aurora-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Maxim Khutornenko <ma...@apache.org>
Subject Re: Aurora performance impact with hourly query runs
Date Thu, 09 Jun 2016 17:06:15 GMT
Scheduler persists its state in the Mesos replicated log regardless of
the in-memory engine. If you change the flag and restart scheduler all
tasks are going to be re-inserted into MemTaskStore instead of
DBTaskStore. No data will be lost.

On Thu, Jun 9, 2016 at 9:55 AM, Shyam Patel <sham.patel04@gmail.com> wrote:
> Thanks Maxim,
>
> If we move to mem task store, restart of aurora would lose the data ? (btw, I’m running
aurora in a container)
>
>
>
>> On Jun 9, 2016, at 8:37 AM, Maxim Khutornenko <maxim@apache.org> wrote:
>>
>> There are plenty of factors that may contribute towards the behavior
>> you're observing. Based on the logs though it appears you are using
>> DBTaskStore (-use_beta_db_task_store=true)? If so, you may want to
>> revert to the default in-mem task store
>> (-use_beta_db_task_store=false) as DBTaskStore is known to perform
>> subpar on large task counts. This is a known issue and we plan to
>> invest into making it faster.
>>
>> On Thu, Jun 9, 2016 at 6:58 AM, Erb, Stephan
>> <Stephan.Erb@blue-yonder.com> wrote:
>>> I am no expert here, but I would assume that slow task store operations could
result from a slow replicated log. Have you tried keeping it on an SSD? (https://github.com/apache/aurora/blob/e89521f1eebd9a5301eb02e2ed6ffebdecd54c9a/docs/operations/configuration.md#-native_log_file_path)
>>>
>>> FWIW, there was a recent RB by Maxim to reduce Master load unter task reconciliation:
https://reviews.apache.org/r/47373/diff/2#index_header
>>> ________________________________________
>>> From: Shyam Patel <sham.patel04@gmail.com>
>>> Sent: Thursday, June 9, 2016 07:48
>>> To: dev@aurora.apache.org
>>> Subject: Re: Aurora performance impact with hourly query runs
>>>
>>> Hi Bill,
>>>
>>> Cluster Set up : AWS
>>>
>>> 1 Mesos , 1 ZK , 1 Aurora instance : 4 CPU, 16G mem
>>>
>>> Aurora : Xmx 14G
>>>
>>> 100 nodes agent cluster : 40 CPU, 160G mem each
>>>
>>> 8000 Jobs, each with 2 instances. So, total ~16K containers
>>>
>>>
>>> Thanks,
>>> Sham
>>>
>>>
>>>
>>>> On Jun 8, 2016, at 9:18 PM, Bill Farner <wfarner@apache.org> wrote:
>>>>
>>>> Can you give some insight into the machine specs and JVM options used?
>>>>
>>>> Also, is it 8000 jobs or tasks?  The terms are often mixed up, but will
>>>> have a big difference here.
>>>>
>>>> On Wednesday, June 8, 2016, Shyam Patel <sham.patel04@gmail.com> wrote:
>>>>
>>>>> Hi,
>>>>>
>>>>> While running LnP testing, I’m spinning of 8K docker jobs. During the
run,
>>>>> I ran into issue where TaskStatUpdate and TaskReconciler queries taking
>>>>> real long times. During the time, Aurora is pretty much freezing and
at a
>>>>> point dying.  Also, tried the same run w/o the docker jobs and faced
the
>>>>> same issue.
>>>>>
>>>>>
>>>>> Is there a way to keep the Aurora performance intact during the query
runs
>>>>> ?
>>>>>
>>>>>
>>>>>
>>>>> Here is snipped from log :
>>>>>
>>>>>
>>>>> I0602 00:53:37.527 [TaskStatUpdaterService RUNNING, DbTaskStore:104]
Query
>>>>> took 1243517 ms: TaskQuery(owner:null, role:null, environment:null,
>>>>> jobName:null, taskIds:null, statuses:[STARTING, THROTTLED, RUNNING,
>>>>> DRAINING, ASSIGNED, KILLING, RESTARTING, PENDING, PREEMPTING],
>>>>> instanceIds:null, slaveHosts:null, jobKeys:null, offset:0, limit:0)
>>>>>
>>>>>
>>>>> I0602 00:56:54.180 [TaskReconciler-0, DbTaskStore:104] Query took 1380169
>>>>> ms: TaskQuery(owner:null, role:null, environment:null, jobName:null,
>>>>> taskIds:null, statuses:[STARTING, RUNNING, DRAINING, ASSIGNED, KILLING,
>>>>> RESTARTING, PREEMPTING], instanceIds:null, slaveHosts:null, jobKeys:null,
>>>>> offset:0, limit:0)
>>>>>
>>>>>
>>>>>
>>>>> Appreciate any insights..
>>>>>
>>>>>
>>>>> Thanks,
>>>>> Sham
>>>>>
>>>>>
>

Mime
View raw message