drill-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Jinfeng Ni <...@apache.org>
Subject Re: TestTpchDistributed unit tests...
Date Wed, 29 Apr 2015 17:50:10 GMT
Mehant raised a good point, that is, whether slice_target = 1 implies plans
with exchanges everywhere.  Is it worthwhile to add a new planner option to
explicitly enforce exchanges inserted during the planning phase, in stead
of relying on slice_target?

Either way, I agree that we should put verification into unit testcase, not
only the query plan, but also the query result. Currently, those testcases
do not verify either of them.



On Wed, Apr 29, 2015 at 9:33 AM, Mehant Baid <baid.mehant@gmail.com> wrote:

> I wanted to point out that a bunch of tests in TestExampleQueries rely on
> the fact that settting the slice target = 1 will generate plans with
> exchanges, which is not true any more. We need to modify those tests as
> well so we have exchanges in there, otherwise we may introduce regressions.
>
> Thanks
> Mehant
>
> On 4/29/15 9:28 AM, Aman Sinha wrote:
>
>> I noticed that the joins in  TestTpchDistributed unit tests are not
>> getting
>> parallelized.  The aggregations and sort are getting parallelized.  I have
>> filed DRILL-2900.
>> We do have test coverage for parallel joins elsewhere, especially with
>> TPCH
>> SF100 tests; so this is an FYI  in case you are relying on
>> TestTpchDistributed for such coverage.
>>
>> This also raises the importance of plan tracking for more tests.  We have
>> been doing plan tracking for partition pruning tests and it has worked out
>> quite well.
>>
>> Aman
>>
>>
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message