> On Feb. 12, 2017, 12:14 a.m., Santhosh Kumar Shanmugham wrote:
> > src/main/java/org/apache/aurora/scheduler/pruning/TaskHistoryPruner.java, 
> > line 62
> > <https://reviews.apache.org/r/56575/diff/1/?file=1630791#file1630791line62>
> >
> >     It is worthwhile to note that we are moving from a workload that was 
> > spread over a duration to a bursty instanteous workload (saw-tooth like), 
> > which can potentially make the situation worse by causing a thundering-herd 
> > at regular intervals.
> 
> Mehrdad Nurolahzade wrote:
>     That's a valid concern; testing can better clarify this.
>     
>     I agree that the existing algorithm offers a better best/average case 
> behavior (due to its scheduled pruning strategy). However, I still think the 
> worst case behavior of this implementation is better for two reasons (1) 
> every task/job is evaluated only once and (2) first prune after restart is 
> similar to other prunes and is not burstier. The burst can better be tamed by 
> reducing the pruning interval (e.g., 5 minutes).
>     
>     I believe the key to get this bursty workload under control is extending 
> `org.apache.aurora.scheduler.base.Query` abstraction. If we add something 
> like `.limit(int)` then we can control the max volume of tasks retrieved == 
> load to be processed == garbage to be collected.
> 
> Stephan Erb wrote:
>     Have you considered to use a control flow in the form of:
>     
>         for job j in all jobs:
>           retrive terminal tasks of j
>           do pruning for retrieved tasks  
>            
>     This would result in less peak memory consumption as only a small portion 
> of terminal tasks will be worked on simultaneously. If you are concerned 
> about heap pressure this may be a favorable setup.

I originally did but then dropped it fearing that it causes overhead. Now in 
hindsight, I liked this alternative better; in addition to reducing heap 
pressure it also elminates the concern regarding `MemTaskStore` full-scans.

Will submit a new patch following this design.


- Mehrdad


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/56575/#review165260
-----------------------------------------------------------


On Feb. 11, 2017, 3:12 p.m., Mehrdad Nurolahzade wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/56575/
> -----------------------------------------------------------
> 
> (Updated Feb. 11, 2017, 3:12 p.m.)
> 
> 
> Review request for Aurora, David McLaughlin, Santhosh Kumar Shanmugham, and 
> Stephan Erb.
> 
> 
> Bugs: AURORA-1837
>     https://issues.apache.org/jira/browse/AURORA-1837
> 
> 
> Repository: aurora
> 
> 
> Description
> -------
> 
> This patch addressed efficiency issues in the current implementation of 
> `TaskHistoryPruner`. The new design is similar to that of 
> `JobUpdateHistoryPruner`: (a) Instead of registering a `DelayExecutor` run 
> upon terminal task state transitions, it runs on preconfigured intervals, 
> finds all terminal state tasks that meet pruning criteria and deletes them. 
> (b) Makes the initial task history pruning delay configurable so that it does 
> not hamper scheduler upon start.
> 
> The new design addressed the following two efficiecy problems:
> 
> 1. Upon scheduler restart/failure, the in-memory state of task history 
> pruning scheduled with `DelayExecutor` is lost. `TaskHistoryPruner` learns 
> about these dead tasks upon restart when log is replayed. These expired tasks 
> are picked up by the second call to `executor.execute()` that performs job 
> level pruning immediately (i.e., without delay). Hence, most task history 
> pruning happens after scheduler restarts and can severely hamper scheduler 
> performance (or cause consecutive fail-overs on test clusters when we put 
> load test on scheduler).
> 
> 2. Expired tasks can be picked up for pruning multiple times. The 
> asynchronous nature of `BatchWorker` which used to process task deletions 
> introduces some delay between delete enqueue and delete execution. As a 
> result, tasks already queued for deletion in a previous evaluation round 
> might get picked up, evaluated and enqueued for deletion again. This is 
> evident in `tasks_pruned` metric which reflects numbers much higher than the 
> actual number of expired tasks deleted.
> 
> 
> Diffs
> -----
> 
>   src/main/java/org/apache/aurora/scheduler/base/Query.java 
> c76b365f43eb6a3b9b0b63a879b43eb04dcd8fac 
>   src/main/java/org/apache/aurora/scheduler/pruning/PruningModule.java 
> 735199ac1ccccab343c24471890aa330d6635c26 
>   src/main/java/org/apache/aurora/scheduler/pruning/TaskHistoryPruner.java 
> f77849498ff23616f1d56d133eb218f837ac3413 
>   
> src/test/java/org/apache/aurora/scheduler/pruning/TaskHistoryPrunerTest.java 
> 14e4040e0b94e96f77068b41454311fa3bf53573 
> 
> Diff: https://reviews.apache.org/r/56575/diff/
> 
> 
> Testing
> -------
> 
> Manual testing under Vagrant
> 
> 
> Thanks,
> 
> Mehrdad Nurolahzade
> 
>

Reply via email to