BUG #15233: Error in estimation leads to very bad parralel plan insimple 2 table join. - Mailing list pgsql-bugs

The following bug has been logged on the website:

Bug reference:      15233
Logged by:          Maxim Boguk
Email address:      mb@dataegret.com
PostgreSQL version: 9.6.8
Operating system:   Linux Ubuntu
Description:

Original query:

SELECT * FROM resume r
INNER JOIN area a ON (r.area_id = a.area_id)
where
      (a.path LIKE '%.1806.%')
      AND r.is_finished IN(1, 2, 9)
      AND r.disabled = false
      AND r.access_type IN (1, 2, 3, 5)
      AND r.desireable_compensation_currency_code = 'RUR' AND
r.desireable_compensation >= 1 AND r.desireable_compensation <= 30000
      AND r.post NOT ILIKE '%Руководитель%' AND r.post NOT ILIKE
'%Директор%' AND r.post NOT ILIKE '%Начальник%' AND r.post NOT ILIKE
'%Заместитель руководителя%' AND r.post NOT ILIKE '%Заместитель начальника%'
AND r.post NOT ILIKE '%Вице-президент%' AND r.post NOT ILIKE '%Заместитель
директора%' AND r.post NOT ILIKE '%Управляющий%' AND r.post NOT ILIKE
'%Заместитель управляющего%'
      AND 2 =  ANY (employments);

Bad plan (normal settings): 

Nested Loop  (cost=100.00..1281887.84 rows=35 width=614) (actual
time=361757.698..2084527.877 rows=4496 loops=1)
   Join Filter: (r.area_id = a.area_id)
   Rows Removed by Join Filter: 9186603
   ->  Seq Scan on area a  (cost=0.00..265.70 rows=1 width=106) (actual
time=0.669..1.756 rows=23 loops=1)
         Filter: ((path)::text ~~ '%.1806.%'::text)
         Rows Removed by Filter: 5076
   ->  Gather  (cost=100.00..1272588.91 rows=178876 width=508) (actual
time=0.205..90588.432 rows=399613 loops=23)
         Workers Planned: 8
         Workers Launched: 8
         ->  Parallel Seq Scan on resume r  (cost=0.00..1270700.15
rows=22360 width=508) (actual time=0.692..41168.395 rows=57806 loops=159)
               Filter: ((NOT disabled) AND (desireable_compensation >= 1)
AND (desireable_compensation <= 30000) AND (post !~~*
'%Руководитель%'::text) AND (post !~~* '%Директор%'::text) AND (post !~~*
'%Начальник%'::text) AND (post !~~* '%Заместитель руководителя%'::text) AND
(post !~~* '%Заместитель начальника%'::text) AND (post !~~*
'%Вице-президент%'::text) AND (post !~~* '%Заместитель директора%'::text)
AND (post !~~* '%Управляющий%'::text) AND (post !~~* '%Заместитель
управляющего%'::text) AND (desireable_compensation_currency_code =
'RUR'::bpchar) AND (is_finished = ANY ('{1,2,9}'::integer[])) AND
(access_type = ANY ('{1,2,3,5}'::integer[])) AND (2 = ANY (employments)))
               Rows Removed by Filter: 12575133
 Planning time: 2.953 ms
 Execution time: 2084537.882 ms

Problem with estimation of selectivity 
   ->  Seq Scan on area a  (cost=0.00..265.70 rows=1 width=106) (actual
time=0.669..1.756 rows=23 loops=1)
         Filter: ((path)::text ~~ '%.1806.%'::text)
leads to very dangerous idea to perform nested loop with very heavy parallel
plan inside.

Normal plan with set enable_nestloop = 0; 
perform almost 100 times better:

 Gather  (cost=365.75..1271083.22 rows=35 width=614) (actual
time=192.640..31459.920 rows=4496 loops=1)
   Workers Planned: 8
   Workers Launched: 8
   ->  Hash Join  (cost=265.75..1270982.87 rows=4 width=614) (actual
time=107.194..31447.124 rows=500 loops=9)
         Hash Cond: (r.area_id = a.area_id)
         ->  Parallel Seq Scan on resume r  (cost=0.00..1270700.15
rows=22360 width=508) (actual time=0.854..31438.614 rows=44401 loops=9)
               Filter: ((NOT disabled) AND (desireable_compensation >= 1)
AND (desireable_compensation <= 30000) AND (post !~~*
'%Руководитель%'::text) AND (post !~~* '%Директор%'::text) AND (post !~~*
'%Начальник%'::text) AND (post !~~* '%Заместитель руководителя%'::text) AND
(post !~~* '%Заместитель начальника%'::text) AND (post !~~*
'%Вице-президент%'::text) AND (post !~~* '%Заместитель директора%'::text)
AND (post !~~* '%Управляющий%'::text) AND (post !~~* '%Заместитель
управляющего%'::text) AND (desireable_compensation_currency_code =
'RUR'::bpchar) AND (is_finished = ANY ('{1,2,9}'::integer[])) AND
(access_type = ANY ('{1,2,3,5}'::integer[])) AND (2 = ANY (employments)))
               Rows Removed by Filter: 9659153
         ->  Hash  (cost=265.70..265.70 rows=1 width=106) (actual
time=1.260..1.260 rows=23 loops=9)
               Buckets: 1024  Batches: 1  Memory Usage: 11kB
               ->  Seq Scan on area a  (cost=0.00..265.70 rows=1 width=106)
(actual time=0.546..1.250 rows=23 loops=9)
                     Filter: ((path)::text ~~ '%.1806.%'::text)
                     Rows Removed by Filter: 5076
 Planning time: 2.914 ms
 Execution time: 31469.176 ms

May be there should be some penalization plans of first kind (with nested
loop over heavy parallel plan) to take into account chance of estimation
error in leading part of nested loop in cases when estimated only 1 row.

PS: I not even knew before that the PostgreSQL can perform parallel seq scan
inside nested loop. Never seen such plans before.