Add forknode method for isNodeAlive

[ralmond]

(Please use https://github.com/HenrikBengtsson/future/discussions for Q&A)

Wish or feature request

isNodeAlive() currently returns NA no matter the state of the cluster, which is forkin' useless.

Suggest adding a method for SOCK0node which checks the connection status.

ccc <- makeForkCluster(2)
> isNodeAlive(ccc)
[1] NA NA
> isNodeAlive.SOCK0node <- function (x, ...) isConnectionValid(x$con)
> isNodeAlive(ccc)
[1] TRUE TRUE
> stopCluster(ccc)
> isNodeAlive(ccc)
[1] FALSE FALSE
> 

See #136

[HenrikBengtsson]

FWIW, it's the same with all cluster:s created directly by the parallel package, e.g.

> cl <- parallel::makePSOCKcluster(2)
> cl
socket cluster with 2 nodes on host 'localhost'
> parallelly::isNodeAlive(cl)
[1] NA NA

I cannot find my notes on using isConnectionValid() to infer whether a cluster node is alive, but I recall I thought about it way back. I might have decided on a conservative approach. For example, what can we infer about the cluster node from isConnectionValid() == TRUE and isConnectionValid() == FALSE? It tells us whether the main R session can communicate with the node. But, I don't think we can safely infer that isConnectionValid() == FALSE means the node process is no longer running. For instance, it could very well be that it's running, but the R connection went down. If it is still running, do we want to conclude that isNodeAlive() == FALSE? If we do, we might launch another parallel worker, which then might fail the same way, and then we risk launching more and more cluster nodes all running 100% in the background, until we fry the current machine. I need to think more about this.

The way

> cl <- parallelly::makeClusterPSOCK(2)
> cl
> Socket cluster with 2 nodes on host 'localhost' (R version 4.5.2 (2025-10-31), platform x86_64-pc-linux-gnu)
> parallelly::isNodeAlive(cl)
[1] TRUE TRUE
> cl[[1]]
RichSOCKnode of a socket cluster on local host 'localhost' with pid 2860414 (R version 4.5.2 (2025-10-31), x86_64-pc-linux-gnu) using socket connection futureverse/parallelly#6 ('<-localhost:11556')
> parallelly::isNodeAlive(cl[[1]])
[1] TRUE

works, is that it checks whether the process ID is still running. This works because we know the node's PID for RichSOCKnode:s. This is gathered at the end of the set up using:

cl <- parallelly:::add_cluster_session_info(cl)

One approach forward could be to export that function so it can be used on:

cl <- parallel::makeForkCluster(2)
cl <- parallel::makePSOCKcluster(2)

Knowing the PID, we could then use the existing machinery to query whether the node process is alive. That way one could add an isNodeAlive() for SOCKnode and SOCK0node that checks if there's a registered PID, and if it is it can return TRUE or FALSE. If there's no PID, it can return NA (as now).

What do you think?

[ralmond]

It occurs to me that the node could be in one of three states:

1. Alive (and connected)
2. Dead (and disconnected)
3. Undead (and disconnected)

If I understand you correctly, isConnectionValid() can tell the difference between 1 and 2 or 3, but not which of 2 or 3 is there. From a user perspective, both dead and undead are bad as no worker is available to do work. If a process dies or unexpectedly looses communication while waiting for results, both conditions need to be debugged. From a sys admin perspective, undead processes require exorcism.

From a user perspective, it would be OK if the function returned TRUE if the connection was present, or NA if not present, as then isTRUE(isNodeAlive()) can be sensibly used to test for functional nodes.