(I work with Marcos) We are starting to believe this issue is related to Pipeline resume capabilities. We can definitely produce a "build queue" hung state (jobs not queuing) after a Jenkins restart if there were pipeline jobs running at the time of the restart that weren't terminated (say a crash or unceremonious shutdown). Upon restart you don't even necessarily see the resuming jobs enter the queue, but we do end up deadlocked.

If we clear the following files:

• /var/lib/jenkins/queue.xml
• /var/lib/jenkins/org.jenkinsci.plugins.workflow.flow.FlowExecutionList.xml
• /var/lib/jenkins/org.jenkinsci.plugins.workflow.support.steps.StageStep.xml 

It often restarts cleanly, but we still get risk the problem described above where intermittantly the queue may sieze 1, 2, even 12 hours later.  We are having some luck then scanning for any jobs that may be trying to resume and cleaning their build history (expensive across a server with 3000+ jobs). In particular the build runs that may need to resume. I'm not clear, but it seems there's a background thread in Jenkins that scans for jobs that need to resume based on their history and resumes them.  This can, potentially, cause the thread deadlocks you see in our thread dump.

If we cull the histories things seem stable.  See Issue: https://issues.jenkins-ci.org/browse/JENKINS-33761 where we believe if we could disable resume capabilities we could avoid this.

This does not appear reproduceable on a small jenkins instance with only a few jobs and only a few needing to resume. But if you have a server like ours, with a few thousand jobs and potentially 20-40 jobs interrupted upon restart, it seems to happen. I don't yet have a garaunteed repro case. 

Versions of things we are running:

• Jenkins 2.63
• Pipeline 2.5
• Pipeline: API 2.16
• Pipeline: Basic Steps 2.5
• Pipeline: Build Step 2.5
• Pipeline: Groovy 2.34
• Pipeline: job 2.12
• Pipeline: Stage Step 2.2
• Pipeline: Step API 2.10
• Pipeline: Shared Groovy Libraries 2.8

It could be specific to Docker Plugin. https://github.com/jenkinsci/docker-plugin/blob/master/docker-plugin/src/main/java/com/nirima/jenkins/plugins/docker/strategy/DockerOnceRetentionStrategy.java#L106 is definitely a bad idea.

The same code exists in the Yet Another Docker Plugin: https://github.com/KostyaSha/yet-another-docker-plugin/blob/master/yet-another-docker-plugin/src/main/java/com/github/kostyasha/yad/strategy/DockerOnceRetentionStrategy.java#L123 . CC integer

There is also logic from the TCB plugin

logic come from https://github.com/jenkinsci/durable-task-plugin/blob/master/src/main/java/org/jenkinsci/plugins/durabletask/executors/OnceRetentionStrategy.java i guess non-pipeline jobs has no issues. 

I was theorizing that the Docker/Jclouds Plugins need to inspect the build queue to decide if they need to provision a new node.  I don't have heap dumps/thread dumps from our Jenkins server that doesn't use the Docker/Jclouds pplugins but we have seen the build queue hang on that server after restart as well. In fact we first saw the behavior there.  Your findings however are hopeful that potentially there's a way to fix this in the future.

integer Would switching to a different Retention strategy avoid this? I have to look at the Yet Another Docker Plugin to see what options exist that might emulate the behavior of provisioning and throwing away the container when done. Or do all the retention strategies have a lock on the build queue?  (I know this doesn't appear specific to the Docker plugins as tghe "Once Retention Strategy" is not just a Docker thing).

oleg_nenashev We definitely use the TCB (Throttle Concurrent builds) plugin on both our Jenkisn environments (non Docker Plugin and Docker Plugin) which is a good lead into why we might see this behavior on our more traditional Jenkins setup as well. 

If a pipeline is trying to resume AND has a throttle associated with it and is queued at the same time for another run, could that cause a deadlock? 

As more info, last night before going home, we crawled all the job folders on our server looking for any job that had any historical message in it's history logs of "attempting to resume". If we found that log message, we purged that build history from disk.

We restarted the jenkins server and have had 14 hours of consecutive up time (our previous high was about 2 hrs). Which is why we believe the culprit is the interaction with resume features of jenkins, but we're not sure exactly what causes Jenkins to believe a pipeline job needs to be resumed.  We know it's more than just purging the xml files in the jenkins home folder that control queue (queue.xml, org.jenkinsci.plugins.workflow.flow.FlowExecutionList.xml org.jenkinsci.plugins.workflow.support.steps.StageStep.xml ) because resumes still occur after clearing those.

After doing our build history purge we appeared to get no resumes (and no locks)

I also see ghprb plugin with some weird locks, maybe try disable it. 

We seem to have stabilized the server.

In our set up, a pipeline build that was interrupted would attempt to resume upon Jenkins restart. However when it does so it uses the exact node name that it had before which in this case is a docker container name with a unique hash that no longer exists. This causes the build to zombie itself and hang.

The only indication on the file system that a pipeline job has not completed is that most recent stage xml file in the workflow directory for a given build does not contain a "<result>" tag. We wrote a python script that we ran on the jenkins master server to look for that and delete any build that did not have a "<result>" tag. We would appreciate any feed back on this approach, the python script is attached to the ticket. 

The directory for a given job is: 

/var/lib/jenkins/jobs/<JOB_NAME>/builds/<BUILD_NUMBER>/workflow

CC abayer and svanoort ^^^

Looks like a clone of https://issues.jenkins-ci.org/browse/JENKINS-36013 to me – I have a fix almost ready to go for that one

Does the fix for JENKINS-36013 address this?

danielbeck yes, it should - albeit after a 5 minute timeout before the build aborts when it can't resume on the same node (safety measure).  mpedreiro can you try installing the latest version of durable task step plugin and let us know if that doesn't settle it out (bear in mind that there is a delay before the build is killed when it cannot resume cleanly). 

Thanks!

Closing as duplicate of the linked issue, since the fix there should cover this as well.