MySQL 8.0.29
Source Code Documentation
Go to the documentation of this file.
1/* Copyright (c) 2020, 2022, Oracle and/or its affiliates.
3 This program is free software; you can redistribute it and/or modify
4 it under the terms of the GNU General Public License, version 2.0,
5 as published by the Free Software Foundation.
7 This program is also distributed with certain software (including
8 but not limited to OpenSSL) that is licensed under separate terms,
9 as designated in a particular file or component or in included license
10 documentation. The authors of MySQL hereby grant you an additional
11 permission to link the program and your derivative works with the
12 separately licensed software that they have included with MySQL.
14 This program is distributed in the hope that it will be useful,
15 but WITHOUT ANY WARRANTY; without even the implied warranty of
17 GNU General Public License, version 2.0, for more details.
19 You should have received a copy of the GNU General Public License
20 along with this program; if not, write to the Free Software
21 Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA */
27 @file
29 The hypergraph join optimizer takes a query block and decides how to
30 execute it as fast as possible (within a given cost model), based on
31 the idea of expressing the join relations as edges in a hypergraph.
32 (See subgraph_enumeration.h for more details on the core algorithm,
33 or FindBestQueryPlan() for more information on overall execution.)
35 It is intended to eventually take over completely from the older join
36 optimizer based on prefix search ( and related code),
37 and is nearly feature complete, but is currently in the early stages
38 with a very simplistic cost model and certain limitations.
39 The most notable ones are that we do not support:
41 - Hints (except STRAIGHT_JOIN).
42 - TRADITIONAL and JSON formats for EXPLAIN (use FORMAT=tree).
43 - UPDATE.
45 There are also have many optimization features it does not yet support;
46 among them:
48 - Aggregation through a temporary table.
49 - Some range optimizer features (notably MIN/MAX optimization).
50 - Materialization of arbitrary access paths (note that nested loop
51 joins against these can enable a limited form of hash join
52 that preserves ordering on the left side).
53 */
55#include <string>
57class Query_block;
58class THD;
59struct AccessPath;
60struct JoinHypergraph;
61struct TABLE;
64 The main entry point for the hypergraph join optimizer; takes in a query
65 block and returns an access path to execute it (or nullptr, for error).
66 It works as follows:
68 1. Convert the query block from MySQL's TABLE_LIST structures into
69 a hypergraph (see make_join_hypergraph.h).
70 2. Find all legal subplans in the hypergraph, calculate costs for
71 them and create access paths -- if there are multiple ways to make a
72 given subplan (e.g. multiple join types, or joining {t1,t2,t3} can be
73 made through either {t1}-{t2,t3} or {t1,t2}-{t3}), keep only the cheapest
74 one. Filter predicates (from WHERE and pushed-down join conditions)
75 are added as soon down as it is legal, which is usually (but not
76 universally) optimal. The algorithm works so that we always see smaller
77 subplans first and then end at the complete join plan containing all the
78 tables in the query block.
79 3. Add an access path for non-pushable filter predicates.
80 4. Add extra access paths for operations done after the joining,
81 such as ORDER BY, GROUP BY, LIMIT, etc..
82 5. Make access paths for the filters in nodes made by #2
83 (see ExpandFilterAccessPaths()).
85 Materializing subqueries need some extra care. (These are typically IN
86 subqueries that for whatever reason could not be rewritten to semijoin,
87 e.g. because they have GROUP BY.) The decision on whether to materialize
88 or not needs to be done cost-based, and depends both on the inner and outer
89 query block, so it needs to be done cost-based. (Materializiation gives
90 a high up-front cost, but each execution is cheaper, so it will depend on
91 how many times we expect to execute the subquery and now expensive it is
92 to run unmaterialized.) Following the flow through the different steps:
94 First of all, these go through a stage known as in2exists, rewriting them
95 from e.g.
97 WHERE t1_outer.x IN ( SELECT t2.y FROM t2 GROUP BY ... )
99 to
101 WHERE EXISTS ( SELECT 1 FROM t2 GROUP BY ... HAVING t2.y = t1_outer.x )
103 This happens before the join optimizer, and the idea is that the HAVING
104 condition (known as a “created_by_in2exists condition”, possibly in WHERE
105 instead of HAVING) can be attempted pushed down into an index or similar,
106 giving more efficient execution. However, if we want to materialize the
107 subquery, these extra conditions need to be removed before materialization;
108 not only do they give the wrong result, but they can also need to wrong
109 costs and a suboptimal join order.
111 Thus, whenever we plan such a subquery, we plan it twice; once as usual,
112 and then a second time with all in2exists conditions removed. This gives
113 EstimateFilterCost() precise cost information for both cases, or at least
114 as precise as the cost model itself is. In the outer query block, we can
115 then weigh the two alternatives against each other when we add a filter
116 with such a subquery; we can choose to materialize it or not, and propose
117 both alternatives as with any other subplan. When we've decided on the
118 final plan, we go through all access paths and actually materialize the
119 subqueries it says to materialize.
121 There are lots of places these conditions can show up; to reduce complexity,
122 we only consider materialization in the most common places (filters on
123 base tables, filters after joins, filters from HAVING) -- in particular,
124 we don't bother checking on join conditions. It is never wrong to not
125 materialize a subquery, though it may be suboptimal.
128 Note that the access path returned by FindBestQueryPlan() is not ready
129 for immediate conversion to iterators; see FinalizePlanForQueryBlock().
130 You may call FindBestQueryPlan() any number of times for a query block,
131 but FinalizePlanForQueryBlock() only once, as finalization generates
132 temporary tables and may rewrite expressions in ways that are incompatible
133 with future planning. The difference is most striking with the planning
134 done twice by in2exists (see above).
136 @param thd Thread handle.
137 @param query_block The query block to find a plan for.
138 @param trace If not nullptr, will be filled with human-readable optimizer
139 trace showing some of the inner workings of the code.
140 */
142 std::string *trace);
144// See comment in .cc file.
145bool FinalizePlanForQueryBlock(THD *thd, Query_block *query_block);
147// Exposed for unit testing only.
148void FindSargablePredicates(THD *thd, std::string *trace,
149 JoinHypergraph *graph);
This class represents a query block, aka a query specification, which is a query consisting of a SELE...
Definition: sql_lex.h:1124
For each client connection we create a separate thread with THD serving as a thread/connection descri...
Definition: sql_class.h:945
bool FinalizePlanForQueryBlock(THD *thd, Query_block *query_block)
void FindSargablePredicates(THD *thd, std::string *trace, JoinHypergraph *graph)
void EstimateAggregateCost(AccessPath *path)
void EstimateMaterializeCost(THD *thd, AccessPath *path)
AccessPath * FindBestQueryPlan(THD *thd, Query_block *query_block, std::string *trace)
The main entry point for the hypergraph join optimizer; takes in a query block and returns an access ...
static char * path
Access paths are a query planning structure that correspond 1:1 to iterators, in that an access path ...
Definition: access_path.h:188
A struct containing a join hypergraph of a single query block, encapsulating the constraints given by...
Definition: make_join_hypergraph.h:70
Definition: table.h:1394