PostgreSQL query is not using an index - sql

Question

Enviroment
My PostgreSQL (9.2) schema looks like this:
CREATE TABLE first
(
id_first bigint NOT NULL,
first_date timestamp without time zone NOT NULL,
CONSTRAINT first_pkey PRIMARY KEY (id_first)
)
WITH (
OIDS=FALSE
);
CREATE INDEX first_first_date_idx
ON first
USING btree
(first_date);
CREATE TABLE second
(
id_second bigint NOT NULL,
id_first bigint NOT NULL,
CONSTRAINT second_pkey PRIMARY KEY (id_second),
CONSTRAINT fk_first FOREIGN KEY (id_first)
REFERENCES first (id_first) MATCH SIMPLE
ON UPDATE NO ACTION ON DELETE NO ACTION
)
WITH (
OIDS=FALSE
);
CREATE INDEX second_id_first_idx
ON second
USING btree
(id_first);
CREATE TABLE third
(
id_third bigint NOT NULL,
id_second bigint NOT NULL,
CONSTRAINT third_pkey PRIMARY KEY (id_third),
CONSTRAINT fk_second FOREIGN KEY (id_second)
REFERENCES second (id_second) MATCH SIMPLE
ON UPDATE NO ACTION ON DELETE NO ACTION
)
WITH (
OIDS=FALSE
);
CREATE INDEX third_id_second_idx
ON third
USING btree
(id_second);
So, I have 3 tables with own PK. First has an index on first_date, Second has a FK from First and index on it. Third as a FK from Second and index on it aswell:
First (0 --> n) Second (0 --> n) Third
First table contains about 10 000 000 records.
Second table contains about 20 000 000 records.
Third table contains about 18 000 000 records.
Date range in column first_date is from 2016-01-01 till today.
random_cost_page is set to 2.0.
default_statistics_target is set to 100.
All FK, PK and first_date STATISTICS are set to 5000
Task to do
I want to count all Third rows connected with First, where first_date < X
My query:
SELECT count(t.id_third) AS count
FROM first f
JOIN second s ON s.id_first = f.id_first
JOIN third t ON t.id_second = s.id_second
WHERE first_date < _my_date
Problem description
Asking for 2 days - _my_date = '2016-01-03'
Everything working pretty well. Query lasts 1-2 seconds.
EXPLAIN ANALYZE:
"Aggregate (cost=8585512.55..8585512.56 rows=1 width=8) (actual time=67.310..67.310 rows=1 loops=1)"
" -> Merge Join (cost=4208477.49..8583088.04 rows=969805 width=8) (actual time=44.277..65.948 rows=17631 loops=1)"
" Merge Cond: (s.id_second = t.id_second)"
" -> Sort (cost=4208477.48..4211121.75 rows=1057709 width=8) (actual time=44.263..46.035 rows=19230 loops=1)"
" Sort Key: s.id_second"
" Sort Method: quicksort Memory: 1670kB"
" -> Nested Loop (cost=0.01..4092310.41 rows=1057709 width=8) (actual time=6.169..39.183 rows=19230 loops=1)"
" -> Index Scan using first_first_date_idx on first f (cost=0.01..483786.81 rows=492376 width=8) (actual time=6.159..12.223 rows=10346 loops=1)"
" Index Cond: (first_date < '2016-01-03 00:00:00'::timestamp without time zone)"
" -> Index Scan using second_id_first_idx on second s (cost=0.00..7.26 rows=7 width=16) (actual time=0.002..0.002 rows=2 loops=10346)"
" Index Cond: (id_first = f.id_first)"
" -> Index Scan using third_id_second_idx on third t (cost=0.00..4316649.89 rows=17193788 width=16) (actual time=0.008..7.293 rows=17632 loops=1)"
"Total runtime: 67.369 ms"
Asking for 10 days or more - _my_date = '2016-01-11' or more
Query is not using a indexscan anymore - replaced by seqscan and last 3-4 minutes...
Query plan:
"Aggregate (cost=8731468.75..8731468.76 rows=1 width=8) (actual time=234411.229..234411.229 rows=1 loops=1)"
" -> Hash Join (cost=4352424.81..8728697.88 rows=1108348 width=8) (actual time=189670.068..234400.540 rows=138246 loops=1)"
" Hash Cond: (t.id_second = o.id_second)"
" -> Seq Scan on third t (cost=0.00..4128080.88 rows=17193788 width=16) (actual time=0.016..124111.453 rows=17570724 loops=1)"
" -> Hash (cost=4332592.69..4332592.69 rows=1208810 width=8) (actual time=98566.740..98566.740 rows=151263 loops=1)"
" Buckets: 16384 Batches: 16 Memory Usage: 378kB"
" -> Hash Join (cost=561918.25..4332592.69 rows=1208810 width=8) (actual time=6535.801..98535.915 rows=151263 loops=1)"
" Hash Cond: (s.id_first = f.id_first)"
" -> Seq Scan on second s (cost=0.00..3432617.48 rows=18752248 width=16) (actual time=6090.771..88891.691 rows=19132869 loops=1)"
" -> Hash (cost=552685.31..552685.31 rows=562715 width=8) (actual time=444.630..444.630 rows=81650 loops=1)"
" -> Index Scan using first_first_date_idx on first f (cost=0.01..552685.31 rows=562715 width=8) (actual time=7.987..421.087 rows=81650 loops=1)"
" Index Cond: (first_date < '2016-01-13 00:00:00'::timestamp without time zone)"
"Total runtime: 234411.303 ms"
For test purposes, I have set:
SET enable_seqscan = OFF;
My queries start using indexscan again and last for 1-10 s (depends on range).
Question
Why this is working like that? How to convince a Query Planner to use a indexscan?
EDIT
After reducing a random_page_cost to 1.1, I can select about 30 days now still using a indexscan. Query plan changed a little bit:
"Aggregate (cost=8071389.47..8071389.48 rows=1 width=8) (actual time=4915.196..4915.196 rows=1 loops=1)"
" -> Nested Loop (cost=0.01..8067832.28 rows=1422878 width=8) (actual time=14.402..4866.937 rows=399184 loops=1)"
" -> Nested Loop (cost=0.01..3492321.55 rows=1551849 width=8) (actual time=14.393..3012.617 rows=436794 loops=1)"
" -> Index Scan using first_first_date_idx on first f (cost=0.01..432541.99 rows=722404 width=8) (actual time=14.372..729.233 rows=236007 loops=1)"
" Index Cond: (first_date < '2016-02-01 00:00:00'::timestamp without time zone)"
" -> Index Scan using second_id_first_idx on second s (cost=0.00..4.17 rows=7 width=16) (actual time=0.008..0.009 rows=2 loops=236007)"
" Index Cond: (second = f.id_second)"
" -> Index Scan using third_id_second_idx on third t (cost=0.00..2.94 rows=1 width=16) (actual time=0.004..0.004 rows=1 loops=436794)"
" Index Cond: (id_second = s.id_second)"
"Total runtime: 4915.254 ms"
However, I still don get it why asking for more couse a seqscan...
Iteresting is that, when I ask for range just above some kind of limit I getting a Query plan like this (here select for 40 days - asking for more will produce full seqscan again):
"Aggregate (cost=8403399.27..8403399.28 rows=1 width=8) (actual time=138303.216..138303.217 rows=1 loops=1)"
" -> Hash Join (cost=3887619.07..8399467.63 rows=1572656 width=8) (actual time=44056.443..138261.203 rows=512062 loops=1)"
" Hash Cond: (t.id_second = s.id_second)"
" -> Seq Scan on third t (cost=0.00..4128080.88 rows=17193788 width=16) (actual time=0.004..119497.056 rows=17570724 loops=1)"
" -> Hash (cost=3859478.04..3859478.04 rows=1715203 width=8) (actual time=5695.077..5695.077 rows=560503 loops=1)"
" Buckets: 16384 Batches: 16 Memory Usage: 1390kB"
" -> Nested Loop (cost=0.01..3859478.04 rows=1715203 width=8) (actual time=65.250..5533.413 rows=560503 loops=1)"
" -> Index Scan using first_first_date_idx on first f (cost=0.01..477985.28 rows=798447 width=8) (actual time=64.927..1688.341 rows=302663 loops=1)"
" Index Cond: (first_date < '2016-02-11 00:00:00'::timestamp without time zone)"
" -> Index Scan using second_id_first_idx on second s (cost=0.00..4.17 rows=7 width=16) (actual time=0.010..0.012 rows=2 loops=302663)"
" Index Cond: (id_first = f.id_first)"
"Total runtime: 138303.306 ms"
UPDATE after Laurenz Able suggestions
After rewritting a query plan as Laurenz Able suggested:
"Aggregate (cost=9102321.05..9102321.06 rows=1 width=8) (actual time=15237.830..15237.830 rows=1 loops=1)"
" -> Merge Join (cost=4578171.25..9097528.19 rows=1917143 width=8) (actual time=9111.694..15156.092 rows=803657 loops=1)"
" Merge Cond: (third.id_second = s.id_second)"
" -> Index Scan using third_id_second_idx on third (cost=0.00..4270478.19 rows=17193788 width=16) (actual time=23.650..5425.137 rows=803658 loops=1)"
" -> Materialize (cost=4577722.81..4588177.38 rows=2090914 width=8) (actual time=9088.030..9354.326 rows=879283 loops=1)"
" -> Sort (cost=4577722.81..4582950.09 rows=2090914 width=8) (actual time=9088.023..9238.426 rows=879283 loops=1)"
" Sort Key: s.id_second"
" Sort Method: external sort Disk: 15480kB"
" -> Merge Join (cost=673389.38..4341477.37 rows=2090914 width=8) (actual time=3662.239..8485.768 rows=879283 loops=1)"
" Merge Cond: (s.id_first = f.id_first)"
" -> Index Scan using second_id_first_idx on second s (cost=0.00..3587838.88 rows=18752248 width=16) (actual time=0.015..4204.308 rows=879284 loops=1)"
" -> Materialize (cost=672960.82..677827.55 rows=973345 width=8) (actual time=3662.216..3855.667 rows=892988 loops=1)"
" -> Sort (cost=672960.82..675394.19 rows=973345 width=8) (actual time=3662.213..3745.975 rows=476519 loops=1)"
" Sort Key: f.id_first"
" Sort Method: external sort Disk: 8400kB"
" -> Index Scan using first_first_date_idx on first f (cost=0.01..568352.90 rows=973345 width=8) (actual time=126.386..3233.134 rows=476519 loops=1)"
" Index Cond: (first_date < '2016-03-03 00:00:00'::timestamp without time zone)"
"Total runtime: 15244.404 ms"

Answer 1

First, it looks like some of the estimates are off.
Try to ANALYZE the tables and see if that changes the query plan chosen.
What might also help is to lower random_page_cost to a value just over 1 and see if that improves the plan.
It is interesting to note that the index scan on third_id_second_idx in the fast query produces only 17632 rows instead of over 17 million, which I can only explain by assuming that from that row on, the values of id_second no longer match any row in the join of first and second, i.e. the merge join is completed after that.
You can try to exploit that with with a rewritten query. Try
JOIN (SELECT id_second, id_third FROM third ORDER BY id_second) t
instead of
JOIN third t
That may result in a better plan since PostgreSQL won't optimize the ORDER BY away, and the planner may decide that since it has to sort third anyway, it may be cheaper to use a merge join. That way you trick the planner into choosing a plan that it wouldn't recognize as ideal. With a different value distribution the planner's original choice would probably be better.