Vahe Evoyan <vahe.evoyan@gmail.com> writes:
> Merge Join (cost=0.00..2513.96 rows=1 width=72) (actual
> time=127.361..473.687 rows=66460 loops=1)
> Merge Cond: ((v1.dft_id = v2.dft_id) AND ((v1.key)::text =
> (v2.key)::text))
> -> Index Scan using qor_value_self_join on qor_value v1
> (cost=0.00..1255.60 rows=275 width=51) (actual time=89.549..97.045
> rows=1388 loops=1)
> Index Cond: ((run_id = 60807) AND (stat_id = 342))
> -> Index Scan using qor_value_self_join on qor_value v2
> (cost=0.00..1255.60 rows=275 width=51) (actual time=37.796..134.286
> rows=66343 loops=1)
> Index Cond: ((run_id = 60875) AND (stat_id = 342))
> Total runtime: 544.646 ms
> (7 rows)
> Note that the second Index Scan has 66343 rows in place of 1388.
That's not a bug. That's a result of rescanning portions of the inner
relation's output due to duplicate mergejoin keys in the outer relation.
The EXPLAIN ANALYZE machinery counts the re-fetches as if they were new
rows, though in some sense they're not.
regards, tom lane