Diff in idx_tup_read and idx_tup_fetch in postgres pg_stat_all_indexes

Question

I have gone through documentation and previous stackoverflow questions but not findings those helpful. Postgres version 10.6

Even for index only scans values of idx_tup_fetch is increasing. So I am not able to understand difference between idx_tup_read and idx_tup_fetch.

Initially understanding was: idx_tup_read: matching records read from index.

idx_tup_fetch: live rows fetched from table after records are matched on index. So my understanding is that for index only scan since no data is fetched from table so this should not increment.

But when I ran the tests then this understanding is not true.

postgres=> select * from pg_stat_all_indexes where indexrelname like 'test_monika_rank';
 relid | indexrelid | schemaname |   relname   |   indexrelname   | idx_scan | idx_tup_read | idx_tup_fetch 
-------+------------+------------+-------------+------------------+----------+--------------+---------------
 16413 |      16418 | public     | test_monika | test_monika_rank |        4 |            2 |             2
(1 row)

postgres=> select * from test_monika where rank=100;
 id | name | rank 
----+------+------
 95 |      |  100
(1 row)

postgres=> select * from pg_stat_all_indexes where indexrelname like 'test_monika_rank';
 relid | indexrelid | schemaname |   relname   |   indexrelname   | idx_scan | idx_tup_read | idx_tup_fetch 
-------+------------+------------+-------------+------------------+----------+--------------+---------------
 16413 |      16418 | public     | test_monika | test_monika_rank |        5 |            3 |             3
(1 row)

postgres=> select rank from test_monika where rank=100;
 rank 
------
  100
(1 row)

postgres=> select * from pg_stat_all_indexes where indexrelname like 'test_monika_rank';
 relid | indexrelid | schemaname |   relname   |   indexrelname   | idx_scan | idx_tup_read | idx_tup_fetch 
-------+------------+------------+-------------+------------------+----------+--------------+---------------
 16413 |      16418 | public     | test_monika | test_monika_rank |        6 |            4 |             4
(1 row)

postgres=> select id from test_monika where rank=100;
 id 
----
 95
(1 row)

postgres=> select * from pg_stat_all_indexes where indexrelname like 'test_monika_rank';
 relid | indexrelid | schemaname |   relname   |   indexrelname   | idx_scan | idx_tup_read | idx_tup_fetch 
-------+------------+------------+-------------+------------------+----------+--------------+---------------
 16413 |      16418 | public     | test_monika | test_monika_rank |        7 |            5 |             5
(1 row)

postgres=> select 1 from test_monika where rank=100;
 ?column? 
----------
        1
(1 row)

postgres=> select * from pg_stat_all_indexes where indexrelname like 'test_monika_rank';
 relid | indexrelid | schemaname |   relname   |   indexrelname   | idx_scan | idx_tup_read | idx_tup_fetch 
-------+------------+------------+-------------+------------------+----------+--------------+---------------
 16413 |      16418 | public     | test_monika | test_monika_rank |        8 |            6 |             6
(1 row)

Answer 1

idx_tup_read is increased whenever an index entry is read.

idx_tup_fetch is increased whenever an index scan fetches a table entry.

If your index-only scan increases idx_tup_fetch, it is not truly index-only. PostgreSQL had to fetch the table rows anyway just to check if they are visible or not, an information that is not stored in the index.

You can see that as "heap tuples fetched" in EXPLAIN (ANALYZE) output.

The solution is to VACUUM the table so that the visibility map gets updated.