Wrong plan with extra parallel workers

Hi,

Having a strange situation, where adding extra parallel workers
(max_parallel_workers_per_gather), the planner chooses a different plan,
with nested loops, which makes the query twice as slow. Strangely with the
COUNT_DISTINCT implementation from Tomas Vondra (
https://github.com/tvondra/count_distinct) it scales nicely (almost
linearly) with the workers.

Can provide the DDL's or more info if needed. Any ideas why this happens?

set max_parallel_workers_per_gather = 2;

EXPLAIN ANALYZE
SELECT
"f_zendesktickets_aaeljtllr5at3el"."via_id" AS "a_1303",
COUNT( DISTINCT "f_zendesktags_aakrjpgq72ad93i"."ticket_id_id" ) AS
"c_fb839a9bd6f2015f"
FROM
"f_zendesktags_aakrjpgq72ad93i"
INNER JOIN
"f_zendesktickets_aaeljtllr5at3el"
ON "f_zendesktags_aakrjpgq72ad93i"."ticket_id_id" =
"f_zendesktickets_aaeljtllr5at3el"."id"
GROUP BY
1 ;

GroupAggregate (cost=31676816.72..32928717.16 rows=8 width=12) (actual
time=124072.467..210958.329 rows=9 loops=1)
Group Key: f_zendesktickets_aaeljtllr5at3el.via_id
-> Sort (cost=31676816.72..32094116.84 rows=166920048 width=8) (actual
time=124072.431..148808.161 rows=166920891 loops=1)
Sort Key: f_zendesktickets_aaeljtllr5at3el.via_id
Sort Method: external merge Disk: 2939944kB
-> Hash Join (cost=1919106.06..6597948.02 rows=166920048
width=8) (actual time=13063.794..85782.027 rows=166920891 loops=1)
Hash Cond: (f_zendesktags_aakrjpgq72ad93i.ticket_id_id =
f_zendesktickets_aaeljtllr5at3el.id)
-> Seq Scan on f_zendesktags_aakrjpgq72ad93i
(cost=0.00..2571476.48 rows=166920048 width=4) (actual
time=0.016..20886.829 rows=166920891 loo
ps=1)
-> Hash (cost=1774033.25..1774033.25 rows=11605825
width=8) (actual time=12981.920..12981.920 rows=11605822 loops=1)
Buckets: 16777216 Batches: 1 Memory Usage: 584425kB
-> Seq Scan on f_zendesktickets_aaeljtllr5at3el
(cost=0.00..1774033.25 rows=11605825 width=8) (actual time=0.045..9262.223
rows=116058
22 loops=1)
Planning Time: 1.426 ms
Execution Time: 211441.893 ms
(13 rows)

---------------------------------

set max_parallel_workers_per_gather = 6;

EXPLAIN ANALYZE
SELECT
"f_zendesktickets_aaeljtllr5at3el"."via_id" AS "a_1303",
COUNT( DISTINCT "f_zendesktags_aakrjpgq72ad93i"."ticket_id_id" ) AS
"c_fb839a9bd6f2015f"
FROM
"f_zendesktags_aakrjpgq72ad93i"
INNER JOIN
"f_zendesktickets_aaeljtllr5at3el"
ON "f_zendesktags_aakrjpgq72ad93i"."ticket_id_id" =
"f_zendesktickets_aaeljtllr5at3el"."id"
GROUP BY
1 ;

GroupAggregate (cost=1001.10..31949141.56 rows=8 width=12) (actual
time=9.125..399880.451 rows=9 loops=1)
Group Key: f_zendesktickets_aaeljtllr5at3el.via_id
-> Gather Merge (cost=1001.10..31114541.24 rows=166920048 width=8)
(actual time=9.037..322148.967 rows=166920891 loops=1)
Workers Planned: 6
Workers Launched: 6
-> Nested Loop (cost=1.01..10826616.89 rows=27820008 width=8)
(actual time=0.150..30867.494 rows=23845842 loops=7)
-> Parallel Index Scan using
f_zendesktickets_aaeljtllr5at3el_via_id_idx on
f_zendesktickets_aaeljtllr5at3el (cost=0.43..5838596.19 rows=193
4304 width=8) (actual time=0.066..3750.816 rows=1657975 loops=7)
-> Index Only Scan using
f_zendesktags_aakrjpgq72ad93i_ticket_id_id_idx on
f_zendesktags_aakrjpgq72ad93i (cost=0.57..2.02 rows=56 width=4) (
actual time=0.005..0.012 rows=14 loops=11605822)
Index Cond: (ticket_id_id =
f_zendesktickets_aaeljtllr5at3el.id)
Heap Fetches: 166920891
Planning Time: 1.395 ms
Execution Time: 400283.994 ms

--------------------------

Strangely with the count_distinct implementation from Tomas Vondra from
https://github.com/tvondra/count_distinct, it doesn' happen, and it scales
nicely with the extra workers.
set max_parallel_workers_per_gather = 6;

EXPLAIN ANALYZE
SELECT
"f_zendesktickets_aaeljtllr5at3el"."via_id" AS "a_1303",
COUNT_DISTINCT ( "f_zendesktags_aakrjpgq72ad93i"."ticket_id_id" ) AS
"c_fb839a9bd6f2015f"
FROM
"f_zendesktags_aakrjpgq72ad93i"
INNER JOIN
"f_zendesktickets_aaeljtllr5at3el"
ON "f_zendesktags_aakrjpgq72ad93i"."ticket_id_id" =
"f_zendesktickets_aaeljtllr5at3el"."id"
GROUP BY
1 ;

Finalize GroupAggregate (cost=6674091.66..6882748.12 rows=8 width=12)
(actual time=24724.265..44768.274 rows=9 loops=1)
Group Key: f_zendesktickets_aaeljtllr5at3el.via_id
-> Gather Merge (cost=6674091.66..6882747.66 rows=48 width=36) (actual
time=22655.677..42954.758 rows=60 loops=1)
Workers Planned: 6
Workers Launched: 6
-> Partial GroupAggregate (cost=6673091.57..6881741.73 rows=8
width=36) (actual time=21427.218..32385.322 rows=9 loops=7)
Group Key: f_zendesktickets_aaeljtllr5at3el.via_id
-> Sort (cost=6673091.57..6742641.59 rows=27820008
width=8) (actual time=20546.722..22817.024 rows=23845842 loops=7)
Sort Key: f_zendesktickets_aaeljtllr5at3el.via_id
Sort Method: quicksort Memory: 1851040kB
Worker 0: Sort Method: quicksort Memory: 1814142kB
Worker 1: Sort Method: quicksort Memory: 1806328kB
Worker 2: Sort Method: quicksort Memory: 1814436kB
Worker 3: Sort Method: quicksort Memory: 1799937kB
Worker 4: Sort Method: quicksort Memory: 1816058kB
Worker 5: Sort Method: quicksort Memory: 1815833kB
-> Parallel Hash Join (cost=1701496.84..3233200.50
rows=27820008 width=8) (actual time=3094.046..15445.013 rows=23845842
loops=7)
Hash Cond:
(f_zendesktags_aakrjpgq72ad93i.ticket_id_id =
f_zendesktickets_aaeljtllr5at3el.id)
-> Parallel Seq Scan on
f_zendesktags_aakrjpgq72ad93i (cost=0.00..1180476.08 rows=27820008
width=4) (actual time=0.014..3673.446
rows=23845842 loops=7)
-> Parallel Hash (cost=1677318.04..1677318.04
rows=1934304 width=8) (actual time=3078.160..3078.160 rows=1657975 loops=7)
Buckets: 16777216 Batches: 1 Memory
Usage: 585248kB
-> Parallel Seq Scan on
f_zendesktickets_aaeljtllr5at3el (cost=0.00..1677318.04 rows=1934304
width=8) (actual time=0.067..
2171.773 rows=1657975 loops=7)
Planning Time: 1.386 ms
Execution Time: 45340.324 ms
(24 rows)