当前位置：首页 > news >正文

Postgresql源码（145）优化器nestloop参数化路径评估不准问题分析

news 来源：原创 2025/5/2 6:15:18

相关
《Postgresql源码（133）优化器动态规划生成连接路径的实例分析》

1 问题

最近遇到一个问题，评估行数和真实行数存在较大差距，导致计划不准的问题。

nestloop内表评估是根据外表的参数来的。因为外表驱动表每取一条，内表才能做查询。所以这里外表的一条数据对内表来说就是参数。内表使用参数化路径来评估行数。

本篇针对这个实例对参数化路径行数评估做一些分析。

postgres=# explain analyze select * from iii, mmm where iii.poid = mmm.poid;QUERY PLAN
-----------------------------------------------------------------------------------------------------------------------------Nested Loop  (cost=0.43..125.97 rows=17 width=27) (actual time=0.063..6.031 rows=1000 loops=1)->  Seq Scan on mmm  (cost=0.00..1.10 rows=10 width=13) (actual time=0.005..0.010 rows=10 loops=1)->  Index Scan using idx_iii_poid on iii  (cost=0.43..12.47 rows=2 width=14) (actual time=0.022..0.563 rows=100 loops=10)Index Cond: (poid = mmm.poid)Planning Time: 0.351 msExecution Time: 6.138 ms
(6 rows)

2 全文总结（便于查询）

文章结论

get_parameterized_baserel_size计算参数化路径的选择率，本例中并没有看到任何特殊的评估方法，完全是按照poid的选择率来评估的。

clauselist_selectivity = 1.7108151058814915e-07
rel->tuples = 10000001nrows = rel->tuples * clauselist_selectivity = 2

pg_statistic记录的poid列唯一值个数占比-0.58516644
pg_class记录行数为10000001条，所以评估poid列的唯一值个数为10000001*0.58516644=5845167（实际9901001）。这里有一半的偏差，但不是主要原因。
2行=10000001*1.71e-07，在一个均匀分布的数据集中，每一个唯一值出现的概率是1.71e-07。所以如果有1千万行数据，随便给一个poid，能选出来两行不一样的。
- 如果唯一值非常少，那么选择率会变大，趋近于1，则10000001行=10000001*1。随便给一个poid，能选出来会非常多。
- 如果唯一值非常多，那么选择率会变小，趋近于0，则0行=10000001*0。随便给一个poid，能选出来会非常少。
评估两行但实际是100行的原因是，数据不均匀但统计信息平均采样，导致pg_statistic中stadistinct不准，进一步导致选择率不准，导致计算出现偏差。

分析思路整理

standard_join_search逐层规划连接顺序，找到参数化path节点。1.1.1 PATH 1：用索引参数化路径，评估2行
create_index_path在make_one_rel→set_base_rel_pathlists中一共调用6次，其中一次是参数化路径，特点是入参required_outer有值，可能拿到外表信息。
进一步分析create_index_path参数化成本计算，create_index_path→get_baserel_parampathinfo函数生成ParamPathInfo，ParamPathInfo中ppi_rows记录估算函数。
进一步分析ppi_rows的计算，在create_index_path→get_baserel_parampathinfo中处理，get_baserel_parampathinfo用行数1千万乘以选择率得到行数。
进一步分析选择率的计算
- clauselist_selectivity → clause_selectivity_ext → restriction_selectivity → eqsel_internal → var_eq_non_const
- 使用stadistinct = stats->stadistinct = -0.58516644 （来自pg_statistic表）
- 使用ntuples = vardata->rel->tuples = 10000001（来自relation结构，来自pg_class表）
- 计算唯一行数：clamp_row_est(-stadistinct * ntuples = 5845167.024516644) = 5845167
- 进而得到选择率selec = selec / ndistinct = 1 / 5845167 = 1.7108151058814915e-07

实例

计划二：优化器认为驱动表每一行，内表有2行能连接上，实际有100行能连接上。为什么差距大？
(cost=0.43..12.47 rows=2 width=14) (actual time=0.022..0.563 rows=100 loops=10)评估不准的原因是什么？

-- 计划二
drop table iii;
CREATE TABLE iii (poid INT NOT NULL, value NUMERIC, status int);
-- 可以和mmm表连上
INSERT INTO iii SELECT t%1000, t, 0 FROM generate_series(1, 100000) t order by random();
-- 干扰数据，占比高担都和mmm表连不上
INSERT INTO iii SELECT t, t, 0 FROM generate_series(100000, 10000000) t order by random();
CREATE INDEX idx_iii_poid ON iii(poid);
analyze iii;drop table mmm;
CREATE TABLE mmm (poid INT NOT NULL, value NUMERIC, status int);
INSERT INTO mmm SELECT t, t, 0 FROM generate_series(1, 10) t order by random();
CREATE INDEX idx_mmm_poid ON mmm(poid);
analyze mmm;set enable_hashjoin to off;
set enable_mergejoin to off;
set enable_bitmapscan to off;
explain analyze select * from iii, mmm where iii.poid = mmm.poid;postgres=# explain analyze select * from iii, mmm where iii.poid = mmm.poid;QUERY PLAN
-----------------------------------------------------------------------------------------------------------------------------Nested Loop  (cost=0.43..125.97 rows=17 width=27) (actual time=0.063..6.031 rows=1000 loops=1)->  Seq Scan on mmm  (cost=0.00..1.10 rows=10 width=13) (actual time=0.005..0.010 rows=10 loops=1)->  Index Scan using idx_iii_poid on iii  (cost=0.43..12.47 rows=2 width=14) (actual time=0.022..0.563 rows=100 loops=10)Index Cond: (poid = mmm.poid)Planning Time: 0.351 msExecution Time: 6.138 ms
(6 rows)

要分析这个问题需要从path生成开始看：

path生成

1 standard_join_search第一层

1.1 第一层第一个RelOptInfo（iii表）

p ((RelOptInfo*)root->join_rel_level[1].elements[0].ptr_value).pathlist

1.1.1 PATH 1：用索引参数化路径，评估2行，代价total_cost = 12.466928526553348

$53 = {path = {type = T_IndexPath,pathtype = T_IndexScan,parent = 0x1e91558,pathtarget = 0x1e697c8,param_info = 0x1e8d340,parallel_aware = false,parallel_safe = true,parallel_workers = 0,rows = 2,startup_cost = 0.435,total_cost = 12.466928526553348,pathkeys = 0x1e8cf90},indexinfo = 0x1e91768,indexclauses = 0x1e8cef0,indexorderbys = 0x0,indexorderbycols = 0x0,indexscandir = ForwardScanDirection,indextotalcost = 4.4499999999999993,indexselectivity = 1.7108151058814915e-07
}

param_info = 0x1e8d340 记录了什么？
在这里插入图片描述

1.1.2 PATH 2：全表扫，代价total_cost = 154055.01000000001

(gdb) tr Path 0x1e8bbc0
$55 = {type = T_Path,pathtype = T_SeqScan,parent = 0x1e91558,pathtarget = 0x1e697c8,param_info = 0x0,parallel_aware = false,parallel_safe = true,parallel_workers = 0,rows = 10000001,startup_cost = 0,total_cost = 154055.01000000001,pathkeys = 0x0
}

1.1.3 PATH 3：用索引全表扫iii表，评估10000001行。代价total_cost = 475019.93093073595

$57 = {path = {type = T_IndexPath,pathtype = T_IndexScan,parent = 0x1e91558,pathtarget = 0x1e697c8,param_info = 0x0,parallel_aware = false,parallel_safe = true,parallel_workers = 0,rows = 10000001,startup_cost = 0.435,total_cost = 475019.93093073595,pathkeys = 0x1e8c7d0},indexinfo = 0x1e91768,indexclauses = 0x0,indexorderbys = 0x0,indexorderbycols = 0x0,indexscandir = ForwardScanDirection,indextotalcost = 158924.44,indexselectivity = 1
}

1.2 第一层第二个RelOptInfo（mmm表）

p ((RelOptInfo*)root->join_rel_level[1].elements[1].ptr_value).pathlist

1.2.1 PATH 1：用索引参数化路径，评估1行，代价total_cost = 0.15250119999987999

$70 = {path = {type = T_IndexPath,pathtype = T_IndexScan,parent = 0x1e90d28,pathtarget = 0x1e698f8,param_info = 0x1eac2d8,parallel_aware = false,parallel_safe = true,parallel_workers = 0,rows = 1,startup_cost = 0.13500000000000001,total_cost = 0.15250119999987999,pathkeys = 0x1eabfd8},indexinfo = 0x1e8b8c8,indexclauses = 0x1eabf38,indexorderbys = 0x0,indexorderbycols = 0x0,indexscandir = ForwardScanDirection,indextotalcost = 0.14250079999991999,indexselectivity = 0.10000000000000001
}

1.2.2 PATH 2：全表扫，评估10行，代价total_cost = 1.1000000000000001

$71 = {type = T_Path,pathtype = T_SeqScan,parent = 0x1e90d28,pathtarget = 0x1e698f8,param_info = 0x0,parallel_aware = false,parallel_safe = true,parallel_workers = 0,rows = 10,startup_cost = 0,total_cost = 1.1000000000000001,pathkeys = 0x0
}

1.2.3 PATH 3：索引全表扫，评估10行，代价otal_cost = 12.285

{path = {type = T_IndexPath,pathtype = T_IndexScan,parent = 0x1e90d28,pathtarget = 0x1e698f8,param_info = 0x0,parallel_aware = false,parallel_safe = true,parallel_workers = 0,rows = 10,startup_cost = 0.13500000000000001,total_cost = 12.285,pathkeys = 0x1eab998},indexinfo = 0x1e8b8c8,indexclauses = 0x0,indexorderbys = 0x0,indexorderbycols = 0x0,indexscandir = ForwardScanDirection,indextotalcost = 8.1850000000000005,indexselectivity = 1
}

2 standard_join_search第二层，只有一个RelOptInfo

p ((RelOptInfo*)root->join_rel_level[2].elements[0].ptr_value).pathlist

只有一个PATH

$91 = {jpath = {path = {type = T_NestPath,pathtype = T_NestLoop,parent = 0x1eacc28,pathtarget = 0x1eace58,param_info = 0x0,parallel_aware = false,parallel_safe = true,parallel_workers = 0,rows = 17,startup_cost = 0.435,total_cost = 125.96928526553347,pathkeys = 0x0},jointype = JOIN_INNER,inner_unique = false,outerjoinpath = 0x1e8d780,  innerjoinpath = 0x1e8d020,joinrestrictinfo = 0x0}
}

outerjoinpath = 0x1e8d780：选择了1.2.2 PATH 2：全表扫，评估10行，代价total_cost = 1.1000000000000001。
innerjoinpath = 0x1e8d020：选择了1.1.1 PATH 1：用索引参数化路径，评估2行，代价total_cost = 12.466928526553348。

为什么iii表的索引参数化路径评估只有2行？

现阶段只有nestloop的内表需要参数化路径，因为内表评估行数时，无法确切知道能连上多少行，所以只能计算出一个行数。
在上文 1.1.1 PATH 1：用索引参数化路径，评估2行，代价total_cost = 12.466928526553348中，参数化node评估的行数是怎么计算出来的？

$53 = {path = {type = T_IndexPath,pathtype = T_IndexScan,parent = 0x1e91558,pathtarget = 0x1e697c8,param_info = 0x1e8d340,parallel_aware = false,parallel_safe = true,parallel_workers = 0,rows = 2,startup_cost = 0.435,total_cost = 12.466928526553348,pathkeys = 0x1e8cf90},indexinfo = 0x1e91768,indexclauses = 0x1e8cef0,indexorderbys = 0x0,indexorderbycols = 0x0,indexscandir = ForwardScanDirection,indextotalcost = 4.4499999999999993,indexselectivity = 1.7108151058814915e-07
}

create_index_path在make_one_rel→set_base_rel_pathlists中一共调用6次：

-- iii表的索引 idx_iii_poid，非参数化路径
create_index_path (root=0x1e68cb8, index=0x1e91768, indexclauses=0x0,       indexorderbys=0x0, indexorderbycols=0x0, pathkeys=0x1e8c7d0, indexscandir=ForwardScanDirection, indexonly=false, required_outer=0x0,       loop_count=1, partial_path=false)
-- iii表的索引 idx_iii_poid，非参数化路径
create_index_path (root=0x1e68cb8, index=0x1e91768, indexclauses=0x0,       indexorderbys=0x0, indexorderbycols=0x0, pathkeys=0x1e8c7d0, indexscandir=ForwardScanDirection, indexonly=false, required_outer=0x0,       loop_count=1, partial_path=true)
-- iii表的索引 idx_iii_poid，required_outer有值，计算参数化路径
create_index_path (root=0x1e68cb8, index=0x1e91768, indexclauses=0x1e8cef0, indexorderbys=0x0, indexorderbycols=0x0, pathkeys=0x1e8cf90, indexscandir=ForwardScanDirection, indexonly=false, required_outer=0x1e8cca0, loop_count=10, partial_path=false)-- mmm表的索引 idx_mmm_poid
create_index_path (root=0x1e68cb8, index=0x1e8b8c8, indexclauses=0x0,       indexorderbys=0x0, indexorderbycols=0x0, pathkeys=0x1eab998, indexscandir=ForwardScanDirection, indexonly=false, required_outer=0x0,       loop_count=1, partial_path=false)
create_index_path (root=0x1e68cb8, index=0x1e8b8c8, indexclauses=0x0,       indexorderbys=0x0, indexorderbycols=0x0, pathkeys=0x1eab998, indexscandir=ForwardScanDirection, indexonly=false, required_outer=0x0,       loop_count=1, partial_path=true)
create_index_path (root=0x1e68cb8, index=0x1e8b8c8, indexclauses=0x1eabf38, indexorderbys=0x0, indexorderbycols=0x0, pathkeys=0x1eabfd8, indexscandir=ForwardScanDirection, indexonly=false, required_outer=0x1e8d2d0, loop_count=10000001, partial_path=false)

required_outer有值，表示参数化的indexpath：

-- iii表的索引 idx_iii_poid，required_outer有值，计算参数化路径
create_index_path (root=0x1e68cb8, index=0x1e91768, indexclauses=0x1e8cef0, indexorderbys=0x0, indexorderbycols=0x0, pathkeys=0x1e8cf90, indexscandir=ForwardScanDirection, indexonly=false, required_outer=0x1e8cca0, loop_count=10, partial_path=false)

在这里插入图片描述

参数化路径计算

create_index_path→get_baserel_parampathinfo函数生成ParamPathInfo

get_baserel_parampathinfo调用get_parameterized_baserel_size评估，当前节点是几行。例如评估出来2行。explain时内表看到的rows就是2行，含义是外表每一行过来，内表有两行能连上。

get_parameterized_baserel_size

double
get_parameterized_baserel_size(PlannerInfo *root, RelOptInfo *rel,List *param_clauses)
{List	   *allclauses;double		nrows;allclauses = list_concat_copy(param_clauses, rel->baserestrictinfo);nrows = rel->tuples *clauselist_selectivity(root,allclauses,rel->relid,	/* do not use 0! */JOIN_INNER,NULL);nrows = clamp_row_est(nrows);/* For safety, make sure result is not more than the base estimate */if (nrows > rel->rows)nrows = rel->rows;return nrows;
}

allclauses合并param_clauses参数化条件和rel->baserestrictinfo当前表自身的约束条件，得到综合条件列表。
使用 clauselist_selectivity 函数计算这些条件的综合选择率，再乘以表的基数rel->tuples，最终得到参数化路径下的预估行数。

当前的allclauses表本身没有条件，只有外表提供的参数化条件：
在这里插入图片描述

上图中opexpr是等号，指向varno=1、varno=2指向两个基表。
nrows = rel->tuples * clauselist_selectivity = 2
- clauselist_selectivity = 1.7108151058814915e-07
- rel->tuples = 10000001

clauselist_selectivity → clause_selectivity_ext的计算逻辑分支较多，这里只给出本例走到的分支：

clause_selectivity_ext......clause = (Node *) rinfo->clause;......else if (is_opclause(clause) || IsA(clause, DistinctExpr))......restriction_selectivity...

restriction_selectivity用于计算条件（where中过滤条件）的选择率：

Selectivity
restriction_selectivity(PlannerInfo *root,Oid operatorid,List *args,Oid inputcollid,int varRelid)
{

这里operatorid=96是等号。
get_oprrest从系统表中拿到oprrest 字段，字段指向一个注册的选择率计算函数oid=101，eqsel函数。

	RegProcedure oprrest = get_oprrest(operatorid);float8		result;/** if the oprrest procedure is missing for whatever reason, use a* selectivity of 0.5*/if (!oprrest)return (Selectivity) 0.5;result = DatumGetFloat8(OidFunctionCall4Coll(oprrest,inputcollid,PointerGetDatum(root),ObjectIdGetDatum(operatorid),PointerGetDatum(args),Int32GetDatum(varRelid)));if (result < 0.0 || result > 1.0)elog(ERROR, "invalid restriction selectivity: %f", result);return (Selectivity) result;
}

进入eqsel函数开始计算选择率，eqsel用于计算等值和不等值操作符的选择率，是优化器代价模型的核心组成部分。

root：优化器上下文（包含统计信息、表元数据等）
operator：操作符的 OID（如 = 或 <>）
args：操作符的参数列表（如 a = 5 中的列 a 和常量 5）
varRelid：关联的关系 ID（用于确定统计信息范围）
collation：排序规则（影响字符串比较）

/** Common code for eqsel() and neqsel()*/
static double
eqsel_internal(PG_FUNCTION_ARGS, bool negate)
{PlannerInfo *root = (PlannerInfo *) PG_GETARG_POINTER(0);Oid			operator = PG_GETARG_OID(1);List	   *args = (List *) PG_GETARG_POINTER(2);int			varRelid = PG_GETARG_INT32(3);Oid			collation = PG_GET_COLLATION();VariableStatData vardata;Node	   *other;bool		varonleft;double		selec;

negate表示不等值。

	if (negate){operator = get_negator(operator);if (!OidIsValid(operator)){/* Use default selectivity (should we raise an error instead?) */return 1.0 - DEFAULT_EQ_SEL;}}

本例中的args：

Var = {xpr = {type = T_Var}, varno = 1, varattno = 1, vartype = 23}
Var = {xpr = {type = T_Var}, varno = 2, varattno = 1, vartype = 23}
varRelid = 1

	/** If expression is not variable = something or something = variable, then* punt and return a default estimate.*/if (!get_restriction_variable(root, args, varRelid,&vardata, &other, &varonleft))return negate ? (1.0 - DEFAULT_EQ_SEL) : DEFAULT_EQ_SEL;

get_restriction_variable将条件拆分为变量和常量或其他表达式。

vardata
- {var = 0x1e6b2c0, rel = 0x1e8b8f8, statsTuple = 0x7f4874e75e08, freefunc = 0xb73cd6 , vartype = 23, atttype = 23, atttypmod = -1, isunique = false, acl_ok = true}
  - var = {xpr = {type = T_Var}, varno = 1, varattno = 1, vartype = 23}
  - statsTuple = pg_stat_statistic对应的行。
other = {{xpr = {type = T_Var}, varno = 2, varattno = 1, vartype = 2}
varonleft = true

开始计算选择率 var_eq_non_const

	/** We can do a lot better if the something is a constant.  (Note: the* Const might result from estimation rather than being a simple constant* in the query.)*/if (IsA(other, Const))selec = var_eq_const(&vardata, operator, collation,((Const *) other)->constvalue,((Const *) other)->constisnull,varonleft, negate);elseselec = var_eq_non_const(&vardata, operator, collation, other,varonleft, negate);ReleaseVariableStats(vardata);return selec;
}

var_eq_non_const用于计算变量与非常量表达式（如其他列或子查询结果）的等值条件（= 或 <>）的选择率，例如 a = b 或 x <> y。核心逻辑是基于统计信息（如唯一性约束、NULL值比例、不同值数量等），结合启发式规则估算满足条件的行数比例

vardata：变量统计信息（如列 a 的直方图、MCV列表等）。
oproid：操作符OID（如 = 或 <>）。
other：非常量表达式（如另一列 b 或表达式 b + 1）。
negate：不等。


/** var_eq_non_const --- eqsel for var = something-other-than-const case** This is exported so that some other estimation functions can use it.*/
double
var_eq_non_const(VariableStatData *vardata, Oid oproid, Oid collation,Node *other,bool varonleft, bool negate)
{double		selec;double		nullfrac = 0.0;bool		isdefault;

先拿到pg_statistic表对应的行。

	if (HeapTupleIsValid(vardata->statsTuple)){Form_pg_statistic stats;stats = (Form_pg_statistic) GETSTRUCT(vardata->statsTuple);nullfrac = stats->stanullfrac;}...else if (HeapTupleIsValid(vardata->statsTuple)){double		ndistinct;AttStatsSlot sslot;

基于统计信息估算：

计算非NULL值比例 1 - nullfrac。
通过 get_variable_numdistinct 获取变量不同值数量 ndistinct。
假设每个不同值均匀分布，选择率为 (1 - nullfrac) / ndistinct。
与MCV（最常见值）的最大频率比较，避免高估。

		selec = 1.0 - nullfrac;ndistinct = get_variable_numdistinct(vardata, &isdefault);

get_variable_numdistinct计算结果：5845167

stadistinct = stats->stadistinct = -0.58516644 （来自pg_statistic表）
stanullfrac = 0
ntuples = vardata->rel->tuples = 10000001（来自relation结构，来自pg_class表）
计算结果：clamp_row_est(-stadistinct * ntuples = 5845167.024516644) = 5845167

get_variable_numdistinct计算方法总结：pg_statistic表的stadistinct中取得选择率乘以评估行数，等于评估出来多少个唯一值。stadistinct含义：stadistinct > 0表示该列中非空唯一值的实际数量。stadistinct < 0表示非空唯一值的数量占总行数的比例。stadistinct = 0表示无法确定该列的非空唯一值数量。

		if (ndistinct > 1)selec /= ndistinct;......return selec;

计算得到选择率
selec = selec / ndistinct = 1 / 5845167 = 1.7108151058814915e-07

回到最初的行数评估函数get_parameterized_baserel_size中：

nrows = rel->tuples * clauselist_selectivity = 2
- clauselist_selectivity = 1.7108151058814915e-07
- rel->tuples = 10000001

总结：

pg_statistic记录的poid列唯一值个数占比-0.58516644
pg_class记录行数为10000001条，所以评估poid列的唯一值个数为10000001*0.58516644=5845167（实际9901001）。这里有一半的偏差，但不是主要原因。
2行=10000001*1.71e-07，在一个均匀分布的数据集中，每一个唯一值出现的概率是1.71e-07。所以如果有1千万行数据，随便给一个poid，能选出来两行不一样的。
- 如果唯一值非常少，那么选择率会变大，趋近于1，则10000001行=10000001*1。随便给一个poid，能选出来会非常多。
- 如果唯一值非常多，那么选择率会变小，趋近于0，则0行=10000001*0。随便给一个poid，能选出来会非常少。
评估两行但实际是100行的原因是，pg_statistic中stadistinct记录的连接条件poid列的唯一值。

create_index_path→cost_index计算参数化路径评估行数（从上一步的ParamPathInfo结果中取得）

cost_index中如果发现参数化路径，会从ParamPathInfo中取评估行数
本例中path->path.param_info->ppi_rows=2。
param_info存在表示当前索引扫描路径是参数化的，依赖外部循环（如嵌套循环连接的外层表）提供的参数值。
行数估算：ppi_rows 表示参数化路径的预估行数，这是优化器根据外层表（如嵌套循环中的驱动表）的约束条件和连接关系动态调整的结果，比如ppi_rows=2表示，优化器评估外表每一条，内表有两条能连得上。

void
cost_index(IndexPath *path, PlannerInfo *root, double loop_count,bool partial_path)......if (path->path.param_info){path->path.rows = path->path.param_info->ppi_rows;/* qpquals come from the rel's restriction clauses and ppi_clauses */qpquals = list_concat(extract_nonindex_conditions(path->indexinfo->indrestrictinfo,path->indexclauses),extract_nonindex_conditions(path->path.param_info->ppi_clauses,path->indexclauses));}else{path->path.rows = baserel->rows;/* qpquals come from just the rel's restriction clauses */qpquals = extract_nonindex_conditions(path->indexinfo->indrestrictinfo,path->indexclauses);}...