Given

• Postgres database
• Perform the task query: taskService.createTaskQuery().taskUnassigned().withoutCandidateGroups().withoutCandidateUsers().count()

Problem

• Consider the following query and the query plan:

  -- QUERY
  select count(distinct RES.ID_) from ACT_RU_TASK RES WHERE (1 = 1 and RES.ASSIGNEE_ IS NULL and
  (RES.ASSIGNEE_ is null and RES.ID_ not in ( select TASK_ID_ from ACT_RU_IDENTITYLINK ARIEL WHERE
  ARIEL.TYPE_ = 'candidate')))
  
  -- QUERY PLAN
  Aggregate  (cost=28712821.87..28712821.88 rows=1 width=8) (actual time=575362.672..575362.672 rows=1 loops=1)
     ->  Gather  (cost=2342.79..28712795.40 rows=10587 width=37) (actual time=293340.304..575364.177 rows=3 loops=1)
           Workers Planned: 2
           Workers Launched: 2
           ->  Parallel Bitmap Heap Scan on act_ru_task res  (cost=1342.79..28710736.70 rows=4411 width=37) (actual time=304801.026..575332.787 rows=1 loops=3)
                 Recheck Cond: ((assignee_ IS NULL) AND (assignee_ IS NULL))
                 Filter: (NOT (SubPlan 1))
                 Rows Removed by Filter: 14033
                 Heap Blocks: exact=1683
                 ->  Bitmap Index Scan on act_idx_task_assignee  (cost=0.00..1340.15 rows=21173 width=0) (actual time=116.300..116.300 rows=42364 loops=1)
                       Index Cond: ((assignee_ IS NULL) AND (assignee_ IS NULL))
                 SubPlan 1
                   ->  Materialize  (cost=0.00..6158.60 rows=139634 width=37) (actual time=0.004..27.630 rows=60595 loops=42102)
                         ->  Seq Scan on act_ru_identitylink ariel  (cost=0.00..4369.43 rows=139634 width=37) (actual time=0.003..29.409 rows=60688 loops=22205)
                               Filter: ((type_)::text = 'candidate'::text)
                               Rows Removed by Filter: 1
   Planning time: 0.632 ms
   Execution time: 575366.077 ms
  

• The query performs not well and the costs are with ~ 28,7 Mio. rather high
• NOT IN (SELECT ...) does not optimize well
  • It uses a plain sub plan which has a bad performance [1]

Solutions
1. Rewrite the query with the help of a correlated subquery: NOT EXISTS (SELECT ...)

-- REWRITTEN QUERY
select count(distinct RES.ID_) 
from ACT_RU_TASK RES
WHERE (1 = 1 and RES.ASSIGNEE_ IS NULL
			 and (RES.ASSIGNEE_ is null
					and NOT EXISTS (
			 			select 1 
			 			from ACT_RU_IDENTITYLINK I 
			 			WHERE I.TYPE_ = 'candidate'
			 			AND   I.TASK_ID_ = RES.ID_
		 			)
			 	 )
		)

-- QUERY PLAN
Aggregate  (cost=13624.60..13624.61 rows=1 width=8) (actual time=138.484..138.484 rows=1 loops=1)
   ->  Hash Anti Join  (cost=7256.58..13615.49 rows=3645 width=37) (actual time=117.066..138.453 rows=3 loops=1)
         Hash Cond: ((res.id_)::text = (i.task_id_)::text)
         ->  Seq Scan on act_ru_task res  (cost=0.00..4569.86 rows=20607 width=37) (actual time=0.007..24.543 rows=42226 loops=1)
               Filter: ((assignee_ IS NULL) AND (assignee_ IS NULL))
               Rows Removed by Filter: 44501
         ->  Hash  (cost=4388.79..4388.79 rows=141183 width=37) (actual time=72.261..72.261 rows=142190 loops=1)
               Buckets: 65536  Batches: 4  Memory Usage: 2911kB
               ->  Seq Scan on act_ru_identitylink i  (cost=0.00..4388.79 rows=141183 width=37) (actual time=0.003..31.867 rows=142190 loops=1)
                     Filter: ((type_)::text = 'candidate'::text)
                     Rows Removed by Filter: 1
 Planning time: 0.473 ms
 Execution time: 138.647 ms

2. Rewrite the query to use a LEFT JOIN with NULL check

-- REWRITTEN QUERY
SELECT COUNT(distinct RES.id_)
FROM ACT_RU_TASK RES
LEFT JOIN ACT_RU_IDENTITYLINK I
ON  I.task_id_ = RES.id_
AND I.type_ = 'candidate'
WHERE   I.id_ IS NULL
AND     RES.assignee_ IS NULL

-- QUERY PLAN
Aggregate  (cost=9856.38..9856.39 rows=1 width=8) (actual time=123.516..123.516 rows=1 loops=1)
   ->  Hash Right Join  (cost=5096.96..9856.38 rows=1 width=37) (actual time=122.414..123.481 rows=3 loops=1)
         Hash Cond: ((link.task_id_)::text = (task.id_)::text)
         Filter: (link.id_ IS NULL)
         Rows Removed by Filter: 57826
         ->  Seq Scan on act_ru_identitylink link  (cost=0.00..4388.79 rows=141183 width=74) (actual time=0.008..31.148 rows=141212 loops=1)
               Filter: ((type_)::text = 'candidate'::text)
               Rows Removed by Filter: 1
         ->  Hash  (cost=4569.86..4569.86 rows=42168 width=37) (actual time=36.236..36.236 rows=42099 loops=1)
               Buckets: 65536  Batches: 1  Memory Usage: 3349kB
               ->  Seq Scan on act_ru_task task  (cost=0.00..4569.86 rows=42168 width=37) (actual time=0.009..22.638 rows=42099 loops=1)
                     Filter: (assignee_ IS NULL)
                     Rows Removed by Filter: 44193
 Planning time: 0.829 ms
 Execution time: 123.567 ms

Thoughts

• Both solutions perform compared to the original query way better
  • The costs are reduced by the factor of ~ 2.000
• Investigate the performance of the solutions for non-postgres databases to avoid performance degradation
• The implementation effort of solution 1 might be less compared to solution 2
  • In the withoutCandidateUser/Group case a LEFT JOIN is needed
  • The query taskService.createTaskQuery().withCandidateUser().withoutCandidateGroup().or().priority(5).dueDate(now).endOr().count() would need an additional LEFT JOIN

[1] https://wiki.postgresql.org/wiki/Don't_Do_This#Don.27t_use_NOT_IN