Slow SQL query (has_many through) - sql

Question

I am facing a slow query on our production machine.
Following Setup:
the join model (polymorphic)
# Table name: util_participants
#
# id :bigint not null, primary key
# assignable_type :string not null
# assignable_id :bigint not null
# contact_id :bigint
#
# Indexes
#
# assignable_index (assignable_id,assignable_type)
#
# Foreign Keys
#
# fk_rails_... (contact_id => client_contacts.id)
This class is for assigning contacts to different elements, like a documentation (assignable).
One query i have is
documentation.contacts
> SELECT COUNT(*) FROM "client_contacts" INNER JOIN "util_participants" ON "client_contacts"."id" = "util_participants"."contact_id" WHERE "client_contacts"."account_id" = $1 AND "util_participants"."assignable_id" = $2 AND "util_participants"."assignable_type" = $3
account is for multi-tenancy
For a specific element i have a query time of 7 seconds. But i cannot track down the cause. If i switch the account_id to another customer, i get fast execution time.
For example
SELECT count(*)
FROM client_contacts
INNER JOIN util_participants
ON client_contacts.id = util_participants.contact_id
WHERE client_contacts.account_id = 35
AND util_participants.assignable_id = 1;
is very slow.
SELECT count(*)
FROM client_contacts
INNER JOIN util_participants
ON client_contacts.id = util_participants.contact_id
WHERE client_contacts.account_id = 27
AND util_participants.assignable_id = 1;
is not but 27 has even more contact rows.
EXPLAIN of postgres results in
Aggregate (cost=3688.79..3688.80 rows=1 width=8)
-> Nested Loop (cost=0.28..3688.79 rows=1 width=0)
Join Filter: (client_contacts.id = util_participants.contact_id)
-> Index Scan using index_client_contacts_on_account_id_and_client_id on client_contacts (cost=0.28..8.30 rows=1 width=8)
Index Cond: (account_id = 35)
-> Seq Scan on util_participants (cost=0.00..3680.48 rows=1 width=8)
Filter: (assignable_id = 1)
(7 rows)
Total row count for participants is 130k and contacts 6500
Update:
Here the output for
slow one:
> EXPLAIN(ANALYZE, VERBOSE, BUFFERS) SELECT count(*) FROM client_contacts INNER JOIN util_participants ON client_contacts.id = util_participants.contact_id WHERE client_contacts.account_id = 35 AND util_participants.assignable_id = 1;
Aggregate (cost=3688.79..3688.80 rows=1 width=8) (actual time=6991.704..6991.706 rows=1 loops=1)
Output: count(*)
Buffers: shared hit=893735
-> Nested Loop (cost=0.28..3688.79 rows=1 width=0) (actual time=6991.699..6991.700 rows=0 loops=1)
Join Filter: (client_contacts.id = util_participants.contact_id)
Rows Removed by Join Filter: 428
Buffers: shared hit=893735
-> Index Scan using index_client_contacts_on_account_id_and_client_id on public.client_contacts (cost=0.28..8.30 rows=1 width=8) (actual time=0.015..1.160 rows=428 loops=1)
Output: ....
Index Cond: (client_contacts.account_id = 35)
Buffers: shared hit=71
-> Seq Scan on public.util_participants (cost=0.00..3680.48 rows=1 width=8) (actual time=0.002..16.325 rows=1 loops=428)
Output: ....
Filter: (util_participants.assignable_id = 1)
Rows Removed by Filter: 127268
Buffers: shared hit=893664
Planning Time: 0.183 ms
Execution Time: 6991.741 ms
fast one (different account_id)
EXPLAIN(ANALYZE, VERBOSE, BUFFERS) SELECT count(*) FROM client_contacts INNER JOIN util_participants ON client_contacts.id = util_participants.contact_id WHERE client_contacts.account_id = 33 AND util_participants.assignable_id = 1;
Aggregate (cost=3688.79..3688.80 rows=1 width=8) (actual time=16.882..16.884 rows=1 loops=1)
Output: count(*)
Buffers: shared hit=2088
-> Nested Loop (cost=0.28..3688.78 rows=1 width=0) (actual time=16.876..16.878 rows=0 loops=1)
Inner Unique: true
Buffers: shared hit=2088
-> Seq Scan on public.util_participants (cost=0.00..3680.48 rows=1 width=8) (actual time=0.007..16.873 rows=1 loops=1)
Output: ...
Filter: (util_participants.assignable_id = 1)
Rows Removed by Filter: 127268
Buffers: shared hit=2088
-> Index Scan using client_contacts_pkey on public.client_contacts (cost=0.28..8.30 rows=1 width=8) (actual time=0.002..0.002 rows=0 loops=1)
Output: ...
Index Cond: (client_contacts.id = util_participants.contact_id)
Filter: (client_contacts.account_id = 33)
Planning Time: 0.176 ms
Execution Time: 16.923 ms
Why the different query plans for almost the same query?