Partition by date range PostgreSQL scans all partitions

Question

I have a table partitioned per month (timestamp column).

when querying the data, explain shows that all partitions are being queried when I'm constructing a date with date functions, whereas when I use hard coded dates only the targeted partitions are being scanned.

So when querying like this:

SELECT * FROM vw_comments 
      WHERE created >= '2019-4-1'
        AND created <= '2019-4-30'
limit 100;

it only scans 1 partition (1 month, good!) but then to make it more dynamic I'm passing in something like this (simplified)

SELECT * FROM vw_comments 
      WHERE created >= (date_trunc('month', now()))::timestamp
        AND created <= (date_trunc('month', now() + interval '1 month') - interval '1 day') ::timestamp
limit 100;

the exact same dates come out of the above date methods as the first query, but EXPLAIN shows all partitions get scanned.

How to make it work?

edit: add table definition and explain

upon request from @a_horse_with_no_name, I added the actual table and explain. Upon doing that I figured out something more: dynamic dates don't work when doing a join. So leaving out the 'users' table in the below query makes dynamic dates work.

CREATE TABLE public.comments
(
    comment_id integer NOT NULL GENERATED BY DEFAULT AS IDENTITY ( INCREMENT 1 START 1 MINVALUE 1 MAXVALUE 2147483647 CACHE 1 ),
    comment_id integer NOT NULL GENERATED BY DEFAULT AS IDENTITY,
    from_user_id integer NOT NULL,
    fk_topic_id integer NOT NULL,
    comment_text text COLLATE pg_catalog."default",
    parent_comment_id integer,
    created timestamp without time zone NOT NULL,
    comment_type integer NOT NULL DEFAULT 0,
    CONSTRAINT comments_pkey PRIMARY KEY (comment_id, created)
) PARTITION BY RANGE (created) 
WITH (
    OIDS = FALSE
)
TABLESPACE pg_default;

ALTER TABLE public.comments
    OWNER to soichat;

CREATE INDEX ix_comments_comment_id
    ON public.comments USING btree
    (comment_id DESC)
    TABLESPACE pg_default;

CREATE INDEX ix_comments_created
    ON public.comments USING btree
    (created DESC)
    TABLESPACE pg_default;

CREATE INDEX ix_comments_fk_topic_id
    ON public.comments USING btree
    (fk_topic_id)
    TABLESPACE pg_default;

CREATE INDEX ix_comments_from_user_id
    ON public.comments USING btree
    (from_user_id)
    TABLESPACE pg_default;

CREATE INDEX ix_comments_parent_comment_id
    ON public.comments USING btree
    (parent_comment_id)
    TABLESPACE pg_default;

-- Partitions SQL

CREATE TABLE public.comments_2019_2 PARTITION OF public.ix_comments_parent_comment_id
    FOR VALUES FROM ('2019-02-01 00:00:00') TO ('2019-03-01 00:00:00');

CREATE TABLE public.comments_2019_3 PARTITION OF public.ix_comments_parent_comment_id
    FOR VALUES FROM ('2019-03-01 00:00:00') TO ('2019-04-01 00:00:00');

CREATE TABLE public.comments_2019_4 PARTITION OF public.ix_comments_parent_comment_id
    FOR VALUES FROM ('2019-04-01 00:00:00') TO ('2019-05-01 00:00:00');

CREATE TABLE public.comments_2019_5 PARTITION OF public.ix_comments_parent_comment_id
    FOR VALUES FROM ('2019-05-01 00:00:00') TO ('2019-06-01 00:00:00');

the query:

explain (analyse, buffers)
 SELECT comments.comment_id,
    comments.from_user_id,
    comments.fk_topic_id,
    comments.comment_text,
    comments.parent_comment_id,
    comments.created,
    users.user_name,
    users.picture_path
   FROM comments
     LEFT JOIN users ON comments.from_user_id = users.user_id
    WHERE comments.created >= (date_trunc('month', now()))::timestamp
        AND comments.created <= (date_trunc('month', now() + interval '1 month') - interval '1 day') ::timestamp
limit 100;

explain (analyze, buffers)

Limit  (cost=1.20..11.93 rows=100 width=126) (actual time=1.441..1.865 rows=100 loops=1)
  Buffers: shared hit=499
  ->  Merge Left Join  (cost=1.20..753901.07 rows=7028011 width=126) (actual time=1.440..1.778 rows=100 loops=1)
        Merge Cond: (comments_2019_2.from_user_id = users.user_id)
        Buffers: shared hit=499
        ->  Merge Append  (cost=0.92..665812.08 rows=7028011 width=51) (actual time=0.017..0.259 rows=100 loops=1)
              Sort Key: comments_2019_2.from_user_id
              Buffers: shared hit=15
              ->  Index Scan using comments_2019_2_from_user_id_idx on comments_2019_2  (cost=0.15..58.95 rows=5 width=56) (actual time=0.002..0.003 rows=0 loops=1)
                    Filter: ((created >= (date_trunc('month'::text, now()))::timestamp without time zone) AND (created <= ((date_trunc('month'::text, (now() + '1 mon'::interval)) - '1 day'::interval))::timestamp without time zone))
                    Buffers: shared hit=1
              ->  Index Scan using comments_2019_3_from_user_id_idx on comments_2019_3  (cost=0.15..9790.24 rows=1 width=51) (actual time=0.002..0.003 rows=0 loops=1)
                    Filter: ((created >= (date_trunc('month'::text, now()))::timestamp without time zone) AND (created <= ((date_trunc('month'::text, (now() + '1 mon'::interval)) - '1 day'::interval))::timestamp without time zone))
                    Buffers: shared hit=1
              ->  Index Scan using comments_2019_4_from_user_id_idx on comments_2019_4  (cost=0.43..550483.74 rows=7028000 width=51) (actual time=0.010..0.162 rows=100 loops=1)
                    Filter: ((created >= (date_trunc('month'::text, now()))::timestamp without time zone) AND (created <= ((date_trunc('month'::text, (now() + '1 mon'::interval)) - '1 day'::interval))::timestamp without time zone))
                    Buffers: shared hit=12
              ->  Index Scan using comments_2019_5_from_user_id_idx on comments_2019_5  (cost=0.15..58.95 rows=5 width=56) (actual time=0.001..0.002 rows=0 loops=1)
                    Filter: ((created >= (date_trunc('month'::text, now()))::timestamp without time zone) AND (created <= ((date_trunc('month'::text, (now() + '1 mon'::interval)) - '1 day'::interval))::timestamp without time zone))
                    Buffers: shared hit=1
        ->  Index Scan using pk_users on users  (cost=0.28..234.83 rows=1606 width=79) (actual time=0.005..0.870 rows=1395 loops=1)
              Buffers: shared hit=484
Planning Time: 0.360 ms
Execution Time: 1.942 ms

@a_horse_with_no_name I added the actual query and explain. Upon doing that I figured out something more: dynamic dates don't work when doing a join. So leaving out the 'users' table in the below query makes dynamic dates work. — Elger Mensonides, Apr 11 '19 at 14:29
@a_horse_with_no_name: Hmm, I used dbeaver to do the create script. I updated the question with the correct create script from pgadmin. Sorry about that. — Elger Mensonides, Apr 11 '19 at 14:37
Looks like this case is too complicated for PostgreSQL to do partition pruning; I don't know the exact cause. But the index scans on the other partitions don't really hurt, do they? — Laurenz Albe, Apr 11 '19 at 14:47
@LaurenzAlbe it seems so, but you're seeing the results of a test db (with almost no records in the other months), usually there are millions of comments per month, we're deleting them afer 1 year, so then it would kind of hurt. Partitioning hasn't currently been implemented in production yet. — Elger Mensonides, Apr 11 '19 at 14:50

score 13 · Answer 1 · answered Apr 12 '19 at 09:50

Found (a great) answer here

Because the planner cannot know what time now() will produce at run-time it goes for the safe option and scan all partitions. Because I don't want to be configuring new functions for every partition I went for an immutable function that creates a date:

CREATE FUNCTION now_immutable()
  RETURNS timestamp AS
$func$
SELECT now() AT TIME ZONE current_setting('TimeZone')
$func$  LANGUAGE sql IMMUTABLE;

so now instead of using now(), I just use this function for functions where the date doesn't change in a transaction:

explain (analyse, buffers)
 SELECT comments.comment_id,
    comments.from_user_id,
    comments.fk_topic_id,
    comments.comment_text,
    comments.parent_comment_id,
    comments.created,
    users.user_name,
    users.picture_path
   FROM comments
     LEFT JOIN users ON comments.from_user_id = users.user_id
    WHERE comments.created >= (date_trunc('month', now_immutable()))
        AND comments.created <= (date_trunc('month', now_immutable() + interval '1 month') - interval '1 day') 
limit 100;

I also created another handy function to call from code months_back:

    CREATE OR REPLACE FUNCTION public.months_back(months_back integer)
     RETURNS timestamp without time zone
     LANGUAGE sql
     IMMUTABLE
    AS $function$
        SELECT cast((date_trunc('month', now()) - (months_back || ' month')::interval)::timestamp AT TIME ZONE current_setting('TimeZone') as timestamp)  
    $function$;

This one is handy when partitioning monthly because you can just call months_back(3) if you know the first comment was 3 months ago and postgres will search only 3 partitions, passing 0 will give you the start of the current month.

Hope this helps someone.

Partition by date range PostgreSQL scans all partitions

1 Answers1