Query optimizer plans with very small selectivity estimates

Поиск
Список
Период
Сортировка
От Matthew Bellew
Тема Query optimizer plans with very small selectivity estimates
Дата
Msg-id CAJnjrPOziOSDWCUgfb1U9QUGdfLFspF6AZ9Xj=x5cOBhyqDFyA@mail.gmail.com
обсуждение исходный текст
Ответы Re: Query optimizer plans with very small selectivity estimates
Список pgsql-performance
Дерево обсуждения
This related to a post in the general bugs forum, but I found this forum, and
 this seems more appropriate.  This is my second attempt to post, I believe
the first attempt last week did not work, apologies if I'm duplicating.


I made have several users encounter performance problems, which all
seem to come down to this problem: multiplying selectivity estimates can 
cause tuple estimates to grow very small very quickly, once the estimator
gets to 1 row, the planner may choose plans that are very good ONLY WHEN 
there is exactly 1 row (maybe even O(N^large)).  Unfortunately, these may
be the worst plans if the estimate is even slightly off (even just returning 
2 or 3 rows versus 1).

Using the patch below, I discovered that clamping relation tuple estimates
to a number as small as 2 seemed to avoid all the catastrophic query 
plans.

In the scenarios I'm seeing, I have several examples of queries
that take >1m to run that should run in <1s. The estimate of 1 row 
(versus thousands actual) leads the planner to tee up several nest loop joins
which causes thousands of table scans.

I have been working on a more complete which tracks uniqueness along
with selectivity so that optimizer can benefit from knowing when a 
relation must have 1 (or fewer) tuples, while clamping all other relations
to 2 rather than 1.

typedef struct
{
double selectivity;
boolean unique;
} Selectivity;

I am interested in hearing discussion about this problem, and if the community
is open to a patch if I continue pursuing the development.

Matt



FIRST ARTIFACT

plan with expensive (80s join) and join estimate of 1
note the first Nested Loop join and 81s join
(I gave up trying to post the full explain, because of the 80 char limit)

"Sort  (cost=7000.04..7000.04 rows=1 width=49) 
(actual time=81739.426..81740.023 rows=5091 loops=1)"
"  Sort Key: c.ten DESC"
"  Sort Method: quicksort  Memory: 948kB"
"  CTE cte"
"    ->  Values Scan on "*VALUES*"  (cost=0.00..0.06 rows=5 width=4) 
(actual time=0.001..0.001 rows=5 loops=1)"
"  ->  Nested Loop  (cost=1.36..6999.97 rows=1 width=49) 
                    (actual time=0.059..81725.475 rows=5091 loops=1)"
"Planning time: 1.912 ms"
"Execution time: 81740.328 ms"


SECOND ARTIFACT

force join row estimate to be minimun of 2
query completes very quickly

"Sort  (cost=7000.06..7000.06 rows=2 width=49) 
(actual time=84.610..85.192 rows=5142 loops=1)"
"  Sort Key: c.ten DESC"
"  Sort Method: quicksort  Memory: 956kB"
"  CTE cte"
"    ->  Values Scan on "*VALUES*"  (cost=0.00..0.06 rows=5 width=4) 
(actual time=0.002..0.003 rows=5 loops=1)"
"  ->  Hash Join  (cost=2518.99..6999.98 rows=2 width=49) 
(actual time=17.629..82.886 rows=5142 loops=1)"

"Planning time: 2.982 ms"
"Execution time: 85.514 ms"


THIRD ARTIFACT

patch I used to make experimenting easier w/o recompiling


index 1b61fd9..444703c 100644
--- a/src/backend/optimizer/path/costsize.c
+++ b/src/backend/optimizer/path/costsize.c
@@ -68,6 +68,12 @@
  *-------------------------------------------------------------------------
  */
 
+
+
+/* These parameters are set by GUC */
+int                     join_row_estimate_clamp=1;
+
+
 #include "postgres.h"
 
 #ifdef _MSC_VER
@@ -175,6 +181,17 @@ clamp_row_est(double nrows)
 }
 
 
+double 
+clamp_join_row_est(double nrows)
+{
+ nrows = clamp_row_est(nrows);
+ if (nrows >= (double)join_row_estimate_clamp)
+ return nrows;
+        return (double)join_row_estimate_clamp;
+}
+
+
+
 /*
  * cost_seqscan
  *  Determines and returns the cost of scanning a relation sequentially.
@@ -3886,7 +3903,7 @@ calc_joinrel_size_estimate(PlannerInfo *root,
  break;
  }
 
- return clamp_row_est(nrows);
+ return clamp_join_row_est(nrows);
 }
 
 /*
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index 71090f2..fabb8ac 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -2664,6 +2664,16 @@ static struct config_int ConfigureNamesInt[] =
  NULL, NULL, NULL
  },
 
+ {
+ {"join_row_estimate_clamp", PGC_USERSET, QUERY_TUNING_OTHER,
+ gettext_noop("Set the minimum estimated size of a join result."),
+                        NULL
+ },
+ &join_row_estimate_clamp,
+ 1, 1, 10000,
+ NULL, NULL, NULL
+ },
+
  /* End-of-list marker */
  {
  {NULL, 0, 0, NULL, NULL}, NULL, 0, 0, 0, NULL, NULL, NULL
diff --git a/src/include/optimizer/cost.h b/src/include/optimizer/cost.h
index 25a7303..0161c4b 100644
--- a/src/include/optimizer/cost.h
+++ b/src/include/optimizer/cost.h
@@ -67,8 +67,10 @@ extern bool enable_material;
 extern bool enable_mergejoin;
 extern bool enable_hashjoin;
 extern int constraint_exclusion;
+extern int join_row_estimate_clamp;
 
 extern double clamp_row_est(double nrows);
+extern double clamp_join_row_est(double nrows);
 extern double index_pages_fetched(double tuples_fetched, BlockNumber pages,
  double index_pages, PlannerInfo *root);
 extern void cost_seqscan(Path *path, PlannerInfo *root, RelOptInfo *baserel,

В списке pgsql-performance по дате отправления:

Предыдущее
От: Alex Ignatov
Дата:
Сообщение: Re: Scalability to more than 64 cores With PG 9.4 and RHEL 7.1 Kernel 3.10
Следующее
От: Tom Lane
Дата:
Сообщение: Re: Query optimizer plans with very small selectivity estimates