Analysis:
When the method JOIN::choose_subquery_plan() decided to apply
the IN-TO-EXISTS strategy, it set the unit and select_lex
uncacheable flag to UNCACHEABLE_DEPENDENT_INJECTED unconditionally.
As result, even if IN-TO-EXISTS injected non-correlated predicates,
the subquery was still treated as correlated.
Solution:
Set the subquery as correlated only if the injected predicate(s) depend
on the outer query.
The patch enables back constant subquery execution during
query optimization after it was disabled during the development
of MWL#89 (cost-based choice of IN-TO-EXISTS vs MATERIALIZATION).
The main idea is that constant subqueries are allowed to be executed
during optimization if their execution is not expensive.
The approach is as follows:
- Constant subqueries are recursively optimized in the beginning of
JOIN::optimize of the outer query. This is done by the new method
JOIN::optimize_constant_subqueries(). This is done so that the cost
of executing these queries can be estimated.
- Optimization of the outer query proceeds normally. During this phase
the optimizer may request execution of non-expensive constant subqueries.
Each place where the optimizer may potentially execute an expensive
expression is guarded with the predicate Item::is_expensive().
- The implementation of Item_subselect::is_expensive has been extended
to use the number of examined rows (estimated by the optimizer) as a
way to determine whether the subquery is expensive or not.
- The new system variable "expensive_subquery_limit" controls how many
examined rows are considered to be not expensive. The default is 100.
In addition, multiple changes were needed to make this solution work
in the light of the changes made by MWL#89. These changes were needed
to fix various crashes and wrong results, and legacy bugs discovered
during development.
- Let fix_semijoin_strategies_for_picked_join_order() set
POSITION::prefix_record_count for POSITION records that it copies from
SJ_MATERIALIZATION_INFO::tables.
(These records do not have prefix_record_count set, because they are optimized
as joins-inside-semijoin-nests, without full advance_sj_state() processing).
Analysis:
The reason for the wrong result is the interaction between constant
optimization (in this case 1-row table) and subquery optimization.
- First the outer query is optimized, and 'make_join_statistics' finds that
table t2 has one row, reads that row, and marks the whole table as constant.
This also means that all fields of t2 are constant.
- Next, we optimize the subquery in the end of the outer 'make_join_statistics'.
The field 'f2' is considered constant, with value '3'. The subquery predicate
is rewritten as the constant TRUE.
- The outer query execution detects early that the whole query result is empty
and calls 'return_zero_rows'. Since the query is with implicit grouping, we
have to produce one row with special values for the aggregates (depending on
each aggregate function), and NULL values for all non-aggregate fields. This
function calls 'no_rows_in_result' to set each aggregate function to the
default value when it aggregates over an empty result, and then calls
'send_data', which in turn evaluates each Item in the SELECT list.
- When evaluation reaches the subquery predicate, it executes the subquery
with field 'f2' having a constant value '3', and the subquery produces the
incorrect result '7'.
Solution:
Implement Item::no_rows_in_result for all subquery predicates. In order to
make this work, it is also needed to make all val_* methods of all subquery
predicates respect the Item_subselect::forced_const flag. Otherwise subqueries
are executed anyways, and override the default value set by no_rows_in_result
with whatever result is produced from the subquery evaluation.
- When doing join optimization, pre-sort the tables so that they mimic the execution
order we've had with 'semijoin=off'.
- That way, we will not get regressions when there are two query plans (the old and the
new) that have indentical costs but different execution times (because of factors that
the optimizer was not able to take into account).
- The problem was with execution strategy for cases where FirstMatch's inner tables
were interleaved with outer-uncorrelated tables.
- I was unable to find any cases where such join orders would be practically useful,
so fixed it by disabling them.
include/mysql_com.h:
remove "shutdown levels" that aren't shutdown levels from mysql_enum_shutdown_level
mysys/my_addr_resolve.c:
my_snprintf in 5.5 (but not in 5.3) supports %p
sql/item_func.cc:
use a method (that exists only in 5.5) instead of directly accessing a member
sql/item_subselect.cc:
use a method (that exists only in 5.5) instead of directly accessing a member
sql/opt_subselect.cc:
use a method (that exists only in 5.5) instead of directly accessing a member
sql/sql_select.cc:
use a method (that exists only in 5.5) instead of directly accessing a member
- Fix equality propagation to work with SJM nests and OR clauses (full descirption of problem and
solution in the comment in the patch)
(The second commit with post-review fixes)
- The problem was that
= we've picked a LooseScan that used full index scan (tab->type==JT_ALL) on certain index.
= there was also a quick select (tab->quick!=NULL), that used other indexes.
= some old code assumes that (tab->type==JT_ALL && tab->quick) -> means that the
quick select should be used, which is not true.
Fixed by discarding the quick select as soon as we know we're using LooseScan
without using the quick select.
- Remove all references of MAX_TABLES from JOIN struct and make these dynamic
- Updated Join_plan_state to allocate just as many elements as it's needed
sql/opt_subselect.cc:
Optimized version of Join_plan_state
sql/sql_select.cc:
Set join->positions and join->best_positions dynamicly
Don't call update_virtual_fields() if table->vfield is not set.
sql/sql_select.h:
Remove all references of MAX_TABLES from JOIN struct and Join_plan_state and make these dynamic
- Avoid needless load/stores in my_hash_sort_simple due to possible aliasing
- Avoid expensive Join_plan_state constructor in choose_subquery_plan when no subquery
- Avoid calling update_virtual_fields for every row when no virtual fields.
- The problem was that convert_subq_to_jtbm() attached the semi-join
TABLE_LIST object into the wrong list: they used to attach it to the
end of parent_lex->leaf_tables.head()->next_local->...->next_local.
This was apparently inccorect, as one can construct an example where
JTBM nest is attached to a table that is inside some mergeable VIEW, which
breaks (causes crash) for name resolution on the subsequent statement
re-execution.
- Solution: Attach to the "right" list. The "wording" was copied from
st_select_lex::handle_derived.
The result of materialization of the right part of an IN subquery predicate
is placed into a temporary table. Each row of the materialized table is
distinct. A unique key over all fields of the temporary table is defined and
created. It allows to perform key look-ups into the table.
The table created for a materialized subquery can be accessed by key as
any other table. The function best_access-path search for the best access
to join a table to a given partial join. With some where conditions this
function considers a possibility of a ref_or_null access. If such access
employs the unique key on the temporary table then when estimating
the cost this access the function tries to use the array rec_per_key. Yet,
such array is not built for this unique key. This causes a crash of the server.
Rows returned by the subquery that contain nulls don't have to be placed
into temporary table, as they cannot be match any row produced by the
left part of the subquery predicate. So all fields of the temporary table
can be defined as non-nullable. In this case any ref_or_null access
to the temporary table does not make any sense and it does not make sense
to estimate such an access.
The fix makes sure that the temporary table for a materialized IN subquery
is defined with columns that are all non-nullable. The also ensures that
any row with nulls returned by the subquery is not placed into the
temporary table.
- If LooseScan is used with quick select, require that quick select produces
data in key order (this disables use of MRR, which can return data in arbitrary order).
- Disable use of join cache when we're using FirstMatch strategy, and the join
order is such that subquery's inner tables are interleaved with outer. Join
buffering code is incapable of handling such join orders.
- The testcase requires use of @@debug_optimizer_prefer_join_prefix to hit the bug,
but I'm pushing it anyway (including the mention of the variable in .test file),
so that it can be found and enabled when/if we get something comparable in the
main tree.
The problem was that LooseScan execution code assumed that tab->key holds
the index used for looseScan. This is only true when range or full index
scan are used. In case of ref access, the index is in tab->ref.key (and
tab->index==0 which explains how LooseScan passed tests with ref access: they
used one index)
Fixed by setting/using loosescan_key, which always the correct index#.
fixed several defects in the greedy optimization:
1) The greedy optimizer calculated the 'compare-cost' (CPU-cost)
for iterating over the partial plan result at each level in
the query plan as 'record_count / (double) TIME_FOR_COMPARE'
This cost was only used locally for 'best' calculation at each
level, and *not* accumulated into the total cost for the query plan.
This fix added the 'CPU-cost' of processing 'current_record_count'
records at each level to 'current_read_time' *before* it is used as
'accumulated cost' argument to recursive
best_extension_by_limited_search() calls. This ensured that the
cost of a huge join-fanout early in the QEP was correctly
reflected in the cost of the final QEP.
To get identical cost for a 'best' optimized query and a
straight_join with the same join order, the same change was also
applied to optimize_straight_join() and get_partial_join_cost()
2) Furthermore to get equal cost for 'best' optimized query and a
straight_join the new code substrcated the same '0.001' in
optimize_straight_join() as it had been already done in
best_extension_by_limited_search()
3) When best_extension_by_limited_search() aggregated the 'best' plan a
plan was 'best' by the check :
'if ((search_depth == 1) || (current_read_time < join->best_read))'
The term '(search_depth == 1' incorrectly caused a new best plan to be
collected whenever the specified 'search_depth' was reached - even if
this partial query plan was more expensive than what we had already
found.
- Let JTBM optimization code handle the case where the subquery is degenerate and doesn't have a
join query plan. Regular materialization would fall back to IN->EXISTS for such cases. Semi-Join
materialization does not have such option, instead we introduce and use "constant JTBM join tabs".
If the duplicate elimination strategy is used for a semi-join and potentially
one of the block-based join algorithms can be employed to join the inner
tables of the semi-join then sorting of the head (first non-constant) table
for a query with ORDER BY / GROUP BY cannot be used.
The function setup_sj_materialization_part1() forgot to set the value
of TABLE::map for any materialized IN subquery.
This could lead to wrong results for queries with subqueries that were
converted to queries with semijoins.
- if we're considering FirstMatch access with one inner table, and
@@optimizer_switch has semijoin_with_cache flag, calculate costs
as if we used join cache (because we will be able to do so)
