[
https://issues.apache.org/jira/browse/NIFI-2848?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15536240#comment-15536240
]
Toivo Adams commented on NIFI-2848:
-----------------------------------
I removed Funnel and tested again.
One queue tends to dominate. This time third GenerateFlowFile queue.
After stopping third GenerateFlowFile, second started dominate.
Once queue gets opportunity it uses forever and don’t give others change to
send anything.
See dominate_1.png and dominate_2.png
> Queues aren't fairly drained when leading to a single component
> ---------------------------------------------------------------
>
> Key: NIFI-2848
> URL: https://issues.apache.org/jira/browse/NIFI-2848
> Project: Apache NiFi
> Issue Type: Improvement
> Components: Core Framework
> Affects Versions: 1.0.0, 0.7.0
> Reporter: Joseph Gresock
> Attachments: Backpressure_prioritization_test.xml, dominate_1.png,
> dominate_2.png, queue_drain.png
>
>
> Consider the scenario where multiple queues lead to a single component and
> all of them are full due to back pressure. With the attached template, it is
> easily observable that once a single queue starts to drain due to relieved
> back pressure, it will continue to drain as long as it has incoming flow
> files. This means that if there's a constant flow of incoming flow files to
> this queue, the other queues will never be drained (at least, that's my
> theory based on several hours of observation).
> To reproduce this:
> # Load the template into NiFi 1.0.0
> # Play all three GenerateFlowFile processors, but not the UpdateAttribute
> processor (this simulates backpressure). Wait until each queue has 1,000
> flow files (max backpressure)
> # Stop the GenerateFlowFile processors, and play the UpdateAttribute
> processor (this relieves the backpressure)
> # Observe which queue has started to drain, and start its GenerateFlowFile
> processor
> # Observe that the other two queues remain full indefinitely, while the
> draining queue continues to replenish and be drained indefinitely
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
