Customized Projections

Hi ,

Below is an  undocumented method  , do not use it in production without support guideline :

SELECT set_optimizer_directives('AvoidUsingProjections=proj1,proj2'); < disable projection

SELECT set_optimizer_directives('AvoidUsingProjections='); <-Enable it back 

Thanks .

You can also check the use of UseOnlyProjections=<your projection name>
like
SELECT set_optimizer_directives('UseOnlyProjections=<your projection name>'); 

Thanks Eli for the response.  (Official disclaimer -- use of these functions is only officially supported if it is under official Support guidance through MyVertica.)

A request, if I may:  We are constantly trying to improve our DBD to create optimal projections for your queries, and our Optimizer to always choose the right projection.  Usually both do.  If you are consistently able to do much better by hand, we would really appreciate it if you're willing to post an example so that we can figure out what our implementation could be doing better.

Adam

Thanks Eli & Adam for responding. First of all, using set_optimizer_directive is nothing different than querying custom projection directly. Does Vertica recommend using custom projection in query directly? [My intention is not to override Optimizers plan, just trying to understand better]. What is the flip side of it?
@Adam: I'm sure DBD is the best bet to create optimal projections . However, as mentioned in documentation, we can create optimized design by modifying or creating the design created by DBD. As i review DBD projections, i thought i should try applying some of the basic tuning principles such as GROUP HASH --> GROUP PIPELINED , Avoiding Data Re-segmentation for GROUP BY queries & Sorting by Low Cardinality to High Cardinality columns etc.


-- Sample SQL query:

Select Prod_Tier_Cd As Prod_Tier_Cd ,
Sum(Total_Store_Inventory_Qty) As Total_Store_Inventory_Qty ,
Sum(Total_Hub_Inventory_Qty) As Total_Hub_Inventory_Qty  ,
Sum(In_Transit_Hub_Unit_Qty) As In_Transit_Hub_Unit_Qty  ,
Sum(In_Transit_Store_Unit_Qty) As In_Transit_Store_Unit_Qty ,
Sum(Avg_Units_On_Hand_Qty) As Avg_Units_On_Hand_Qty ,
Sum(Prod_For_Sale_Prev_2_Week_Days) As Prod_For_Sale_Prev_2_Week_Days,
Sum(Day_1_To_7_Sales_Qty) As Day_1_To_7_Sales_Qty
FROM TABLEA
WHERE
Rptg_Dt='2013-05-13' AND
Region_Name='APAC' AND
Country_Name='APAC' AND
Prod_Categ_Cd IN  ('XYZ','ABC')
GROUP BY Prod_Tier_Cd

--DBD Recommended Projection:

CREATE PROJECTION TABLEA /*+basename(TABLEA),createtype(L)*/ 
(
 Rptg_Dt ENCODING RLE, 
 Country_Name ENCODING RLE, 
 Prod_Id ENCODING AUTO, 
 Prod_Categ_Cd ENCODING RLE, 
 Region_Name ENCODING RLE, 
 Metric_Categ_Name ENCODING AUTO, 
 Prod_Class_Name ENCODING AUTO, 
 Prod_Tier_Cd ENCODING AUTO, 
 Planner_Cd ENCODING AUTO, 
 Prod_Row_Id ENCODING AUTO, 
 Day_1_To_7_Sales_Qty ENCODING DELTARANGE_COMP 
)
AS
 SELECT TABLEA.Rptg_Dt,
        TABLEA.Country_Name,
        TABLEA.Prod_Id,
        TABLEA.Prod_Categ_Cd,
        TABLEA.Region_Name,
        TABLEA.Metric_Categ_Name,
        TABLEA.Prod_Class_Name,
        TABLEA.Prod_Tier_Cd,
        TABLEA.Planner_Cd,
        TABLEA.Prod_Row_Id,
        TABLEA.Day_1_To_7_Sales_Qty
 FROM retail_ops.TABLEA
 ORDER BY Rptg_Dt,
          Country_Name,
          Prod_Categ_Cd,
          Prod_Class_Nr,
          Prod_Row_Id,
          Row_Order_Id,
          Prod_Tier_Cd,
          Prod_Id
SEGMENTED BY hash(TABLEA.Metric_Categ_Name) ALL NODES
KSAFE 1 ;


--Custom Projection with consideration of below tuning principles:

-- ALL GROUP COLUMNS IN ORDER BY CLAUSE (GROUP PIPELINED INSTEAD OF GROUP BY HASH )
-- ALL SEGMENTED COLUMNS ARE INCLUDED IN GROUP BY (TO AVOID RUN TIME RE-SEGMENTATION)
-- STATS Collected.

CREATE PROJECTION TABLEA_t /*+basename(TABLEA),createtype(L)*/ 
(
 Rptg_Dt ENCODING RLE, 
 Country_Name ENCODING RLE, 
 Prod_Id ENCODING AUTO, 
 Prod_Categ_Cd ENCODING RLE, 
 Region_Name ENCODING RLE, 
 Metric_Categ_Name ENCODING AUTO, 
 Prod_Class_Name ENCODING AUTO, 
 Prod_Tier_Cd ENCODING AUTO, 
 Planner_Cd ENCODING AUTO, 
 Prod_Row_Id ENCODING AUTO, 
 Day_1_To_7_Sales_Qty ENCODING DELTARANGE_COMP 
)
AS
 SELECT TABLEA.Rptg_Dt,
        TABLEA.Country_Name,
        TABLEA.Prod_Id,
        TABLEA.Prod_Categ_Cd,
        TABLEA.Region_Name,
        TABLEA.Metric_Categ_Name,
        TABLEA.Prod_Class_Name,
        TABLEA.Prod_Tier_Cd,
        TABLEA.Planner_Cd,
        TABLEA.Prod_Row_Id,
        TABLEA.Day_1_To_7_Sales_Qty
 FROM retail_ops.TABLEA
 ORDER BY TABLEA.Prod_Tier_Cd,
          TABLEA.Prod_Categ_Cd,
          TABLEA.Region_Name,
          TABLEA.Country_Name,
 TABLEA.Rptg_Dt
SEGMENTED BY hash(TABLEA.Prod_Tier_Cd) ALL NODES
KSAFE 1 ;



-- Trace back in execution_engine_profiles

-- CLOCK TIME

--DBD Recommended

App_owner=> SELECT operator_name,path_id,SUM(counter_value) FROM V_monitor.execution_engine_profiles where transaction_id=63050394795945858 and statement_id=9 and counter_name='clock time (us)' GROUP BY 1,2 ORDER BY path_id ;
 operator_name | path_id |  SUM   
---------------+---------+--------
 NewEENode     |      -1 |     93
 ExprEval      |      -1 |    100
 Root          |      -1 |    144
 ExprEval      |       1 |    151
 Sort          |       1 |    476
 ExprEval      |       3 |   1591
 GroupByHash   |       3 |  17451
 ParallelUnion |       3 | 566618
 GroupByPipe   |       3 |   1219
 ExprEval      |       4 |    313
 Scan          |       4 |  12623
(11 rows)

-- CUSTOM

Time: First fetch (11 rows): 1027.724 ms. All rows formatted: 1027.777 ms
App_owner=> SELECT operator_name,path_id,SUM(counter_value) FROM V_monitor.execution_engine_profiles where transaction_id=63050394795945858 and statement_id=21 and counter_name='clock time (us)' GROUP BY 1,2 ORDER BY path_id ;
 operator_name | path_id |  SUM   
---------------+---------+--------
 NewEENode     |      -1 |     67
 ExprEval      |      -1 |     90
 Root          |      -1 |    107
 ExprEval      |       1 |    161
 Sort          |       1 |    444
 ExprEval      |       3 |    111
 GroupByPipe   |       3 |   1127
 ExprEval      |       4 |   1200
 Scan          |       4 | 104084
(9 rows)


--EXECUTION TIME (CPU)

--DBD Recommended

App_owner=> SELECT operator_name,path_id,SUM(counter_value) FROM V_monitor.execution_engine_profiles where transaction_id=63050394795945858 and statement_id=9 and counter_name='execution time (us)' GROUP BY 1,2 ORDER BY path_id ;
 operator_name | path_id |  SUM  
---------------+---------+-------
 NewEENode     |      -1 |    71
 ExprEval      |      -1 |    86
 Root          |      -1 |   140
 ExprEval      |       1 |   147
 NetworkSend   |       1 |    58
 Sort          |       1 |   442
 NetworkRecv   |       1 |   244
 ExprEval      |       3 |   577
 GroupByHash   |       3 | 13809
 NetworkRecv   |       3 |  1297
 ParallelUnion |       3 |  5797
 GroupByPipe   |       3 |  1154
 StorageUnion  |       3 |  3337
 NetworkSend   |       3 |   187
 ExprEval      |       4 |   264
 Scan          |       4 | 12505
(16 rows)


-- CUSTOM

Time: First fetch (16 rows): 972.249 ms. All rows formatted: 972.314 ms
App_owner=> SELECT operator_name,path_id,SUM(counter_value) FROM V_monitor.execution_engine_profiles where transaction_id=63050394795945858 and statement_id=21 and counter_name='execution time (us)' GROUP BY 1,2 ORDER BY path_id ;
 operator_name | path_id |  SUM  
---------------+---------+-------
 NewEENode     |      -1 |    58
 ExprEval      |      -1 |    77
 Root          |      -1 |   104
 ExprEval      |       1 |   159
 NetworkSend   |       1 |    47
 Sort          |       1 |   399
 NetworkRecv   |       1 |   256
 ExprEval      |       3 |    93
 StorageMerge  |       3 |   547
 GroupByPipe   |       3 |   536
 ExprEval      |       4 |   470
 Scan          |       4 | 72454
(12 rows)

--OBSERVATIONS:
CUSTOM has better stats than DBD recommended . However, Optimizer is showing DBD recommended/Super projection in Explain plan.

Hi Mahesh,

Thanks for the detailed reply.  Hm...  When you ran the DBD, did you provide it with this as one of its example queries?  That doesn't look to me like a design that we would produce for that query.

Thanks,
Adam

Did you ensure there were no tuple mover operations before executing each test? I would strongly encourage getting at least 5 trials per projection.

Adding up the counter values, the DBD recommended profile appears to be 35,000 counters less than your custom projection.

What we really need to see is the explain plan for the query, using the DBD projection, and your custom projection.