In MariaDB up to 10.11, the test_if_cheaper_ordering() code (that tries
to optimizer how GROUP BY is executed) assumes that if a table scan is used
then if there is any index usable by GROUP BY it will be used.
The reason MySQL 10.4 provides a better plan is because of two differences:
- Plans using 'ref' has a cost of 1/10 of what it should be (as a
protection against table scans). This is why 'ref' is used in 10.4
and not in 10.5.
- When 'ref' is used, then GROUP BY will not use an index for GROUP BY.
In MariaDB 10.5 the chosen plan is a table scan (as it calculated to be
faster) but as 'ref' is not used, the test_if_cheaper_ordering()
optimizer phase decides (as ref is not usd) to use an index for GROUP BY,
which has bad performance.
Description of fix:
- All new code is protected by the "optimizer_adjust_secondary_key_costs"
variable, which is now a bit map, and is only executed if the option
"disable_forced_index_in_group_by" set.
- Corrects GROUP BY handling in test_if_cheaper_ordering() by making
the choise of using and index with GROUP BY cost based instead of rule
based.
- Adds TIME_FOR_COMPARE to all costs, when using group by, to make
read_time, index_scan_time and range_cost comparable.
Other things:
- Made optimizer_adjust_secondary_key_costs a bit map (compatible with old
code).
Notes:
Current code ignores costs for the algorithm used when doing GROUP
BY on the first table:
- Create an in-memory temporary table for handling group by and doing a
filesort of the result file
We can probably in 10.6 continue to ignore this cost.
This patch should NOT be merged to 11.0 series (not needed in 11.0).
When executing a statement of the form
SELECT AGGR_FN(DISTINCT c1, c2,..,cn) FROM t1,
where AGGR_FN is an aggregate function such as COUNT(), AVG() or SUM(),
and a unique index exists on table t1 covering some or all of the
columns (c1, c2,..,cn), the retrieved values are inherently unique.
Consequently, the need for de-duplication imposed by the DISTINCT
clause can be eliminated, leading to optimization of aggregation
operations.
This optimization applies under the following conditions:
- only one table involved in the join (not counting const tables)
- some arguments of the aggregate function are fields
(not functions/subqueries)
This optimization extends to queries of the form
SELECT AGGR_FN(c1, c2,..,cn) GROUP BY cx,..cy
when a unique index covers some or all of the columns
(c1, c2,..cn, cx,..cy)
This patch fixes the issue with passing the DEFAULT or IGNORE values to
positional parameters for some kind of SQL statements to be executed
as prepared statements.
The main idea of the patch is to associate an actual value being passed
by the USING clause with the positional parameter represented by
the Item_param class. Such association must be performed on execution of
UPDATE statement in PS/SP mode. Other corner cases that results in
server crash is on handling CREATE TABLE when positional parameter
placed after the DEFAULT clause or CALL statement and passing either
the value DEFAULT or IGNORE as an actual value for the positional parameter.
This case is fixed by checking whether an error is set in diagnostics
area at the function pack_vcols() on return from the function pack_expression()
optimizer-adjust_secondary_key_costs is added to provide 2 small
adjustments to the 10.x optimizer cost model. This can be used in the
case where the optimizer wrongly uses a secondary key instead of a
clustered primary key.
The reason behind this change is that MariaDB 10.x does not take into
account that for engines like InnoDB, that scanning a primary key can be
up to 7x faster than scanning a secondary key + read the row data trough
the primary key.
The different values for optimizer_adjust_secondary_key_costs are:
optimizer_adjust_secondary_key_costs=0
- No changes to current model
optimizer_adjust_secondary_key_costs=1
- Ensure that the cost of of secondary indexes has a cost of at
least 5x times the cost of a clustered primary key (if one exists).
This disables part of the worst_seek optimization described below.
optimizer_adjust_secondary_key_costs=2
- Disable "worst_seek optimization" and adjust filter cost slightly
(add cost of 1 if filter is used).
The idea behind 'worst_seek optimization' is that we limit the
cost for all non clustered ref access to the least of:
- best-rows-by-range (or all rows in no range found) / 10
- scan-time-table (roughly number of file blocks to scan table) * 3
In addition we also do not try to use rowid_filter if number of rows
estimated for 'ref' access is less than the worst_seek limitation.
The idea is that worst_seek is trying to take into account that if
we do a lot of accesses through a key, this is likely to be cached.
However it only does this for secondary keys, and not for clustered
keys or index only reads.
The effect of the worst_seek are:
- In some cases 'ref' will have a much lower cost than range or using
a clustered key.
- Some possible rowid filters for secondary keys will be ignored.
When implementing optimizer_adjust_secondary_key_costs=2, I noticed
that there is a slightly different costs for how ref+filter and
range+filter are calculated. This caused a lot of range and
range+filter to change to ref+filter, which is not good as
range+filter provides the optimizer a better estimate of how many
accepted rows there will be in the result set.
Adding a extra small cost (1 seek) when using filter mitigated the
above problems in almost all cases.
This patch should not be applied to MariaDB 11.0 as worst_seeks is
removed in 11.0 and the cost calculation for clustered keys, secondary
keys, index scan and filter is more exact.
Test case changes for --optimizer-adjust_secondary_key_costs=1
(Fix secondary key costs to be 5x of primary key):
- stat_tables_innodb:
- Complex change (probably ok as number of rows are really small)
- ref over 1 row changed to range over 10 rows with join buffer
- ref over 5 rows changed to eq_ref
- secondary ref over 1 row changed to ref of primary key over 4 rows
- Change of key to use longer key with index pushdown (a little
bit worse but not significant).
- Change to use secondary (1 row) -> primary (4 rows)
- rowid_filter_innodb:
- index_merge (2 rows) & ref (1) -> all (23 rows) -> primary eq_ref.
Test case changes for --optimizer-adjust_secondary_key_costs=2
(remove of worst_seeks & adjust filter cost):
- stat_tables_innodb:
- Join order change (probably ok as number of rows are really small)
- ref (5 rows) & ref(1 row) changed to range (10 rows & join buffer)
& eq_ref.
- selectivity_innodb:
- ref -> ref|filter (ok)
- rowid_filter_innodb:
- ref -> ref|filter (ok)
- range|filter (64 rows) changed to ref|filter (128 rows).
ok as ref|filter outputs wrong number of rows in explain.
- range, range_mrr_icp:
-ref (500 rows -> ALL (1000 rows) (ok)
- select_pkeycache, select, select_jcl6:
- ref|filter (2 rows) -> ref (2 rows) (ok)
- selectivity:
- ref -> ref_filter (ok)
- range:
- Change of 'filtered' but no stat or plan change (ok)
- selectivity:
- ref -> ref+filter (ok)
- Change of filtered but no plan change (ok)
- join_nested_jcl6:
- range -> ref|filter (ok as only 2 rows)
- subselect3, subselect3_jcl6:
- ref_or_null (4 rows) -> ALL (10 rows) (ok)
- Index_subquery (4 rows) -> ALL (10 rows) (ok)
- partition_mrr_myisam, partition_mrr_aria and partition_mrr_innodb:
- Uses ALL instead of REF for a key value that is the same for > 50%
of rows. (good)
order_by_innodb:
- range (200 rows) -> ref (20 rows)+filesort (ok)
- subselect_sj2_mat:
- One test changed. One ALL removed and replaced with eq_ref. Likely
to be better.
- join_cache:
- Changed ref over 60% of the rows to use hash join (ok)
- opt_tvc:
- Changed to use eq_ref instead of ref with plan change (probably ok)
- opt_trace:
- No worst/max seeks clipping (good).
- Almost double range_scan_time and index_scan_time (ok).
- rowid_filter:
- ref -> ref|filtered (ok)
- range|filter (77 rows) changed to ref|filter (151 rows). Proably
ok as ref|filter outputs wrong number of rows in explain.
Reviewer: Sergei Petrunia <sergey@mariadb.com>
In the case of calcuating cost for a ref access for which there exists
a usable range, the variable keyread_tmp would always be 0.
If there is also another index that could be used as a filter, the cost
of that filter would be wrong.
In many cases 'the worst_seeks optimzation' would disable the filter
from getting used, which is why this bug has not been noticed before.
The next commit, which allows one to disable worst_seeks, will have a
test case for this bug.
Statements affect by this bug are all SQL statements that
1) prefixed with "EXPLAIN"
2) have a lower level join structure created for a union subquery.
A bug in select_describe() passed an incorrect "result" object to
mysql_explain_union(), resulting in unpredictable behaviour and
out of context calls.
Reviewed by: Oleksandr Byelkin, sanja@mariadb.com
Enable unusable key notes for non-equality predicates:
<, <=, =>, >, BETWEEN, IN, LIKE
Note, in some scenarios it displays duplicate notes, e.g.
for queries with ORDER BY:
SELECT * FROM t1
WHERE indexed_string_column >= 10
ORDER BY indexed_string_column
LIMIT 5;
This should be tolarable. Getting rid of the diplicate note
completely would need a much more complex patch, which is
not desiable in 10.6.
Details:
- Changing RANGE_OPT_PARAM::note_unusable_keys from bool
to a new data type Item_func::Bitmap, so the caller can
choose with a better granuality which predicates
should raise unusable key notes inside the range optimizer:
a. all predicates (=, <=>, <, <=, =>, >, BETWEEN, IN, LIKE)
b. all predicates except equality (=, <=>)
c. none of the predicates
"b." is needed because in some scenarios equality predicates (=, <=>)
send unusable key notes at an earlier stage, before the range optimizer,
during update_ref_and_keys(). Calling the range optimizer with
"all predicates" would produce duplicate notes for = and <=> in such cases.
- Fixing get_quick_record_count() to call the range optimizer
with "all predicates except equality" instead of "none of the predicates".
Before this change the range optimizer suppressed all notes for
non-equality predicates: <, <=, =>, >, BETWEEN, IN, LIKE.
This actually fixes the reported problem.
- Fixing JOIN::make_range_rowid_filters() to call the range optimizer
with "all predicates except equality" instead of "all predicates".
Before this change the range optimizer produced duplicate notes
for = and <=> during a rowid_filter optimization.
- Cleanup:
Adding the op_collation argument to Field::raise_note_cannot_use_key_part()
and displaying the operation collation rather than the argument collation
in the unusable key note. This is important for operations with more than
two arguments: BETWEEN and IN, e.g.:
SELECT * FROM t1
WHERE column_utf8mb3_general_ci
BETWEEN 'a' AND 'b' COLLATE utf8mb3_unicode_ci;
SELECT * FROM t1
WHERE column_utf8mb3_general_ci
IN ('a', 'b' COLLATE utf8mb3_unicode_ci);
The note for 'a' now prints utf8mb3_unicode_ci as the collation.
which is the collation of the entire operation:
Cannot use key key1 part[0] for lookup:
"`column_utf8mb3_general_ci`" of collation `utf8mb3_general_ci` >=
"'a'" of collation `utf8mb3_unicode_ci`
Before this change it printed the collation of 'a',
so the note was confusing:
Cannot use key key1 part[0] for lookup:
"`column_utf8mb3_general_ci`" of collation `utf8mb3_general_ci` >=
"'a'" of collation `utf8mb3_general_ci`"
During the 10.5->10.6 merge please use the 10.6 code on conflicts.
This is the 10.5 version of the patch (a backport of the 10.6 version).
Unlike 10.6 version, it makes changes in plugin/type_inet/sql_type_inet.*
rather than in sql/sql_type_fixedbin.h
Item_bool_rowready_func2, Item_func_between, Item_func_in
did not check if a not-NULL argument of an arbitrary data type
can produce a NULL value on conversion to INET6.
This caused a crash on DBUG_ASSERT() in conversion failures,
because the function returned SQL NULL for something that
has Item::maybe_null() equal to false.
Adding setting NULL-ability in such cases.
Details:
- Removing the code in Item_func::setup_args_and_comparator()
performing character set aggregation with optional narrowing.
This aggregation is done inside Arg_comparator::set_cmp_func_string().
So this code was redundant
- Removing Item_func::setup_args_and_comparator() as it git simplified to
just to two lines:
convert_const_compared_to_int_field(thd);
return cmp->set_cmp_func(thd, this, &args[0], &args[1], true);
Using these lines directly in:
- Item_bool_rowready_func2::fix_length_and_dec()
- Item_func_nullif::fix_length_and_dec()
- Adding a new virtual method:
- Type_handler::Item_bool_rowready_func2_fix_length_and_dec().
- Adding tests detecting if the data type conversion can return SQL NULL into
the following methods of Type_handler_inet6:
- Item_bool_rowready_func2_fix_length_and_dec
- Item_func_between_fix_length_and_dec
- Item_func_in_fix_comparator_compatible_types
Fix the issue introduced in ec2574fd8f, fix for MDEV-31983:
get_quick_record_count() must set quick_count=0 when it got
IMPOSSIBLE_RANGE from test_quick_select.
Failure to do so will cause an assertion in 11.0, when the number of
quick select rows (0) is checked to be lower than the number of
found_records (which is capped up to 1).
This commit addresses column naming issues with CTEs in the use of prepared
statements and stored procedures. Usage of either prepared statements or
procedures with Common Table Expressions and column renaming may be affected.
There are three related but different issues addressed here.
1) First execution issue. Consider the following
prepare s from "with cte (col1, col2) as (select a as c1, b as c2 from t
order by c1) select col1, col2 from cte";
execute s;
After parsing, items in the select are named (c1,c2), order by (and group by)
resolution is performed, then item names are set to (col1, col2).
When the statement is executed, context analysis is again performed, but
resolution of elements in the order by statement will not be able to find c1,
because it was renamed to col1 and remains this way.
The solution is to save the names of these items during context resolution
before they have been renamed. We can then reset item names back to those after
parsing so first execution can resolve items referred to in order and group by
clauses.
2) Second Execution Issue
When the derived table contains more than one select 'unioned' together we could
reasonably think that dealing with only items in the first select (which
determines names in the resultant table) would be sufficient. This can lead to
a different problem. Consider
prepare st from "with cte (c1,c2) as
(select a as col1, sum(b) as col2 from t1 where a > 0 group by col1
union select a as col3, sum(b) as col4 from t2 where b > 2 group by col3)
select * from cte where c1=1";
When the optimizer (only run during the first execution) pushes the outside
condition "c1=1" into every select in the derived table union, it renames the
items to make the condition valid. In this example, this leaves the first item
in the second select named 'c1'. The second execution will now fail 'group by'
resolution.
Again, the solution is to save the names during context analysis, resetting
before subsequent resolution, but making sure that we save/reset the item
names in all the selects in this union.
3) Memory Leak
During parsing Item::set_name() is used to allocate memory in the statement
arena. We cannot use this call during statement execution as this represents
a memory leak. We directly set the item list names to those in the column list
of this CTE (also allocated during parsing).
Approved by Igor Babaev <igor@mariadb.com>
ref->null_rejecting is a key_part_map. we need to check
the bit corresponding to the particular store_key.
Note that there are no store_key objects for const ref parts.
...errors, then failing ASSERT.
UPDATE queries treat warnings as errors. In this case, an invalid
condition "datetime_key_col >= '2012-01'" caused warning-as-error inside
SQL_SELECT::test_quick_select().
The code that called test_quick_select() ignored this error and continued
join optimization. Then it eventually reached a thd->is_error() check
and failed to setup SJ-Materialization which failed an assert.
Fixed this by making SQL_SELECT::test_quick_select() return error in
its return value, and making any code that calls it to check for error
condition and abort the query if the error is returned.
Places in the code that didn't check for errors from
SQL_SELECT::test_quick_select but now do:
- get_quick_record_count() call in make_join_statistics(),
- test_if_skip_sort_order(),
- "Range checked for each record" code.
Extra error handling fixes and commit text wording by Sergei Petrunia,
Reviewed-by: Sergei Petrunia, Oleg Smirnov
test_if_skip_sort_order() should catch the join types JT_EQ_REF,
JT_CONST and JT_SYSTEM and skip sort order for these.
Such join types imply retrieving of a single row of data, and sorting
of a single row can always be skipped.
(Variant#3: Allow cross-charset comparisons, use a special
CHARSET_INFO to create lookup keys. Review input addressed.)
Equalities that compare utf8mb{3,4}_general_ci strings, like:
WHERE ... utf8mb3_key_col=utf8mb4_value (MB3-4-CMP)
can now be used to construct ref[const] access and also participate
in multiple-equalities.
This means that utf8mb3_key_col can be used for key-lookups when
compared with an utf8mb4 constant, field or expression using '=' or
'<=>' comparison operators.
This is controlled by optimizer_switch='cset_narrowing=on', which is
OFF by default.
IMPLEMENTATION
Item value comparison in (MB3-4-CMP) is done using utf8mb4_general_ci.
This is valid as any utf8mb3 value is also an utf8mb4 value.
When making index lookup value for utf8mb3_key_col, we do "Charset
Narrowing": characters that are in the Basic Multilingual Plane (=BMP) are
copied as-is, as they can be represented in utf8mb3. Characters that are
outside the BMP cannot be represented in utf8mb3 and are replaced
with U+FFFD, the "Replacement Character".
In utf8mb4_general_ci, the Replacement Character compares as equal to any
character that's not in BMP. Because of this, the constructed lookup value
will find all index records that would be considered equal by the original
condition (MB3-4-CMP).
Approved-by: Monty <monty@mariadb.org>
Problem was that JOIN_TAB::cleanup() was not run because
JOIN::top_join_tab_count was not set in case of early errors.
Fixed by setting JOIN::tab_join_tab_count when JOIN_TAB's are allocated.
Something that should eventually be fixed:
- Cleaning up JOIN_TAB's is now done in 3 different loops.
JOIN_TAB::cleanup() is only doing a partial cleanup. Other cleanups
are done outside of JOIN_TAB::cleanup().
The above should be fixed so that JOIN_TAB::cleanup() is freeing
everything related to it's own memory, including all its sub JOIN_ TAB's.
JOIN::cleanup() should only loop over all it's top JOIN_TAB's and call
JOIN_TAB::cleanup() on these.
This will greatly simplify and speedup the current code (as we now do some
cleanup's twice).
The MDEV-29693 conflict resolution is from Monty, as well as is
a bug fix where ANALYZE TABLE wrongly built histograms for
single-column PRIMARY KEY.
Also includes a fix for safe_malloc error reporting.
Other things:
- Copied main.log_slow from 10.4 to avoid mtr issue
Disabled test:
- spider/bugfix.mdev_27239 because we started to get
+Error 1429 Unable to connect to foreign data source: localhost
-Error 1158 Got an error reading communication packets
- main.delayed
- Bug#54332 Deadlock with two connections doing LOCK TABLE+INSERT DELAYED
This part is disabled for now as it fails randomly with different
warnings/errors (no corruption).
The problem was that we did not handle errors properly in
JOIN::get_best_combination. In case an early error, JOIN->join_tab would
contain unintialized values, which would cause errors on cleanup().
The error in question was reported earlier, but not noticed until later.
One cause of this is that most of the sql_select.cc code just checks
thd->fatal_error and not thd->is_error().
Fixed by changing of checks of fatal_error to is_error().
Raise notes if indexes cannot be used:
- in case of data type or collation mismatch (diferent error messages).
- in case if a table field was replaced to something else
(e.g. Item_func_conv_charset) during a condition rewrite.
Added option to write warnings and notes to the slow query log for
slow queries.
New variables added/changed:
- note_verbosity, with is a set of the following options:
basic - All old notes
unusable_keys - Print warnings about keys that cannot be used
for select, delete or update.
explain - Print unusable_keys warnings for EXPLAIN querys.
The default is 'basic,explain'. This means that for old installations
the only notable new behavior is that one will get notes about
unusable keys when one does an EXPLAIN for a query. One can turn all
of all notes by either setting note_verbosity to "" or setting sql_notes=0.
- log_slow_verbosity has a new option 'warnings'. If this is set
then warnings and notes generated are printed in the slow query log
(up to log_slow_max_warnings times per statement).
- log_slow_max_warnings - Max number of warnings written to
slow query log.
Other things:
- One can now use =ALL for any 'set' variable to set all options at once.
For example using "note_verbosity=ALL" in a config file or
"SET @@note_verbosity=ALL' in SQL.
- mysqldump will in the future use @@note_verbosity=""' instead of
@sql_notes=0 to disable notes.
- Added "enum class Data_type_compatibility" and changing the return type
of all Field::can_optimize*() methods from "bool" to this new data type.
Reviewer & Co-author: Alexander Barkov <bar@mariadb.com>
- The code that prints out the notes comes mainly from Alexander
The fix is to return 3-state value from Range_rowid_filter::build()
call:
1. The filter was built successfully;
2. The filter was not built, but the error was not fatal, i.e. there is
no need to rollback transaction. For example, if the size of container to
storevrow ids is not enough;
3. The filter was not built because of fatal error, for example,
deadlock or lock wait timeout from storage engine. In this case we
should stop query plan execution and roll back transaction.
Reviewed by: Sergey Petrunya
Before this patch, the code in Item_field::print() used
this convention (described in sql_explain.h:ExplainDataStructureLifetime):
- By default, the table that Item_field refers to is accessible.
- ANALYZE and SHOW {EXPLAIN|ANALYZE} may print Items after some
temporary tables have been dropped. They use
QT_DONT_ACCESS_TMP_TABLES flag. When it is ON, Item_field::print
will not access the table it refers to, if it is a temp.table
The bug was that EXPLAIN statement also may compute subqueries (depending
on subquery context and @@expensive_subquery_limit setting). After the
computation, the subquery calls JOIN::cleanup(true) which drops some of
its temporary tables. Calling Item_field::print() that refer to such table
will cause an access to free'd memory.
In this patch, we take into account that query optimization can compute
a subquery and discard its temporary tables. Item_field::print() now
assumes that any temporary table might have already been dropped.
This means QT_DONT_ACCESS_TMP_TABLES flag is not needed - we imply it is
always present.
But we also make one exception: derived tables are not freed in
JOIN::cleanup() call. They are freed later in close_thread_tables(),
at the same time when regular tables are closed.
Because of that, Item_field::print may assume that temp.tables
representing derived tables are available.
Initial patch by: Rex Jonston
Reviewed by: Monty <monty@mariadb.org>
The problem is that the first execution of the prepared statement makes
a permanent optimization of converting the LEFT JOIN to an INNER JOIN.
This is based on the assumption that all the user parameters (?) are
always constants and that parameters to Item_cond() will not change value
from true and false between different executions.
(The example was using IS NULL, which will change value if parameter
depending on if the parameter is NULL or not).
The fix is to change Item_cond::fix_fields() and
Item_cond::eval_not_null_tables() to not threat user parameters as
constants. This will ensure that we don't do the LEFT_JOIN -> INNER
JOIN conversion that causes problems.
There is also some things that needs to be improved regarding
calculations of not_null_tables_cache as we get a different value for
WHERE 1 or t1.a=1
compared to
WHERE t1.a= or 1
Changes done:
- Mark Item_param with the PARAM flag to be able to quickly check
in Item_cond::eval_not_null_tables() if an item contains a
prepared statement parameter (just like we check for stored procedure
parameters).
- Fixed that Item_cond::not_null_tables_cache is not depending on
order of arguments.
- Don't call item->eval_const_cond() for items that are NOT on the top
level of the WHERE clause. This removed a lot of unnecessary
warnings in the test suite!
- Do not reset not_null_tables_cache for not top level items.
- Simplified Item_cond::fix_fields by calling eval_not_null_tables()
instead of having duplication of all the code in
eval_not_null_tables().
- Return an error if Item_cond::fix_field() generates an error
The old code did generate an error in some cases, but not in all
cases.
- Fixed all handling of the above error in make_cond_for_tables().
The error handling by the callers did not exists before which
could lead to asserts in many different places in the old code).
- All changes in sql_select.cc are just checking the return value of
fix_fields() and make_cond_for_tables() and returning an error
value if fix_fields() returns true or make_cond_for_tables()
returns NULL and is_error() is set.
- Mark Item_cond as const_item if all arguments returns true for
can_eval_in_optimize().
Reviewer: Sergei Petrunia <sergey@mariadb.com>
After MDEV-21580 the truncation of SORT_FIELD::length
set_if_smaller(sortorder->length, thd->variables.max_sort_length)
became conditional:
if (is_variable_sized())
set_if_smaller(length, thd->variables.max_sort_length)
To provide correct functioning of is_variable_sized() SORT_FIELD::type
must be set properly. This commit adds the necessary initialization
of SORT_FIELD::type to JOIN_TAB::remove_duplicates() as it is done
in filesort's sortlength() function.
DBUG_ASSERT is added to sortlength() just in case to prevent
a possible uint32 overflow
