30x size used by index vs tables about pgvector

[ncoder]

FYI. i used this command to see disk usage:

SELECT pg_relation_filepath(oid), relpages*8 as kb, relname FROM pg_class order by kb desc;

[ankane]

Hey @ncoder, use pg_table_size to get the table size.

[ankane]

fwiw, test script:

CREATE TABLE items (embedding vector(1500));
INSERT INTO items (embedding)
    SELECT (
        SELECT array_agg(i) FROM generate_series(1, 1500) i
    ) FROM generate_series(1, 100000) n;
SELECT pg_size_pretty(pg_table_size('items')) AS table_size;
SET maintenance_work_mem = '500MB';
CREATE INDEX my_index ON items USING ivfflat (embedding) WITH (lists = 1000);
SELECT pg_size_pretty(pg_total_relation_size('my_index')) AS index_size;

and output:

CREATE TABLE
INSERT 0 100000
 table_size 
------------
 795 MB
(1 row)

SET
CREATE INDEX
 index_size 
------------
 797 MB
(1 row)

[ncoder]

db=# SELECT pg_size_pretty(pg_table_size('qa')) as pg_table_size, pg_size_pretty(pg_total_relation_size('qa')) as pg_total_relation_size;
 pg_table_size | pg_total_relation_size
---------------+------------------------
 25 GB         | 48 GB

[ncoder]

wait, i made a second index to test on this one... hold up.

(Good, now Edited...)

[ncoder]

Using your script, I replicate your results exactly.

corrected my query to be equivalent on my data:

SELECT pg_size_pretty(pg_table_size('qa')) as pg_table_size, pg_size_pretty(pg_total_relation_size('qa_embedding_idx')) as pg_total_relation_size;
 pg_table_size | pg_total_relation_size
---------------+------------------------
 25 GB         | 24 GB
(1 row)

[ncoder]

So I accept that nothing is wrong, except my understanding of this page:
https://www.postgresql.org/docs/current/disk-usage.html

given I have 3 million rows, vectors alone should take 18GB, so pg_table_size() is in the right ballpark

Thanks for the pointers, @ankane.