited with non-zero status, has
#> crashed or was killed'. Post-mortem diagnostic: The parallel worker
#> (PID 686807) started at 2023-08-08T09:08:46+0000 finished with exit
#> code -15. The total size of the 2 globals exported is 6.52 KiB. There
#> are two globals: 'segfault' (6.47 KiB of class 'function') and 'ii'
#> (56 bytes of class 'numeric')

> vs <- value(fs)
#> Error: CallrFuture (<none>) failed. The reason reported was '! callr
#> subprocess failed: could not start R, exited with non-zero status, has
#> crashed or was killed'. Post-mortem diagnostic: The parallel worker
#> (PID 686807) started at 2023-08-08T09:08:46+0000 finished with exit
#> code -15. The total size of the 2 globals exported is 6.52 KiB. There
#> are two globals: 'segfault' (6.47 KiB of class 'function') and 'ii'
#> (56 bytes of class 'numeric')

Suggestion

Harmonize the behavior. This is related to releasing future slots for failed futures.

• future() should attempt to release future slots, if possible. If there are available slots, it should use those and only produce a FutureError if no more slots are available.

• resolved() should never give an error. may throw a FutureError per https://future.futureverse.org/articles/future-6-future-api-backend-specification.html#resolved-method.

See also

This is related to futureverse/future.callr#11.