Bug 2073847

Description of problem:
When postgres service is restarted (i.e. as part of all services restart or alone) when dynflow is about to complete a task, then the task can end up hung in a few invalid situations forever.

"Invalid situation" means e.g.:
- foreman sees the task as stopped/pending while dynflow sees it as stopped/succes
- or foreman sees the task as running/pending while dynflow sees it as stopped/success

"Forever" means there is no user action to fix the status, like:
- services restart doesnt help
- force unlock can move foreman task from running/pending to stopped/pending, but nothing else

Also, until force unlock is done, such stuck task can have acquired its object(s) lock(s).


Version-Release number of selected component (if applicable):
Sat6.10.4


How reproducible:
100% within a few attempts


Steps to Reproduce:
One particular reproducer is to Destroy a CV and just at the end, restart postgres service. It can be VERY tricky to guess the "at the end", so the script below checks for number of completed pulp tasks - for a CV with one repo, the ContentView::Destroy task triggers one pulp task. So whenever the script detects as many new completed pulp tasks as the number of being-destroyed CVs is, the script restarts postgres.

Script itself:

--------8<----------------8<----------------8<--------
CONCUR=${1:-5}   
REPOIDS=${2:-51}
hmr="hammer shell"

prepare_cv_to_delete() {
        CVID=$1
(       echo "content-view create --organization-id=1 --name cv_zoos_${CVID} --repository-ids ${REPOIDS}"
        echo "content-view publish --organization-id=1 --name cv_zoos_${CVID}"
        echo "content-view remove-from-environment --organization-id=1 --name=cv_zoos_${CVID} --lifecycle-environment-id=1"
        echo "content-view version delete --content-view=cv_zoos_${CVID} --version 1.0 --organization-id 1"
) | $hmr
}

for i in $(seq 1 $CONCUR); do
        prepare_cv_to_delete $i &
done

echo "waiting for CVs create+almost-delete"
time wait

for i in $(seq 1 $CONCUR); do
        hammer content-view delete --name=cv_zoos_${i} --organization-id 1 &
done

echo "$(date): waiting for CVs delete"
tasks=$(su - postgres -c "psql pulpcore -c \"copy (select count(*) from core_task) to stdout;\"")
echo "$(date): waiting for CVs delete, pulp tasks=${tasks}"
expected=$((tasks+CONCUR))
tasks=0
while [ $tasks -lt $expected ]; do
        tasks=$(su - postgres -c "psql pulpcore -c \"copy (select count(*) from core_task) to stdout;\"")
        sleep 0.5
done
#su - postgres -c "psql pulpcore -c \"select count(*) from core_task;\""
echo "$(date): restarting postgres as having tasks=${tasks}"
systemctl restart rh-postgresql12-postgresql.service
date
time wait
su - postgres -c "psql pulpcore -c \"select count(*) from core_task;\""
--------8<----------------8<----------------8<--------

Usage:

./create_delete_cv_restart_postgres.sh 5 REPOID

where REPOID is an id of a small repo


Actual results:
Random tasks tuck forever, optionally with acquired locks.

As an example, see attached task export.


Expected results:
No such stuck tasks forever. Tasks should be recoverable by a restart or manual (Skip&)Resume.


Additional info: