PostgreSQL的set_base_rel_sizes函数分析

这篇文章主要讲解了“postgresql的set_base_rel_sizes函数分析”，文中的讲解内容简单清晰，易于学习与理解，下面请大家跟着小编的思路慢慢深入，一起来研究和学习“Postgresql的set_base_rel_sizes函数分析”吧！

在set_base_rel_sizes函数调用过程中，如RTE为子查询，则生成子查询的访问路径，通过调用函数set_subquery_pathlist实现。

make_one_rel源代码:

 RelOptInfo *
 make_one_rel(PlannerInfo *root, List *joinlist)
 {
     //...

     
     set_base_rel_sizes(root);//估算Relation的Size并且设置consider_parallel标记
     //...
 }

一、数据结构

RelOptInfo
如前所述,RelOptInfo数据结构贯穿整个优化过程.

 typedef struct RelOptInfo
 {
     nodeTag     type;//节点标识
 
     RelOptKind  reloptkind;//RelOpt类型
 
     
     Relids      relids;         
 
     
     double      rows;           
 
     
     bool        consider_startup;   
     bool        consider_param_startup; 
     bool        consider_parallel;  
 
     
     struct PathTarget *reltarget;   
 
     
     List       *pathlist;       
     List       *ppilist;        
     List       *partial_pathlist;   
     struct Path *cheapest_startup_path;//代价最低的启动路径
     struct Path *cheapest_total_path;//代价最低的整体路径
     struct Path *cheapest_unique_path;//代价最低的获取唯一值的路径
     List       *cheapest_parameterized_paths;//代价最低的参数化路径链表
 
     
     
     Relids      direct_lateral_relids;  
     Relids      lateral_relids; 
 
     
     //reloptkind=RELOPT_BASEREL时使用的数据结构
     Index       relid;          
     Oid         reltablespace;  
     RTEKind     rtekind;        
     AttrNumber  min_attr;       
     AttrNumber  max_attr;       
     Relids     *attr_needed;    
     int32      *attr_widths;    
     List       *lateral_vars;   
     Relids      lateral_referencers;    
     List       *indexlist;      
     List       *statlist;       
     BlockNumber pages;          
     double      tuples;         
     double      allvisfrac;     
     PlannerInfo *subroot;       
     List       *subplan_params; 
     int         rel_parallel_workers;   
 
     
     //FDW相关信息
     Oid         serverid;       
     Oid         userid;         
     bool        useridiscurrent;    
     
     struct FdwRoutine *fdwroutine;
     void       *fdw_private;
 
     
     //已知的,可保证唯一元组返回的Relids链表
     List       *unique_for_rels;    
     List       *non_unique_for_rels;    
 
     
     List       *baserestrictinfo;   
     QualCost    baserestrictcost;   
     Index       baserestrict_min_security;  
     List       *joininfo;       
     bool        has_eclass_joins;   
 
     
       //是否尝试partitionwise连接,这是PG 11的一个新特性.
     bool        consider_partitionwise_join;    
     Relids      top_parent_relids;  
 
     
     //分区表使用
     PartitionScheme part_scheme;    
     int         nparts;         
     struct PartitionBoundInfoData *boundinfo;   
     List       *partition_qual; 
     struct RelOptInfo **part_rels;  
     List      **partexprs;      
     List      **nullable_partexprs; 
     List       *partitioned_child_rels; 
 } RelOptInfo;

二、源码解读

set_rel_size
如前所述,set_rel_size函数估算关系的大小,如RTE为子查询,则调用set_subquery_pathlist方法生成子查询访问路径,相关代码如下:

 
 static void
 set_rel_size(PlannerInfo *root, RelOptInfo *rel,
              Index rti, RangeTblEntry *rte)
 {
     if (rel->reloptkind == RELOPT_BASEREL &&
         relation_excluded_by_constraints(root, rel, rte))
     {
        //...
     }
     else
     {
         switch (rel->rtekind)
         {
             case //...
                 //...
             case RTE_SUBQUERY://子查询
 
                 
                 set_subquery_pathlist(root, rel, rti, rte);//生成子查询访问路径
                 break;
             case ...
          }
        //...
      }
      //...
}

set_subquery_pathlist
生成子查询的扫描路径,在生成过程中尝试下推外层的限制条件(减少参与运算的元组数量),并调用subquery_planner生成执行计划.

 
 static void
 set_subquery_pathlist(PlannerInfo *root, RelOptInfo *rel,
                       Index rti, RangeTblEntry *rte)
 {
     Query      *parse = root->parse;
     Query      *subquery = rte->subquery;
     Relids      required_outer;
     pushdown_safety_info safetyInfo;
     double      tuple_fraction;
     RelOptInfo *sub_final_rel;
     ListCell   *lc;
 
     
     subquery = copyObject(subquery);//拷贝
 
     
     required_outer = rel->lateral_relids;//外层的Relids
 
     
     memset(&safetyInfo, 0, sizeof(safetyInfo));
     safetyInfo.unsafeColumns = (bool *)
         palloc0((list_length(subquery->targetList) + 1) * sizeof(bool));
 
     
     safetyInfo.unsafeLeaky = rte->security_barrier;
 
     
     if (rel->baserestrictinfo != NIL &&
         subquery_is_pushdown_safe(subquery, subquery, &safetyInfo))
     {
     //可以下推限制条件
         
         List       *upperrestrictlist = NIL;
         ListCell   *l;
 
         foreach(l, rel->baserestrictinfo)//遍历子查询上的限制条件
         {
             RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
             Node       *clause = (Node *) rinfo->clause;
 
             if (!rinfo->pseudoconstant &&
                 qual_is_pushdown_safe(subquery, rti, clause, &safetyInfo))
             {
                 
                 subquery_push_qual(subquery, rte, rti, clause);//下推限制条件
             }
             else
             {
                 
                 upperrestrictlist = lappend(upperrestrictlist, rinfo);//保留在上层中
             }
         }
         rel->baserestrictinfo = upperrestrictlist;
         
     }
 
     pfree(safetyInfo.unsafeColumns);
 
     
     remove_unused_subquery_outputs(subquery, rel);
 
     
     if (parse->hasAggs ||
         parse->groupClause ||
         parse->groupingSets ||
         parse->havingQual ||
         parse->distinctClause ||
         parse->sortClause ||
         has_multiple_baserels(root))
         tuple_fraction = 0.0;   
     else
         tuple_fraction = root->tuple_fraction;
 
     
     Assert(root->plan_params == NIL);
 
     
     rel->subroot = subquery_planner(root->glob, subquery,
                                     root,
                                     false, tuple_fraction);//调用subquery_planner获取子查询的执行计划
 
     
     rel->subplan_params = root->plan_params;
     root->plan_params = NIL;
 
     
     sub_final_rel = fetch_upper_rel(rel->subroot, UPPERREL_FINAL, NULL);//子查询返回的最终关系
 
     if (IS_DUMMY_REL(sub_final_rel))
     {
         set_dummy_rel_pathlist(rel);
         return;
     }
 
     
     set_subquery_size_estimates(root, rel);//设置子查询的估算信息
 
     
     foreach(lc, sub_final_rel->pathlist)//遍历最终关系的访问路径
     {
         Path       *subpath = (Path *) lfirst(lc);
         List       *pathkeys;
 
         
     //转换pathkeys为外层的表示法
         pathkeys = convert_subquery_pathkeys(root,
                                              rel,
                                              subpath->pathkeys,
                                              make_tlist_from_pathtarget(subpath->pathtarget));
 
         
         add_path(rel, (Path *)
                  create_subqueryscan_path(root, rel, subpath,
                                           pathkeys, required_outer));//通过子查询路径生成外层访问路径
     }
 
     
     if (rel->consider_parallel && bms_is_empty(required_outer))//是否可以并行处理
     {
         
         Assert(sub_final_rel->consider_parallel ||
                sub_final_rel->partial_pathlist == NIL);
 
         
         foreach(lc, sub_final_rel->partial_pathlist)
         {
             Path       *subpath = (Path *) lfirst(lc);
             List       *pathkeys;
 
             
             pathkeys = convert_subquery_pathkeys(root,
                                                  rel,
                                                  subpath->pathkeys,
                                                  make_tlist_from_pathtarget(subpath->pathtarget));
 
             
             add_partial_path(rel, (Path *)
                              create_subqueryscan_path(root, rel, subpath,
                                                       pathkeys,
                                                       required_outer));
         }
     }
 }

//-------------------------------------------------------- create_subqueryscan_path
 
 SubqueryScanPath *
 create_subqueryscan_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath,
                          List *pathkeys, Relids required_outer)
 {
     SubqueryScanPath *pathnode = makeNode(SubqueryScanPath);
 
     pathnode->path.pathtype = T_SubqueryScan;//路径类型:子查询扫描
     pathnode->path.parent = rel;//父RelOptInfo
     pathnode->path.pathtarget = rel->reltarget;//投影列
     pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
                                                           required_outer);//参数化信息
     pathnode->path.parallel_aware = false;//并行相关参数
     pathnode->path.parallel_safe = rel->consider_parallel &&
         subpath->parallel_safe;
     pathnode->path.parallel_workers = subpath->parallel_workers;
     pathnode->path.pathkeys = pathkeys;//排序键
     pathnode->subpath = subpath;//子访问路径
 
     cost_subqueryscan(pathnode, root, rel, pathnode->path.param_info);//子查询的成本
 
     return pathnode;
 }
 
//-------------------------------------------------------- cost_subqueryscan

 
 void
 cost_subqueryscan(SubqueryScanPath *path, PlannerInfo *root,
                   RelOptInfo *baserel, ParamPathInfo *param_info)
 {
     Cost        startup_cost;
     Cost        run_cost;
     QualCost    qpqual_cost;
     Cost        cpu_per_tuple;
 
     
     Assert(baserel->relid > 0);
     Assert(baserel->rtekind == RTE_SUBQUERY);
 
     
     if (param_info)
         path->path.rows = param_info->ppi_rows;
     else
         path->path.rows = baserel->rows;
 
     
     path->path.startup_cost = path->subpath->startup_cost;
     path->path.total_cost = path->subpath->total_cost;
 
     get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
 
     startup_cost = qpqual_cost.startup;
     cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
     run_cost = cpu_per_tuple * baserel->tuples;
 
     
     startup_cost += path->path.pathtarget->cost.startup;
     run_cost += path->path.pathtarget->cost.per_tuple * path->path.rows;
 
     path->path.startup_cost += startup_cost;
     path->path.total_cost += startup_cost + run_cost;
 }

三、跟踪分析

测试脚本如下:

select t1.*,t2.dwbh,t2.counter 
from t_dwxx t1,
    (select dwbh,count(*) as counter from t_grxx group by dwbh) t2 
where t1.dwbh = t2.dwbh and t1.dwbh = '1001';

启动gdb:

(gdb) c
Continuing.

Breakpoint 1, set_subquery_pathlist (root=0x2d749b0, rel=0x2d34dd0, rti=2, rte=0x2d341a0) at allpaths.c:2082
2082    Query    *parse = root->parse;

进入函数set_subquery_pathlist,输入参数中的root->simple_rel_array[2],rtekind为RTE_SUBQUERY子查询

(gdb) p *root->simple_rel_array[2]
$13 = {type = T_RelOptInfo, reloptkind = RELOPT_BASEREL, relids = 0x2d6a428, rows = 0, consider_startup = false, 
  consider_param_startup = false, consider_parallel = true, reltarget = 0x2d6a440, pathlist = 0x0, ppilist = 0x0, 
  partial_pathlist = 0x0, cheapest_startup_path = 0x0, cheapest_total_path = 0x0, cheapest_unique_path = 0x0, 
  cheapest_parameterized_paths = 0x0, direct_lateral_relids = 0x0, lateral_relids = 0x0, relid = 2, reltablespace = 0, 
  rtekind = RTE_SUBQUERY, min_attr = 0, max_attr = 2, attr_needed = 0x2d69b00, attr_widths = 0x2d69b50, lateral_vars = 0x0, 
  lateral_referencers = 0x0, indexlist = 0x0, statlist = 0x0, pages = 0, tuples = 0, allvisfrac = 0, subroot = 0x0, 
  subplan_params = 0x0, rel_parallel_workers = -1, serverid = 0, userid = 0, useridiscurrent = false, fdwroutine = 0x0, 
  fdw_private = 0x0, unique_for_rels = 0x0, non_unique_for_rels = 0x0, baserestrictinfo = 0x2d6b648, baserestrictcost = {
    startup = 0, per_tuple = 0}, baserestrict_min_security = 0, joininfo = 0x0, has_eclass_joins = true, 
  top_parent_relids = 0x0, part_scheme = 0x0, nparts = 0, boundinfo = 0x0, partition_qual = 0x0, part_rels = 0x0, 
  partexprs = 0x0, nullable_partexprs = 0x0, partitioned_child_rels = 0x0}
(gdb) p *rte
$10 = {type = T_RangeTblEntry, rtekind = RTE_SUBQUERY, relid = 0, relkind = 0 '\000', tablesample = 0x0, 
  subquery = 0x2d342b0, security_barrier = false, jointype = JOIN_INNER, joinaliasvars = 0x0, functions = 0x0, 
  funcordinality = false, tablefunc = 0x0, values_lists = 0x0, ctename = 0x0, ctelevelsup = 0, self_reference = false, 
  coltypes = 0x0, coltypmods = 0x0, colcollations = 0x0, enrname = 0x0, enrtuples = 0, alias = 0x2c82728, eref = 0x2d35328, 
  lateral = false, inh = false, inFromCl = true, requiredPerms = 0, checkAsUser = 0, selectedCols = 0x0, 
  insertedCols = 0x0, updatedCols = 0x0, securityQuals = 0x0}

下推限制条件:

...
(gdb) n
2157          qual_is_pushdown_safe(subquery, rti, clause, &safetyInfo))
(gdb) 
2156        if (!rinfo->pseudoconstant &&
(gdb) 
2160          subquery_push_qual(subquery, rte, rti, clause);

tuple_fraction设置为0.0

...
(gdb) 
2193      tuple_fraction = 0.0; 

调用subquery_planner获取执行计划:

(gdb) n
2201    rel->subroot = subquery_planner(root->glob, subquery,

获取子查询生成的最终关系,reloptkind为RELOPT_UPPER_REL

...
(gdb) 
2214    sub_final_rel = fetch_upper_rel(rel->subroot, UPPERREL_FINAL, NULL);
(gdb) 
2216    if (IS_DUMMY_REL(sub_final_rel))
(gdb) p *sub_final_rel
$16 = {type = T_RelOptInfo, reloptkind = RELOPT_UPPER_REL, relids = 0x0, rows = 0, consider_startup = false, 
  consider_param_startup = false, consider_parallel = true, reltarget = 0x2d7bd50, pathlist = 0x2d7be10, ppilist = 0x0, 
  partial_pathlist = 0x0, cheapest_startup_path = 0x2d7aaa8, cheapest_total_path = 0x2d7aaa8, cheapest_unique_path = 0x0, 
  cheapest_parameterized_paths = 0x2d7be60, direct_lateral_relids = 0x0, lateral_relids = 0x0, relid = 0, 
  reltablespace = 0, rtekind = RTE_RELATION, min_attr = 0, max_attr = 0, attr_needed = 0x0, attr_widths = 0x0, 
  lateral_vars = 0x0, lateral_referencers = 0x0, indexlist = 0x0, statlist = 0x0, pages = 0, tuples = 0, allvisfrac = 0, 
  subroot = 0x0, subplan_params = 0x0, rel_parallel_workers = 0, serverid = 0, userid = 0, useridiscurrent = false, 
  fdwroutine = 0x0, fdw_private = 0x0, unique_for_rels = 0x0, non_unique_for_rels = 0x0, baserestrictinfo = 0x0, 
  baserestrictcost = {startup = 0, per_tuple = 0}, baserestrict_min_security = 0, joininfo = 0x0, has_eclass_joins = false, 
  top_parent_relids = 0x0, part_scheme = 0x0, nparts = 0, boundinfo = 0x0, partition_qual = 0x0, part_rels = 0x0, 
  partexprs = 0x0, nullable_partexprs = 0x0, partitioned_child_rels = 0x0}

成本最低的路径

(gdb) p *sub_final_rel->cheapest_total_path
$17 = {type = T_AggPath, pathtype = T_Agg, parent = 0x2d7b6d0, pathtarget = 0x2d7adc8, param_info = 0x0, 
  parallel_aware = false, parallel_safe = true, parallel_workers = 0, rows = 10, startup_cost = 0.29249999999999998, 
  total_cost = 20.143376803383145, pathkeys = 0x0}

通过子查询访问路径生成外层的访问路径

(gdb) n
2227    set_subquery_size_estimates(root, rel);
(gdb) 
2233    foreach(lc, sub_final_rel->pathlist)
(gdb) 
2235      Path     *subpath = (Path *) lfirst(lc);
(gdb) 
2239      pathkeys = convert_subquery_pathkeys(root,
(gdb) 
2246           create_subqueryscan_path(root, rel, subpath,
(gdb) 
2245      add_path(rel, (Path *)
(gdb) 
2233    foreach(lc, sub_final_rel->pathlist)

完成函数调用,结束处理

(gdb) 
set_rel_size (root=0x2d749b0, rel=0x2d34dd0, rti=2, rte=0x2d341a0) at allpaths.c:380
380         break;

执行计划如下:

testdb=# explain verbose select t1.*,t2.dwbh,t2.counter from t_dwxx t1,(select dwbh,count(*) as counter from t_grxx group by dwbh) t2                                           
where t1.dwbh = t2.dwbh and t1.dwbh = '1001';
                                               QUERY PLAN                                               
--------------------------------------------------------------------------------------------------------
 Nested Loop  (cost=0.58..28.65 rows=10 width=32)
   Output: t1.dwmc, t1.dwbh, t1.dwdz, t_grxx.dwbh, (count(*))
   ->  Index Scan using t_dwxx_pkey on public.t_dwxx t1  (cost=0.29..8.30 rows=1 width=20)
         Output: t1.dwmc, t1.dwbh, t1.dwdz
         Index Cond: ((t1.dwbh)::text = '1001'::text)
   ->  GroupAggregate  (cost=0.29..20.14 rows=10 width=12)
         Output: t_grxx.dwbh, count(*)
         Group Key: t_grxx.dwbh
         ->  Index Only Scan using idx_t_dwxx_grbh on public.t_grxx  (cost=0.29..19.99 rows=10 width=4)
               Output: t_grxx.dwbh
               Index Cond: (t_grxx.dwbh = '1001'::text)
(11 rows)

感谢各位的阅读，以上就是“PostgreSQL的set_base_rel_sizes函数分析”的内容了，经过本文的学习后，相信大家对PostgreSQL的set_base_rel_sizes函数分析这一问题有了更深刻的体会，具体使用情况还需要大家实践验证。这里是编程网，小编将为大家推送更多相关知识点的文章，欢迎关注！