While looking into a customer performance problem, i saw this in
costsize.c, cost_index() (9.3.6, but it looks the same in HEAD):
/* Mark the path with the correct row estimate */if (path->path.param_info){ path->path.rows =
path->path.param_info->ppi_rows; /* also get the set of clauses that should be enforced by the scan */ allclauses
=list_concat(list_copy(path->path.param_info->ppi_clauses),
baserel->baserestrictinfo);}else{ path->path.rows = baserel->rows; /* allclauses should just be the rel's
restrictionclauses */ allclauses = baserel->baserestrictinfo;}
What i'm wondering is the else branch, where the baserel row estimate is
assigned to the
IndexPath. However, it occurs to me that in conjunction with a partial
index, the overall row estimate will be constrained by the row estimate
from the partial index itself, no? I see that this doesn't have an impact
on the cost estimation of the index scan itself, since cost_index() does
this later:
/* estimate number of main-table tuples fetched */tuples_fetched = clamp_row_est(indexSelectivity * baserel->tuples);
I stumpled across this, since i see heavy misestimates in the EXPLAIN
output, where the estimated row count from the partial index vs. the real
row count heavily differs. In the customers case, there are two tables,
where one of the relation has many tuples in the JOIN condition which are
NULLs, like:
CREATE TABLE a2(id integer);
CREATE TABLE b2(id integer);
INSERT INTO a2 SELECT NULL FROM generate_series(1, 9800) AS t(id);
INSERT INTO a2 SELECT t.id FROM generate_series(1, 200) AS t(id);
INSERT INTO b2 SELECT t.id FROM generate_series(1, 200) AS t(id);
CREATE INDEX ON a2(id) WHERE id IS NOT NULL;
CREATE INDEX ON b2(id);
Here i get the following plan:
EXPLAIN ANALYZE SELECT * FROM b2 INNER JOIN a2 ON a2.id = b2.id ;
Merge Join (cost=10.79..12.63 rows=4 width=8) (actual time=0.084..0.291
rows=200 loops=1) Merge Cond: (b2.id = a2.id) -> Sort (cost=10.64..11.14 rows=200 width=4) (actual
time=0.069..0.082
rows=200 loops=1) Sort Key: b2.id Sort Method: quicksort Memory: 35kB -> Seq Scan on b2
(cost=0.00..3.00rows=200 width=4) (actual
time=0.010..0.027 rows=200 loops=1) -> Index Only Scan using a2_id_idx on a2 (cost=0.14..15.14 rows=10000
width=4) (actual time=0.012..0.074 rows=200 loops=1) Heap Fetches: 200Total runtime: 0.350 ms
EXPLAIN ANALYZE SELECT * FROM b2 INNER JOIN a2 ON a2.id = b2.id WHERE a2.id
IS NOT NULL;
Merge Join (cost=10.79..12.12 rows=1 width=8) (actual time=0.080..0.281
rows=200 loops=1) Merge Cond: (b2.id = a2.id) -> Sort (cost=10.64..11.14 rows=200 width=4) (actual
time=0.063..0.070
rows=200 loops=1) Sort Key: b2.id Sort Method: quicksort Memory: 35kB -> Seq Scan on b2
(cost=0.00..3.00rows=200 width=4) (actual
time=0.010..0.034 rows=200 loops=1) -> Index Only Scan using a2_id_idx on a2 (cost=0.14..15.64 rows=200
width=4) (actual time=0.012..0.052 rows=200 loops=1) Index Cond: (id IS NOT NULL) Heap Fetches: 200Total
runtime:0.335 ms
With the partial index predicate explicitly specified the row estimate is
accurate, without the predicate the row estimate of the index scan on
a2_id_idx assumes 10000.
It's very likely i miss something really important here, could someone shed
some light on this?
--
Thanks
Bernd
