Query performance changes significantly depending on limit value

Hi guys

I'm a bit stuck on a query that performs fantastically up to a certain
limit value, after which the planner goes off in a completely different
direction and performance gets dramatically worse. Am using Postgresql 9.3

You can see all the relevant schemas at http://pastebin.com/PNEqw2id and in
the test database there are 1,000,000 records in contacts_contact, and
about half of those will match the subquery on values_value.

The query in question is:

SELECT "contacts_contact".* FROM "contacts_contact"
INNER JOIN "contacts_contactgroup_contacts" ON ("contacts_contact"."id" =
"contacts_contactgroup_contacts"."contact_id")
WHERE ("contacts_contactgroup_contacts"."contactgroup_id" = 1
AND "contacts_contact"."id" IN (
SELECT U0."contact_id" FROM "values_value" U0 WHERE
(U0."contact_field_id" = 1 AND UPPER(U0."string_value"::text) = UPPER('F'))
)
) ORDER BY "contacts_contact"."id" DESC LIMIT 222;

With that limit of 222, it performs like:

Limit (cost=3.09..13256.36 rows=222 width=88) (actual time=0.122..3.358
rows=222 loops=1)
Buffers: shared hit=708 read=63
-> Nested Loop (cost=3.09..59583.10 rows=998 width=88) (actual
time=0.120..3.304 rows=222 loops=1)
Buffers: shared hit=708 read=63
-> Merge Semi Join (cost=2.65..51687.89 rows=2004 width=92)
(actual time=0.103..1.968 rows=227 loops=1)
Merge Cond: (contacts_contact.id = u0.contact_id)
Buffers: shared hit=24 read=63
-> Index Scan Backward using contacts_contact_pkey on
contacts_contact (cost=0.42..41249.43 rows=1000000 width=88) (actual
time=0.008..0.502 rows=1117 loops=1)
Buffers: shared hit=22 read=2
-> Index Scan using values_value_field_string_value_contact
on values_value u0 (cost=0.43..7934.72 rows=2004 width=4) (actual
time=0.086..0.857 rows=227 loops=1)
Index Cond: ((contact_field_id = 1) AND
(upper(string_value) = 'F'::text))
Buffers: shared hit=2 read=61
-> Index Only Scan using
contacts_contactgroup_contacts_contactgroup_id_0f909f73_uniq on
contacts_contactgroup_contacts (cost=0.43..3.93 rows=1 width=4) (actual
time=0.005..0.005 rows=1 loops=227)
Index Cond: ((contactgroup_id = 1) AND (contact_id =
contacts_contact.id))
Heap Fetches: 0
Buffers: shared hit=684
Total runtime: 3.488 ms

https://explain.depesz.com/s/iPPJ

But if increase the limit to 223 then it performs like:

Limit (cost=8785.68..13306.24 rows=223 width=88) (actual
time=2685.830..2686.534 rows=223 loops=1)
Buffers: shared hit=767648 read=86530
-> Merge Join (cost=8785.68..29016.70 rows=998 width=88) (actual
time=2685.828..2686.461 rows=223 loops=1)
Merge Cond: (contacts_contact.id =
contacts_contactgroup_contacts.contact_id)
Buffers: shared hit=767648 read=86530
-> Sort (cost=8784.44..8789.45 rows=2004 width=92) (actual
time=2685.742..2685.804 rows=228 loops=1)
Sort Key: contacts_contact.id
Sort Method: quicksort Memory: 34327kB
Buffers: shared hit=767648 read=86524
-> Nested Loop (cost=6811.12..8674.53 rows=2004 width=92)
(actual time=646.573..2417.291 rows=200412 loops=1)
Buffers: shared hit=767648 read=86524
-> HashAggregate (cost=6810.70..6813.14 rows=244
width=4) (actual time=646.532..766.200 rows=200412 loops=1)
Buffers: shared read=51417
-> Bitmap Heap Scan on values_value u0
(cost=60.98..6805.69 rows=2004 width=4) (actual time=92.016..433.709
rows=200412 loops=1)
Recheck Cond: ((contact_field_id = 1) AND
(upper(string_value) = 'F'::text))
Buffers: shared read=51417
-> Bitmap Index Scan on
values_value_field_string_value_contact (cost=0.00..60.47 rows=2004
width=0) (actual time=70.647..70.647 rows=200412 loops=1)
Index Cond: ((contact_field_id = 1)
AND (upper(string_value) = 'F'::text))
Buffers: shared read=770
-> Index Scan using contacts_contact_pkey on
contacts_contact (cost=0.42..7.62 rows=1 width=88) (actual
time=0.007..0.007 rows=1 loops=200412)
Index Cond: (id = u0.contact_id)
Buffers: shared hit=767648 read=35107
-> Index Only Scan Backward using
contacts_contactgroup_contacts_contactgroup_id_0f909f73_uniq on
contacts_contactgroup_contacts (cost=0.43..18967.29 rows=497992 width=4)
(actual time=0.073..0.273 rows=550 loops=1)
Index Cond: (contactgroup_id = 1)
Heap Fetches: 0
Buffers: shared read=6
Total runtime: 2695.301 ms

https://explain.depesz.com/s/gXS

I've tried running ANALYZE but that actually reduced the limit at which
things get worse. Any insight into the reasoning of the query planner would
be much appreciated.

Thanks

--
*Rowan Seymour* | +260 964153686 | @rowanseymour