PostgreSQL: Is there a way to improve performance of SELECT queries using JSONB or HSTORE keys?

Question

I have a large table with many rows (millions) with a column of type JSONB / HSTORE , which contains many fields (hundreds). For illustration, I use the following smaller and less complex table:

-- table with HSTORE column
CREATE TABLE test_hstore (id BIGSERIAL PRIMARY KEY, data HSTORE);
INSERT INTO test_hstore (data)
SELECT hstore(
    '  key_1=>' || trunc(2 * random()) ||
    ', key_2=>' || trunc(2 * random()) ||
    ', key_3=>' || trunc(2 * random()))
FROM generate_series(0, 9999999) i;

-- table with JSONB column
CREATE TABLE test_jsonb (id BIGSERIAL PRIMARY KEY, data JSONB);
INSERT INTO test_jsonb (data)
SELECT (
    '{ "key_1":' || trunc(2 * random()) ||
    ', "key_2":' || trunc(2 * random()) ||
    ', "key_3":' || trunc(2 * random()) || '}')::JSONB
FROM generate_series(0, 9999999) i;

I would like to simply SELECT one or more fields within the data column without using a WHERE clause. I get a decrease in performance with an increasing number of selected fields:

EXPLAIN ANALYSE
SELECT id FROM test_hstore;
--Seq Scan on test_hstore  (cost=0.00..213637.56 rows=10000056 width=8) (actual time=0.049..3705.852 rows=10000000 loops=1)
--Planning time: 0.419 ms
--Execution time: 5445.654 ms

EXPLAIN ANALYSE
SELECT data FROM test_hstore;
--Seq Scan on test_hstore  (cost=0.00..213637.56 rows=10000056 width=56) (actual time=0.083..2424.334 rows=10000000 loops=1)
--Planning time: 0.082 ms
--Execution time: 3856.972 ms

EXPLAIN ANALYSE
SELECT data->'key_1' FROM test_hstore;
--Seq Scan on test_hstore  (cost=0.00..238637.70 rows=10000056 width=32) (actual time=0.122..3263.937 rows=10000000 loops=1)
--Planning time: 0.052 ms
--Execution time: 5390.803 ms


EXPLAIN ANALYSE
SELECT data->'key_1', data->'key_2' FROM test_hstore;
--Seq Scan on test_hstore  (cost=0.00..263637.84 rows=10000056 width=64) (actual time=0.089..3621.768 rows=10000000 loops=1)
--Planning time: 0.051 ms
--Execution time: 5334.452 ms

EXPLAIN ANALYSE
SELECT data->'key_1', data->'key_2', data->'key_3' FROM test_hstore;
--Seq Scan on test_hstore  (cost=0.00..288637.98 rows=10000056 width=96) (actual time=0.086..4291.111 rows=10000000 loops=1)
--Planning time: 0.067 ms
--Execution time: 6375.229 ms

Same trend (even more pronounced) for JSONB column type:

EXPLAIN ANALYSE
SELECT id FROM test_jsonb;
--Seq Scan on test_jsonb  (cost=0.00..233332.28 rows=9999828 width=8) (actual time=0.028..4009.841 rows=10000000 loops=1)
--Planning time: 0.878 ms
--Execution time: 5867.604 ms

EXPLAIN ANALYSE
SELECT data FROM test_jsonb;
--Seq Scan on test_jsonb  (cost=0.00..233332.28 rows=9999828 width=68) (actual time=0.074..2371.212 rows=10000000 loops=1)
--Planning time: 0.061 ms
--Execution time: 3787.308 ms

EXPLAIN ANALYSE
SELECT data->'key_1' FROM test_jsonb;
--Seq Scan on test_jsonb  (cost=0.00..258331.85 rows=9999828 width=32) (actual time=0.106..4677.026 rows=10000000 loops=1)
--Planning time: 0.066 ms
--Execution time: 6382.469 ms

EXPLAIN ANALYSE
SELECT data->'key_1', data->'key_2' FROM test_jsonb;
--Seq Scan on test_jsonb  (cost=0.00..283331.42 rows=9999828 width=64) (actual time=0.094..6888.904 rows=10000000 loops=1)
--Planning time: 0.047 ms
--Execution time: 8593.060 ms

EXPLAIN ANALYSE
SELECT data->'key_1', data->'key_2', data->'key_3' FROM test_jsonb;
--Seq Scan on test_jsonb  (cost=0.00..308330.99 rows=9999828 width=96) (actual time=0.173..9567.699 rows=10000000 loops=1)
--Planning time: 0.171 ms
--Execution time: 11262.135 ms

This becomes even more pronounced when the table contains many more fields. Is there a workaround?

Adding a GIN INDEX doesn't seem to help:

CREATE INDEX ix_test_hstore ON test_hstore USING GIN (data);
EXPLAIN ANALYSE
SELECT data->'key_1', data->'key_2', data->'key_3' FROM test_hstore;
--Seq Scan on test_hstore  (cost=0.00..288637.00 rows=10000000 width=96) (actual time=0.045..4650.447 rows=10000000 loops=1)
--Planning time: 2.100 ms
--Execution time: 6746.631 ms

CREATE INDEX ix_test_jsonb ON test_jsonb USING GIN (data);
EXPLAIN ANALYSE
SELECT data->'key_1', data->'key_2', data->'key_3' FROM test_jsonb;
--Seq Scan on test_jsonb  (cost=0.00..308334.00 rows=10000000 width=96) (actual time=0.149..9807.012 rows=10000000 loops=1)
--Planning time: 0.131 ms
--Execution time: 11739.948 ms

Answer 1

There's actually not much you can do to improve access to one key within a data store, or a property of a JSON piece of data (which could be an array , or a string or number ; which might be the reason why retrieving it is more difficult than retrieving it from an hstore ).

An index could help you if you need to use data->key_1 in a WHERE clause, but it will not make it any easier to retrieve the property from the data.

The best course of action, if you always (or frequently) use a certain key_1 , would be to normalise your data and make a column named key_1 . If your data source makes it very easy for you to store data , but not so easy to store key_1 , you could have a trigger function take care (on insert or update) to populate the column key_1 from the value of data :

CREATE TABLE test_jsonb 
(
    id BIGSERIAL PRIMARY KEY, 
    data JSONB, 
    key_1 integer
);

CREATE OR REPLACE FUNCTION ins_upd_test_data() 
RETURNS trigger AS
$$
BEGIN
    new.key_1 = (new.data->>'key_1')::integer ;
    RETURN new ;
END ;
$$
LANGUAGE plpgsql VOLATILE LEAKPROOF;

CREATE TRIGGER ins_upd_test_jsonb_trigger 
    BEFORE INSERT OR UPDATE OF data
    ON test_jsonb FOR EACH ROW
    EXECUTE PROCEDURE ins_upd_test_data();

This way, you can retrieve key_1 with the same efficiency that you can retrieve id .