Originally suggested in JENKINS-26130 but I felt it was better to split this out. JENKINS-26130 does at least provide a much more comprehensible diagnosis for the problem.

Sounds related to what R. Tyler Croy is experiencing with JENKINS-41569.

FWIW, I am experiencing the same issue with the mesos cloud provider.

Jesse Glick did you have any WIP on this or is someone free to take a run at this? as it is biting a few people people now (not mentioned on this ticket)

I had nothing in progress; up for grabs. The tricky bit is identifying the case you are running under reliably. If you are actually dealing with an EphemeralNode, probably the ExecutorPickle could record that fact and fail promptly upon rehydration. The harder cases are nodes that get reattached by some outside process sometime after start—you need to give them a grace period.

thanks Jesse Glick yes on a hangout that conclusion was made to - to have some grace period for things that happen outside, that should be fine. 

I have a testcase that reproduces the issue and a code fix kills the build.  Working on adding the timeout to that solution, and looking at ways to limit the scope of builds that are killed to avoid it getting overeager.

nice Sam Van Oort!

FYI I've cut beta 2.14-beta to try out of this if people want try it

On second thought, hold on, I'm going to cut a different one. 

Fix released with durable task step 2.14

After upgrading from 2.13 to 2.14, my pipeline jobs that use the swarm plugin can't survive a restart anymore.

Under 2.13, my pipeline jobs were able to survive a restart if I used the -deleteExistingClients option that the swarm plugin has. Without -deleteExistingClients, I used to get this:

SEVERE: RetryException occurred
hudson.plugins.swarm.RetryException: Failed to create a slave on Jenkins, response code: 409
A slave called 'create-dc-slave-1-eff65ede' already exists and legacy clients do not support name disambiguation

        at hudson.plugins.swarm.SwarmClient.createSwarmSlave(SwarmClient.java:448)
        at hudson.plugins.swarm.Client.run(Client.java:134)
        at hudson.plugins.swarm.Client.main(Client.java:87)

But this was basically working once I added -deleteExistingClients:

Resuming build at Fri Aug 25 23:54:47 UTC 2017 after Jenkins restart
Waiting to resume part of test #5: create-dc-slave-1-eff65ede is offline
Waiting to resume part of test #5: There are no nodes with the label ‘create-dc-slave-1-eff65ede’
Waiting to resume part of test #5: Jenkins doesn’t have label create-dc-slave-1-eff65ede
create-dc-slave-1-eff65ede is offline
Ready to run at Fri Aug 25 23:54:58 UTC 2017

Now, after upgrading to 2.14, this doesn't work at all:

Resuming build at Sat Aug 26 00:57:51 UTC 2017 after Jenkins restart
[Pipeline] End of Pipeline
ERROR: Killed hudson.model.Queue$WaitingItem:ExecutorStepExecution.PlaceholderTask{runId=test#6,label=create-dc-slave-1-eff65ede,context=CpsStepContext[3:null]:Owner[test/6:test #6],cookie=da72341a-9bff-4c9d-9607-8ff0c2e81595}:29 because EphemeralNode create-dc-slave-1-eff65ede is never going to reappear, by definition!
Finished: FAILURE

Since I rely on the swarm plugin in my pipeline jobs, I view this as a major regression. It would be nice to have this functionality restored.

Also note that I'm using the latest version of the swarm plugin and swarm client (3.4 on both ends).

Basil Crow I am sorry to hear that this caused a regression for you – it appears to be an unanticipated case where EphemeralNodes generated by the Swarm Plugin aren't really following the contract of that interface, since they can reconnect and be recreated but will not do so immediately. 

 I have a fix supplied here – https://github.com/jenkinsci/workflow-durable-task-step-plugin/pull/48 – pending review by Jesse Glick.  This will apply a 5 minute timeout, restoring my original implementation strategy. 

Basil Crow sounds like a bug in the Swarm plugin to me. The implementation of JENKINS-34593 ought to have removed the EphemeralNode marker if I understand it correctly. The description of -deleteExistingClients does not seem to talk about retaining agents across restarts but it seems that you have discovered that as a use case—probably not a tested one. (As an aside, it seems the plugin contains no tests which actually run Jenkins, much less tests of Pipeline interoperability or of restart behavior.)

Sam Van Oort and Jesse Glick, I wanted to say thank you for releasing version 2.15 which restores this functionality. I tested it, and my pipeline jobs that use the swarm plugin once again survive Jenkins restarts. Thanks!

hi five Sam Van Oort (but make sure he has washed his hands, he just finished coming second in a chilli eating competition)!

Basil Crow Thanks!  I'm glad you're finding it works well for you now