[
https://issues.apache.org/jira/browse/SLING-12078?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Stefan Egli updated SLING-12078:
--------------------------------
Description:
Two regular cases where a job is stored as part of JobManager.addJob():
* when a topology is defined, it directly gets stored to the appropriate
assigned/target slingId subtree. This is the most frequent case by far.
* if no topology is defined (no TOPOLOGY_INIT received) yet, it gets put into
the unassigned subtree. Later upon receiving TOPOLOGY_INIT
CheckTopologyTask.fullRun() finds such unassigned jobs and moves them to the
corresponding assigned subtree.
There is a suspect race condition (test case to be provided), which happens
between the thread doing JobManager.addJob() and the thread handling the
TOPOLOGY_INIT:
* JobManager.addJob determines the target slingId - which is not yet defined,
as TOPOLOGY_INIT is just being handled concurrently
* CheckTopologyTask.fullRun(), as part of TOPOLOGY_INIT handling, however does
not yet find the above new job in unassigned, as the job is just being stored
concurrently.
The result is a job in the unassigned subtree, which waits until the next
TopologyEvent happens - which then invokes CheckTopologyTask.fullRun() - which
then finds the unassigned job and re/assigns it accordingly. So the job is
never lost, but substantially delayed due to this. (the frequency of
TopologyEvents depends on actual cluster/property changes happening in the
topology and can thus vary).
Tasks:
* provide a test case to reproduce
* fix the race-condition
* undo
https://github.com/apache/sling-org-apache-sling-event/commit/d16686705908099b26d0a3233f61c4e209880f93
was:
Two regular cases where a job is stored as part of JobManager.addJob():
* when a topology is defined, it directly gets stored to the appropriate
assigned/target slingId subtree. This is the most frequent case by far.
* if no topology is defined (no TOPOLOGY_INIT received) yet, it gets put into
the unassigned subtree. Later upon receiving TOPOLOGY_INIT
CheckTopologyTask.fullRun() finds such unassigned jobs and moves them to the
corresponding assigned subtree.
There is a suspect race condition (test case to be provided), which happens
between the thread doing JobManager.addJob() and the thread handling the
TOPOLOGY_INIT:
* JobManager.addJob determines the target slingId - which is not yet defined,
as TOPOLOGY_INIT is just being handled concurrently
* CheckTopologyTask.fullRun(), as part of TOPOLOGY_INIT handling, however does
not yet find the above new job in unassigned, as the job is just being stored
concurrently.
The result is a job in the unassigned subtree, which waits until the next
TopologyEvent happens - which then invokes CheckTopologyTask.fullRun() - which
then finds the unassigned job and re/assigns it accordingly. So the job is
never lost, but substantially delayed due to this. (the frequency of
TopologyEvents depends on actual cluster/property changes happening in the
topology and can thus vary).
Tasks:
* provide a test case to reproduce
* fix the race-condition
* undo
> Suspected race condition between TOPOLOGY_INIT and JobManager.addJob
> --------------------------------------------------------------------
>
> Key: SLING-12078
> URL: https://issues.apache.org/jira/browse/SLING-12078
> Project: Sling
> Issue Type: Bug
> Components: Event
> Affects Versions: Event 4.3.12
> Reporter: Stefan Egli
> Priority: Major
>
> Two regular cases where a job is stored as part of JobManager.addJob():
> * when a topology is defined, it directly gets stored to the appropriate
> assigned/target slingId subtree. This is the most frequent case by far.
> * if no topology is defined (no TOPOLOGY_INIT received) yet, it gets put
> into the unassigned subtree. Later upon receiving TOPOLOGY_INIT
> CheckTopologyTask.fullRun() finds such unassigned jobs and moves them to the
> corresponding assigned subtree.
> There is a suspect race condition (test case to be provided), which happens
> between the thread doing JobManager.addJob() and the thread handling the
> TOPOLOGY_INIT:
> * JobManager.addJob determines the target slingId - which is not yet
> defined, as TOPOLOGY_INIT is just being handled concurrently
> * CheckTopologyTask.fullRun(), as part of TOPOLOGY_INIT handling, however
> does not yet find the above new job in unassigned, as the job is just being
> stored concurrently.
> The result is a job in the unassigned subtree, which waits until the next
> TopologyEvent happens - which then invokes CheckTopologyTask.fullRun() -
> which then finds the unassigned job and re/assigns it accordingly. So the job
> is never lost, but substantially delayed due to this. (the frequency of
> TopologyEvents depends on actual cluster/property changes happening in the
> topology and can thus vary).
> Tasks:
> * provide a test case to reproduce
> * fix the race-condition
> * undo
> https://github.com/apache/sling-org-apache-sling-event/commit/d16686705908099b26d0a3233f61c4e209880f93
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
