MySQL 9.1.0
Source Code Documentation
join_optimizer.h File Reference

The hypergraph join optimizer takes a query block and decides how to execute it as fast as possible (within a given cost model), based on the idea of expressing the join relations as edges in a hypergraph. More...

Go to the source code of this file.

Functions

AccessPathFindBestQueryPlan (THD *thd, Query_block *query_block)
 The main entry point for the hypergraph join optimizer; takes in a query block and returns an access path to execute it (or nullptr, for error). More...
 
bool FinalizePlanForQueryBlock (THD *thd, Query_block *query_block)
 
void FindSargablePredicates (THD *thd, JoinHypergraph *graph)
 
void EstimateAggregateCost (AccessPath *path)
 
void EstimateMaterializeCost (THD *thd, AccessPath *path)
 

Detailed Description

The hypergraph join optimizer takes a query block and decides how to execute it as fast as possible (within a given cost model), based on the idea of expressing the join relations as edges in a hypergraph.

(See subgraph_enumeration.h for more details on the core algorithm, or FindBestQueryPlan() for more information on overall execution.)

It is intended to eventually take over completely from the older join optimizer based on prefix search (sql_planner.cc and related code), and is nearly feature complete, but is currently in the early stages with a very simplistic cost model and certain limitations. The most notable ones are that we do not support:

  • Hints (except STRAIGHT_JOIN).
  • TRADITIONAL and JSON formats for EXPLAIN (use FORMAT=tree).
  • UPDATE.

There are also have many optimization features it does not yet support; among them:

  • Aggregation through a temporary table.
  • Some range optimizer features (notably MIN/MAX optimization).
  • Materialization of arbitrary access paths (note that nested loop joins against these can enable a limited form of hash join that preserves ordering on the left side).

Function Documentation

◆ EstimateAggregateCost()

void EstimateAggregateCost ( AccessPath path)

◆ EstimateMaterializeCost()

void EstimateMaterializeCost ( THD thd,
AccessPath path 
)

◆ FinalizePlanForQueryBlock()

bool FinalizePlanForQueryBlock ( THD thd,
Query_block query_block 
)

◆ FindBestQueryPlan()

AccessPath * FindBestQueryPlan ( THD thd,
Query_block query_block 
)

The main entry point for the hypergraph join optimizer; takes in a query block and returns an access path to execute it (or nullptr, for error).

It works as follows:

  1. Convert the query block from MySQL's Table_ref structures into a hypergraph (see make_join_hypergraph.h).
  2. Find all legal subplans in the hypergraph, calculate costs for them and create access paths – if there are multiple ways to make a given subplan (e.g. multiple join types, or joining {t1,t2,t3} can be made through either {t1}-{t2,t3} or {t1,t2}-{t3}), keep only the cheapest one. Filter predicates (from WHERE and pushed-down join conditions) are added as soon down as it is legal, which is usually (but not universally) optimal. The algorithm works so that we always see smaller subplans first and then end at the complete join plan containing all the tables in the query block.
  3. Add an access path for non-pushable filter predicates.
  4. Add extra access paths for operations done after the joining, such as ORDER BY, GROUP BY, LIMIT, etc..
  5. Make access paths for the filters in nodes made by #2 (see ExpandFilterAccessPaths()).

Materializing subqueries need some extra care. (These are typically IN subqueries that for whatever reason could not be rewritten to semijoin, e.g. because they have GROUP BY.) The decision on whether to materialize or not needs to be done cost-based, and depends both on the inner and outer query block, so it needs to be done cost-based. (Materializiation gives a high up-front cost, but each execution is cheaper, so it will depend on how many times we expect to execute the subquery and now expensive it is to run unmaterialized.) Following the flow through the different steps:

First of all, these go through a stage known as in2exists, rewriting them from e.g.

WHERE t1_outer.x IN ( SELECT t2.y FROM t2 GROUP BY ... )

to

WHERE EXISTS ( SELECT 1 FROM t2 GROUP BY ... HAVING t2.y = t1_outer.x )

This happens before the join optimizer, and the idea is that the HAVING condition (known as a “created_by_in2exists condition”, possibly in WHERE instead of HAVING) can be attempted pushed down into an index or similar, giving more efficient execution. However, if we want to materialize the subquery, these extra conditions need to be removed before materialization; not only do they give the wrong result, but they can also need to wrong costs and a suboptimal join order.

Thus, whenever we plan such a subquery, we plan it twice; once as usual, and then a second time with all in2exists conditions removed. This gives EstimateFilterCost() precise cost information for both cases, or at least as precise as the cost model itself is. In the outer query block, we can then weigh the two alternatives against each other when we add a filter with such a subquery; we can choose to materialize it or not, and propose both alternatives as with any other subplan. When we've decided on the final plan, we go through all access paths and actually materialize the subqueries it says to materialize.

There are lots of places these conditions can show up; to reduce complexity, we only consider materialization in the most common places (filters on base tables, filters after joins, filters from HAVING) – in particular, we don't bother checking on join conditions. It is never wrong to not materialize a subquery, though it may be suboptimal.

Note that the access path returned by FindBestQueryPlan() is not ready for immediate conversion to iterators; see FinalizePlanForQueryBlock(). You may call FindBestQueryPlan() any number of times for a query block, but FinalizePlanForQueryBlock() only once, as finalization generates temporary tables and may rewrite expressions in ways that are incompatible with future planning. The difference is most striking with the planning done twice by in2exists (see above).

Parameters
thdThread handle.
query_blockThe query block to find a plan for.

◆ FindSargablePredicates()

void FindSargablePredicates ( THD thd,
JoinHypergraph graph 
)