[
https://issues.apache.org/jira/browse/OOZIE-348?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13101901#comment-13101901
]
Hadoop QA commented on OOZIE-348:
---------------------------------
mislam77 remarked:
Queue uniqueness is already implemented. Surely it reduces the occurrence of
problem.
However, didn't eliminate that as you mentioned.
As part of concurrency control, we are re-queuing the same command with 500ms
delay at the head of the queue. In a high loaded system, the same command could
be re-queued and causes livelock like situation.
Consider an example where there are nearly 10K unique coordinator input checks.
And the maximum concurrency is 40. After first 40, all of them will get
re-queued until one command is done. This type of situation continues for
sometime.
The similar situation has created a big trouble in production.
> GH-561: Redesign oozie internal queue
> -------------------------------------
>
> Key: OOZIE-348
> URL: https://issues.apache.org/jira/browse/OOZIE-348
> Project: Oozie
> Issue Type: Bug
> Reporter: Hadoop QA
>
> We had a lot of issues related to oozie internal queue. It includes queue
> overflow as well as re-queuing the same overly used commands to avoid
> starvation. There are other situations too. This problem becomes very obvious
> in very high-load case.
> I would like to open-up the discussion to find out a better architectural
> design for longer term considering a very high-load situation.
> The following proposals are to initiate the discussion that varied from
> complete overhaul to adjusting the current design:
> 1. Implement the queue idea into DB:
> Pros: Persistence. In hot-hot or load balancing situation it useful.
> Single place of truth. Different level of ordering could be done as needed
> through SQL. Don't bother about queue size. Don't need to recreate in every
> restart -- recovery service might be less busy.
> Cons: Extra DB access overhead.
> Middle approach could be to keep a memory cache with strict conditions. The
> details could be discussed later.
> 2. Re-queuing the same commands (that is used for throttling) -- should be
> redesigned. In this case, make sure queuing happens in the *same* place --
> not at the end of queue. I know this will break the queue meaning. In this
> case, we might need to use a different data structure.
> Currently queuing the same command at the end created starvation ( live-lock)
> like situation.
> 3. Multiple queues. One for coordinator input check that is used 99% of time.
> Comments?
> Regards,
> Mohammad
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
