mariadb/sql/opt_group_by_cardinality.cc
Sergei Petrunia ef966af801 MDEV-30877: Output cardinality for derived table ignores GROUP BY
(Variant 3) (commit in 11.4)
When a derived table has a GROUP BY clause:

  SELECT ...
    FROM  (SELECT ... GROUP BY col1, col2) AS tbl

The optimizer would use inner join's output cardinality as an estimate
of derived table size, ignoring the fact that GROUP BY operation would
produce much fewer groups.

Add code to produce tighter bounds:
- The GROUP BY list is split into per-table lists. If GROUP BY list has
  expressions that refer to multiple tables, we fall back to join output
  cardinality.
- For each table, the first cardinality estimate is join_tab->read_records.
- Then, we try to get a tighter bound by using index statistics.
- If indexes do not cover all GROUP BY columns, we try to use per-column
  EITS statistics.
2025-02-10 22:06:49 +02:00

376 lines
11 KiB
C++

/*
Copyright (c) 2025, MariaDB
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; version 2 of the License.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program; if not, write to the Free Software
Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1335 USA */
/**
@file
Contains estimate_post_group_cardinality() which estimates cardinality
after GROUP BY operation is applied.
*/
#include "mariadb.h"
#include "sql_priv.h"
#include "sql_select.h"
#include "sql_statistics.h"
#include "opt_trace.h"
static
double estimate_table_group_cardinality(JOIN *join, Item ***group_list,
Item* const *end);
inline bool has_one_bit_set(table_map val)
{
return val && !(val & (val-1));
}
/*
@brief
Sort the Items that refer to one table (so have only one bit in
used_tables()). Used to get the items that refer to the same table
to be next to each other.
*/
int cmp_items_by_used_tables(const void *a_val, const void *b_val)
{
table_map v1= (*((Item**)a_val))->used_tables();
table_map v2= (*((Item**)b_val))->used_tables();
return v1 > v2 ? 1 : (v1 < v2 ? -1 : 0);
}
/*
@brief
Given a SELECT with GROUP BY clause, estimate the cardinality of output
after the grouping operation is performed.
@detail
Consider a query
SELECT ...
FROM t1, t2, t3 ...
WHERE ...
GROUP BY
col1, col2, ...
Join optimizer produces an estimate of number of record combinations we'll
get after all join operations are performed (denote this join_output_card).
This function produces a conservative (i.e. upper bound) estimate of how
many groups will be produced by the GROUP BY operation.
It does it as follows:
* Split the GROUP BY clause into per-table lists.
(if there are GROUP BY items that refer to multiple tables, refuse
to work and return join_output_card).
* Compute n_groups estimate for each table and its GROUP BY sub-list.
* Compute a product of these estimates, n_groups_prod.
* Return MIN(join_ouput_card, n_groups_prod).
@param
join_output_card Number of rows after join operation
@return
Number of rows that will be left after grouping operation
*/
double estimate_post_group_cardinality(JOIN *join, double join_output_card)
{
Dynamic_array<Item*> group_cols(join->thd->mem_root);
ORDER *cur_group;
Json_writer_object wrapper(join->thd);
Json_writer_object trace(join->thd, "materialized_output_cardinality");
trace.add("join_output_cardinality", join_output_card);
/*
Walk the GROUP BY list and put items into group_cols array. Array is
easier to work with: we will sort it and then produce estimates for
sub-arrays that refer to just one table.
Also check that each item depends on just one table (if not, bail out).
*/
for (cur_group= join->group_list; cur_group; cur_group= cur_group->next)
{
Item *item= *cur_group->item;
table_map map= item->used_tables();
if ((map & PSEUDO_TABLE_BITS) || !has_one_bit_set(map))
{
/* Can't estimate */
return join_output_card;
}
group_cols.append(item);
}
DBUG_ASSERT(group_cols.size());
group_cols.sort(cmp_items_by_used_tables);
double new_card= 1.0;
Item **pos= group_cols.front();
Json_writer_array trace_steps(join->thd, "estimation");
while (pos != group_cols.end())
{
new_card *= estimate_table_group_cardinality(join, &pos, group_cols.end());
if (new_card > join_output_card)
return join_output_card;
}
trace_steps.end();
trace.add("post_group_cardinality", new_card);
return new_card;
}
/*
@brief
Compute number of groups for a GROUP BY list that refers to a single table
@detail
Consider a query:
SELECT ...
FROM t1, t2, t3 ...
WHERE ...
GROUP BY
t1.col1, ... t1.colN -- expressions only refer to t1.
The number of groups is estimated using the following:
== 1. Use found_records ==
There cannot be more rows than the number of records in t1 that match the
WHERE clause, that is, JOIN_TAB(t1)->found_records.
This estimate doesn't depend on the expressions in the GROUP BY list, so we
use it as a fall-back estimate.
== 2. Use index statistics ==
If t1 has an INDEX(col1, ... colN) then the number of different
combinations of {col1, ..., colN} can be obtained from index statistics.
It is possible to cover the GROUP BY list with several indexes (without
overlaps) and use a product of n_distinct statistics. For example, for
GROUP BY key1part1, key1part2, key2part1, key2part2, key2part3
the estimate would be:
n_groups= n_distinct(key1, parts=2) * n_distinct(key2, parts=3)
There can be multiple ways one can cover GROUP BY list with different
indexes. We try to use indexes that cover more GROUP BY columns, first.
This may cause us to fail later. For example, for
GROUP BY a, b, c, d
and indexes
INDEX idx1(a,b,c)
INDEX idx2(a,b)
INDEX idx3(c,d)
We will use idx1 and then will be unable to get any estimate for column d.
We could have used idx2 and idx3, instead, and could have covered all
columns. We ignore such cases.
Note that when using index statistics, we ignore the WHERE condition
selectivity. That's because we cannot tell how the WHERE affects index
stats. Does it
A. reduce the number of GROUP BY groups, or
B. make each GROUP BY group smaller ?
We conservatively assume that B holds.
== 3 Use per-column EITS statistics ==
If we fail to cover GROUP BY with indexes, we try to use column statistics
for the remaining columns.
@param join the Join object we're computing for
@param group_list INOUT Array of Item* from GROUP BY clause, ordered
by table. This function should process the table
it is pointing to, and advance the pointer so it
points at 'end' or at the next table.
@param end IN End of the above array.
*/
double estimate_table_group_cardinality(JOIN *join, Item ***group_list,
Item* const *end)
{
TABLE *table= NULL;
key_map possible_keys;
Dynamic_array<int> columns(join->thd->mem_root);
double card= 1.0;
double table_records_after_where= DBL_MAX; // Safety
table_map table_bit= (**group_list)->used_tables();
/*
join->map2table is not set yet, so find our table in JOIN_TABs.
*/
for (JOIN_TAB *tab= join->join_tab;
tab < join->join_tab + join->top_join_tab_count;
tab++)
{
if (tab->table->map == table_bit)
{
table= tab->table;
table_records_after_where= rows2double(tab->found_records);
break;
}
}
DBUG_ASSERT(table);
Json_writer_object trace_obj(join->thd);
trace_obj.add_table_name(table);
Json_writer_array trace_steps(join->thd, "steps");
possible_keys.clear_all();
bool found_complex_item= false;
/*
Walk through the group list and collect references to fields.
If there are other kinds of items, return table's cardinality.
*/
Item **p;
for (p= *group_list;
p != end && (*p)->used_tables() == table_bit;
p++)
{
Item *real= (*p)->real_item();
if (real->type() == Item::FIELD_ITEM)
{
Field *field= ((Item_field*)real)->field;
possible_keys.merge(field->part_of_key);
columns.append(field->field_index);
}
else
found_complex_item= true;
}
/* Tell the caller where group_list ended */
*group_list= p;
if (found_complex_item)
goto whole_table;
possible_keys.intersect(table->keys_in_use_for_query);
/*
Ok, group_list has only columns and we've got them in 'columns'.
*/
while (!possible_keys.is_clear_all())
{
/* Find the index which has the longest prefix covered by columns. */
uint longest_key= UINT_MAX;
int longest_len= 0;
key_map::Iterator key_it(possible_keys);
uint key;
while ((key= key_it++) != key_map::Iterator::BITMAP_END)
{
const KEY *keyinfo= table->key_info + key;
/* Find the length of index prefix covered by GROUP BY columns */
int part;
for (part= 0; part < (int)keyinfo->usable_key_parts; part++)
{
uint field_index= keyinfo->key_part[part].field->field_index;
if (columns.find_first(field_index) == columns.NOT_FOUND)
break;
}
if (part > 0) // At least one column is covered
{
/* Make sure the index has statistics available */
if (!keyinfo->actual_rec_per_key(part - 1))
{
possible_keys.clear_bit(key);
continue;
}
if (part > longest_len)
{
longest_len= part;
longest_key= key;
}
}
else
{
/*
The index can't cover even one-column prefix. Remove it from
consideration.
*/
possible_keys.clear_bit(key);
}
}
if (longest_key == UINT_MAX)
break; // No indexes are usable, stop.
possible_keys.clear_bit(longest_key);
/* Multiply cardinality by index prefix's cardinality */
const KEY *keyinfo= table->key_info + longest_key;
double index_card= (rows2double(table->stat_records()) /
keyinfo->actual_rec_per_key(longest_len-1));
/* Safety in case of inconsistent statistics: */
set_if_bigger(index_card, 1.0);
Json_writer_object trace_idx(join->thd);
trace_idx.add("index_name", keyinfo->name)
.add("cardinality", index_card);
card *= index_card;
if (card > table_records_after_where)
goto whole_table;
/* Remove the columns we've handled from consideration */
for (int part= 0; part < longest_len; part++)
{
uint field_index= keyinfo->key_part[part].field->field_index;
size_t idx= columns.find_first(field_index);
if (idx != columns.NOT_FOUND)
columns.del(idx);
else
DBUG_ASSERT(0); // Can't happen, we've found it above.
}
if (!columns.size())
break; // If we've covered all columns, stop.
}
/*
If there are some columns left for which we couldn't get cardinality
from index statistics, try getting it from columns' histograms
*/
for (size_t i=0; i < columns.size(); i++)
{
double freq;
Field *field= table->field[columns.at(i)];
if (!field->read_stats ||
(freq= field->read_stats->get_avg_frequency()) == 0.0)
goto whole_table;
double column_card= rows2double(table->stat_records()) / freq;
Json_writer_object trace_col(join->thd);
trace_col.add("column", field->field_name)
.add("cardinality", column_card);
card *= column_card;
if (card > table_records_after_where)
goto whole_table;
}
normal_exit:
trace_steps.end();
trace_obj.add("cardinality", card);
return card;
whole_table:
card= table_records_after_where;
goto normal_exit;
}