Queuing queries rather killing

[Fawze Abujaber <fa...@gmail.com>]

Hi Guys,

I just upgraded my CDH version to include impala 2.11, and I applied the
impala admission control, I have few resource pools with the default
configuration.

I see when one query consuming all the impala resources, and new query is
submitted, the query is killed because no free memory to assign to the new
query.

How I can enforce the new query to queued instead of being killed.

Re: Queuing queries rather killing

[Fawze Abujaber <fa...@gmail.com>]

Hi Tim,

Since I just upgraded the cluster, I left all parameters in the resource
pool as the default to be able learn the usage trend, but when I saw that
the queries I killed, then i applied the default memory for query and the
max running query, and this was done for all resource pools.

There quick  questions:

1- according to the documentation the best way is to set the max memory and
default memory per query and let the impala determines that the max running
query, in my side I setup the max running queries per pool and the default
memory, will this enforce the queries without resources to be queued or I
have to set the max memory also? Which I think will make things more
complicated.

2- The case i described was for queries from different pools, for example a
query from the default pool consumed around 34 GB per node and the new
query from the other pool asked for 2 GB but since the impala Damon memory
limit is 35 GB, the other resource pool query failed, is setting the memory
configs in this case will enforce the other resource pool query to queued
even the consumption was from other pools?

3- is there any benchmark testing that shows the relationship between HDFS
scan data, number of joins in the query and the memory used per node? My
problem here that I’m in learning phase of my resource pools and I want to
avoid business queries to fail.

On Wed, 21 Feb 2018 at 21:37 Tim Armstrong <ta...@cloudera.com> wrote:

> Do you have memory-based admission control configured? If you set that up
> with default memory limits then it should help to avoid that.
>
> See
> https://impala.apache.org/docs/build/html/topics/impala_admission.html#admission_memory
> or
> https://www.cloudera.com/documentation/enterprise/latest/topics/impala_admission.html#admission_memory
>
> On Wed, Feb 21, 2018 at 10:45 AM, Fawze Abujaber <fa...@gmail.com>
> wrote:
>
>> Hi Guys,
>>
>> I just upgraded my CDH version to include impala 2.11, and I applied the
>> impala admission control, I have few resource pools with the default
>> configuration.
>>
>> I see when one query consuming all the impala resources, and new query is
>> submitted, the query is killed because no free memory to assign to the new
>> query.
>>
>> How I can enforce the new query to queued instead of being killed.
>>
>
>

Re: Queuing queries rather killing

[Tim Armstrong <ta...@cloudera.com>]

Do you have memory-based admission control configured? If you set that up
with default memory limits then it should help to avoid that.

See
https://impala.apache.org/docs/build/html/topics/impala_admission.html#admission_memory
or
https://www.cloudera.com/documentation/enterprise/latest/topics/impala_admission.html#admission_memory

On Wed, Feb 21, 2018 at 10:45 AM, Fawze Abujaber <fa...@gmail.com> wrote:

> Hi Guys,
>
> I just upgraded my CDH version to include impala 2.11, and I applied the
> impala admission control, I have few resource pools with the default
> configuration.
>
> I see when one query consuming all the impala resources, and new query is
> submitted, the query is killed because no free memory to assign to the new
> query.
>
> How I can enforce the new query to queued instead of being killed.
>