Atomic CREATE OR REPLACE allows to keep an old table intact if the
command fails or during the crash. That is done through creating
a table with a temporary name and filling it with the data
(for CREATE OR REPLACE .. SELECT), then renaming the original table
to another temporary (backup) name and renaming the replacement table
to original table. The backup table is kept until the last chance of
failure and if that happens, the replacement table is thrown off and
backup recovered. When the command is complete and logged the backup
table is deleted.
Atomic replace algorithm
Two DDL chains are used for CREATE OR REPLACE:
ddl_log_state_create (C) and ddl_log_state_rm (D).
1. (C) Log CREATE_TABLE_ACTION of TMP table (drops TMP table);
2. Create new table as TMP;
3. Do everything with TMP (like insert data);
finalize_atomic_replace():
4. Link chains: (D) is executed only if (C) is closed;
5. (D) Log DROP_ACTION of BACKUP;
6. (C) Log RENAME_TABLE_ACTION from ORIG to BACKUP (replays BACKUP -> ORIG);
7. Rename ORIG to BACKUP;
8. (C) Log CREATE_TABLE_ACTION of ORIG (drops ORIG);
9. Rename TMP to ORIG;
finalize_ddl() in case of success:
10. Close (C);
11. Replay (D): BACKUP is dropped.
finalize_ddl() in case of error:
10. Close (D);
11. Replay (C):
1) ORIG is dropped (only after finalize_atomic_replace());
2) BACKUP renamed to ORIG (only after finalize_atomic_replace());
3) drop TMP.
If crash happens (C) or (D) is replayed in reverse order. (C) is
replayed if crash happens before it is closed, otherwise (D) is
replayed.
Temporary table for CREATE OR REPLACE
Before dropping "old" table, CREATE OR REPLACE creates "tmp" table.
ddl_log_state_create holds the drop of the "tmp" table. When
everything is OK (data is inserted, "tmp" is ready) ddl_log_state_rm
is written to replace "old" with "tmp". Until ddl_log_state_create
is closed ddl_log_state_rm is not executed.
After the binlogging is done ddl_log_state_create is closed. At that
point ddl_log_state_rm is executed and "tmp" is replaced with
"old". That is: final rename is done by the DDL log.
With that important role of DDL log for CREATE OR REPLACE operation
replay of ddl_log_state_rm must fail at the first hit error and
print the error message if possible. F.ex. foreign key error is
discovered at this phase: InnoDB rejects to drop the "old" table and
returns corresponding foreign key error code.
Additional notes
- CREATE TABLE without REPLACE is not affected by this commit.
- Engines having HTON_EXPENSIVE_RENAME flag set are not affected by
this commit.
- CREATE TABLE .. SELECT XID usage is fixed and now there is no need
to log DROP TABLE via DDL_CREATE_TABLE_PHASE_LOG (see comments in
do_postlock()). XID is now correctly updated so it disables
DDL_LOG_DROP_TABLE_ACTION. Note that binary log is flushed at the
final stage when the table is ready. So if we have XID in the
binary log we don't need to drop the table.
- Three variations of CREATE OR REPLACE handled:
1. CREATE OR REPLACE TABLE t1 (..);
2. CREATE OR REPLACE TABLE t1 LIKE t2;
3. CREATE OR REPLACE TABLE t1 SELECT ..;
- Test case uses 6 combinations for engines (aria, aria_notrans,
myisam, ib, lock_tables, expensive_rename) and 2 combinations for
binlog types (row, stmt). Combinations help to check differences
between the results. Error failures are tested for the above three
variations.
- expensive_rename tests CREATE OR REPLACE without atomic
replace. The effect should be the same as with the old behaviour
before this commit.
- Triggers mechanism is unaffected by this change. This is tested in
create_replace.test.
- LOCK TABLES is affected. Lock restoration must be done after "rm"
chain is replayed.
- Moved ddl_log_complete() from send_eof() to finalize_ddl(). This
checkpoint was not executed before for normal CREATE TABLE but is
executed now.
- CREATE TABLE will now rollback also if writing to the binary
logging failed. See rpl_gtid_strict.test
Rename and drop via DDL log
We replay ddl_log_state_rm to drop the old table and rename the
temporary table. In that case we must throw the correct error
message if ddl_log_revert() fails (f.ex. on FK error).
If table is deleted earlier and not via DDL log and the crash
happened, the create chain is not closed. Linked drop chain is not
executed and the new table is not installed. But the old table is
already deleted.
ddl_log.cc changes
Now we can place action before DDL_LOG_DROP_INIT_ACTION and it will
be replayed after DDL_LOG_DROP_TABLE_ACTION.
report_error parameter for ddl_log_revert() allows to fail at first
error and print the error message if possible.
ddl_log_execute_action() now can print error message.
Since we now can handle errors from ddl_log_execute_action() (in
case of non-recovery execution) unconditional setting "error= TRUE"
is wrong (it was wrong anyway because it was overwritten at the end
of the function).
On XID usage
Like with all other atomic DDL operations XID is used to avoid
inconsistency between master and slave in the case of a crash after
binary log is written and before ddl_log_state_create is closed. On
recovery XIDs are taken from binary log and corresponding DDL log
events get disabled. That is done by
ddl_log_close_binlogged_events().
On linking two chains together
Chains are executed in the ascending order of entry_pos of execute
entries. But entry_pos assignment order is undefined: it may assign
bigger number for the first chain and then smaller number for the
second chain. So the execution order in that case will be reverse:
second chain will be executed first.
To avoid that we link one chain to another. While the base chain
(ddl_log_state_create) is active the secondary chain
(ddl_log_state_rm) is not executed. That is: only one chain can be
executed in two linked chains.
The interface ddl_log_link_chains() was done in "MDEV-22166
ddl_log_write_execute_entry() extension".
More on CREATE OR REPLACE .. SELECT
We use create_and_open_tmp_table() like in ALTER TABLE to create
temporary TABLE object (tmp_table is (NON_)TRANSACTIONAL_TMP_TABLE).
After we created such TABLE object we use create_info->tmp_table()
instead of table->s->tmp_table when we need to check for
parser-requested tmp-table.
External locking is required for temporary table created by
create_and_open_tmp_table(). F.ex. that disables logging for Aria
transactional tables and without that (when no mysql_lock_tables()
is done) it cannot work correctly.
For making external lock the patch requires Aria table to work in
non-transactional mode. That is usually done by
ha_enable_transaction(false). But we cannot disable transaction
completely because: 1. binlog rollback removes pending row events
(binlog_remove_pending_rows_event()). The row events are added
during CREATE .. SELECT data insertion phase. 2. replication slave
highly depends on transaction and cannot work without it.
So we put temporary Aria table into non-transactional mode with
"thd->transaction->on hack". See comment for on_save variable.
Note that Aria table has internal_table mode. But we cannot use it
because:
if (!internal_table)
{
mysql_mutex_lock(&THR_LOCK_myisam);
old_info= test_if_reopen(name_buff);
}
For internal_table test_if_reopen() is not called and we get a new
MARIA_SHARE for each file handler. In that case duplicate errors are
missed because insert and lookup in CREATE .. SELECT is done via two
different handlers (see create_lookup_handler()).
For temporary table before dropping TABLE_SHARE by
drop_temporary_table() we must do ha_reset(). ha_reset() releases
storage share. Without that the share is kept and the second CREATE
OR REPLACE .. SELECT fails with:
HA_ERR_TABLE_EXIST (156): MyISAM table '#sql-create-b5377-4-t2' is
in use (most likely by a MERGE table). Try FLUSH TABLES.
HA_EXTRA_PREPARE_FOR_DROP also removes MYISAM_SHARE, but that is
not needed as ha_reset() does the job.
ha_reset() is usually done by
mark_tmp_table_as_free_for_reuse(). But we don't need that mechanism
for our temporary table.
Atomic_info in HA_CREATE_INFO
Many functions in CREATE TABLE pass the same parameters. These
parameters are part of table creation info and should be in
HA_CREATE_INFO (or whatever). Passing parameters via single
structure is much easier for adding new data and
refactoring.
InnoDB changes (revised by Marko Mäkelä)
row_rename_table_for_mysql(): Specify the treatment of FOREIGN KEY
constraints in a 4-valued enum parameter. In cases where FOREIGN KEY
constraints cannot exist (partitioned tables, or internal tables of
FULLTEXT INDEX), we can use the mode RENAME_IGNORE_FK.
The mod RENAME_REBUILD is for any DDL operation that rebuilds the
table inside InnoDB, such as TRUNCATE and native ALTER TABLE
(or OPTIMIZE TABLE). The mode RENAME_ALTER_COPY is used solely
during non-native ALTER TABLE in ha_innobase::rename_table().
Normal ha_innobase::rename_table() will use the mode RENAME_FK.
CREATE OR REPLACE will rename the old table (if one exists) along
with its FOREIGN KEY constraints into a temporary name. The replacement
table will be initially created with another temporary name.
Unlike in ALTER TABLE, all FOREIGN KEY constraints must be renamed
and not inherited as part of these operations, using the mode RENAME_FK.
dict_get_referenced_table(): Let the callers convert names when needed.
create_table_info_t::create_foreign_keys(): CREATE OR REPLACE creates
the replacement table with a temporary name table, so for
self-references foreign->referenced_table will be a table with
temporary name and charset conversion must be skipped for it.
Reviewed by:
Michael Widenius <monty@mariadb.org>
Part #2: make sure we allocate space for two JOIN_TABs that
use temporary tables.
The dbug_join_tab_array_size is still set to catch cases where
we try to access more JOIN_TAB object than we thought we would have.
The problem was caused by use of COLLATION(AVG('x')). This is an
item whose value is a constant.
Name Resolution code called convert_const_to_int() which removed AVG('x').
However, the item representing COLLATION(...) still had with_sum_func=1.
This inconsistent state confused the code that handles grouping and
DISTINCT: JOIN::get_best_combination() decided to use one temporary
table and allocated one JOIN_TAB for it, but then
JOIN::make_aggr_tables_info() attempted to use two and made writes
beyond the end of the JOIN::join_tab array.
The fix:
- Do not replace constant expressions which contain aggregate functions.
- Add JOIN::dbug_join_tab_array_size to catch attempts to use more
JOIN_TAB objects than we've allocated.
This bug could cause a crash of the server when executing queries containing
ANY/ALL predicands with redundant subqueries in GROUP BY clauses.
These subqueries are eliminated by remove_redundant_subquery_clause()
together with elimination of GROUP BY list containing these subqueries.
However the references to the elements of the GROUP BY remained in the
JOIN::all_fields list of the right operand of of the ALL/ANY predicand.
Later these references confused make_aggr_tables_info() when forming
proper execution structures after ALL/ANY predicands had been replaced
with expressions containing MIN/MAX set functions.
The patch just removes these references from JOIN::all_fields list used
by the subquery of the ALL/ANY predicand when its GROUP BY clause is
eliminated.
Approved by Oleksandr Byelkin <sanja@mariadb.com>
Part #2: Extend heuristic pruning to use multiple tables as the
"Model tables".
Before the patch, heuristic pruning uses only one "Model table":
The table which had the best cost AND record became the "Model table".
After that, if a table's cost and record were both worse than
those of the Model Table, the table would be pruned away.
This didn't work well when the first table (the optimizer sorts them
by record_count) had low record_count but relatively high cost: nothing
could be pruned afterwards.
The patch adds the two additional "Model tables": one with the least
cost and the other with the least record_count.
(In both cases, a table can be pruned away if BOTH its cost and
record_count are worse than those of a Model table)
The new pruning is active when the number of tables to consider for
the prefix is higher than @@optimizer_extra_pruning_depth.
One can see the new pruning in the Optimizer Trace as
- "pruned_by_heuristic":"min_record_count", or
- "pruned_by_heuristic":"min_read_time".
Old heuristic pruning shows as "pruned_by_heuristic":1.
Replaced Cost_estimate prefix_cost with a double as prefix_cost was
only used to store and retrive total prefix cost.
This also speeds up things (a bit) as don't have to call
Cost_estimate::total_cost() for every access to the prefix_cost.
Sizeof POSITION decreased from 304 to 256.
MDEV-28073 Slow query performance in MariaDB when using many table
The idea is to prefer and chain EQ_REF tables (tables that uses an
unique key to find a row) when searching for the best table combination.
This significantly reduces row combinations that has to be examined.
This is optimization is enabled when setting optimizer_prune_level=2
(which is now default).
Implementation:
- optimizer_prune_level has a new level, 2, which enables EQ_REF
optimization in addition to the pruning done by level 1.
Level 2 is now default.
- Added JOIN::eq_ref_tables that contains bits of tables that could use
potentially use EQ_REF access in the query. This is calculated
in sort_and_filter_keyuse()
Under optimizer_prune_level=2:
- When the greedy_optimizer notices that the preceding table was an
EQ_REF table, it tries to add an EQ_REF table next. If an EQ_REF
table exists, only this one will be considered at this level.
We also collect all EQ_REF tables chained by the next levels and these
are ignored on the starting level as we have already examined these.
If no EQ_REF table exists, we continue as normal.
This optimization speeds up the greedy_optimizer combination test with
~25%
Other things:
- I ported the changes in MySQL 5.7 to greedy_optimizer.test to MariaDB
to be able to ensure we can handle all cases that MySQL can do.
- I have run all tests with --mysqld=--optimizer_prune_level=1 to verify that
there where no test changes.
The dependency checking is not needed when using
best_extension_by_limited_search() as this function ensures
that we don't use tables that are depending on any of the remaning
tables.
MDEV-28073 Slow query performance in MariaDB when using many tables
The faster we can find a good query plan, the more options we have for
finding and pruning (ignoring) bad plans.
This patch adds sorting of plans to best_extension_by_limited_search().
The plans, from best_access_path() are sorted according to the numbers
of found rows. This allows us to faster find 'good tables' and we are
thus able to eliminate 'bad plans' faster.
One side effect of this patch is that if two tables have equal cost,
the table that which was used earlier in the query is preferred.
This allows users to improve plans by reordering eq_ref tables in the
order they would like them to be uses.
Result changes caused by the patch:
- Traces are different as now we print the cost for using tables before
we start considering them in the plan.
- Table order are changed for some plans. In most cases this is because
the plans are equal and tables are in this case sorted according to
their usage in the original query.
- A few plans was changed as the optimizer was able to find a better
plan (that was pruned by the original code).
Other things:
- Added a new statistic variable: "optimizer_join_prefixes_check_calls",
which counts number of calls to best_extension_by_limited_search().
This can be used to check the prune efficiency in greedy_search().
- Added variable "JOIN_TAB::embedded_dependent" to be able to handle
XX IN (SELECT..) in the greedy_optimizer. The idea is that we
should prune a table if any of the tables in embedded_dependent is
not yet read.
- When using many tables in a query, there will be some additional
memory usage as we need to pre-allocate table of
table_count*table_count*sizeof(POSITION) objects (POSITION is 312
bytes for now) to hold the pre-calculated best_access_path()
information. This memory usage is offset by the expected
performance improvement when using many tables in a query.
- Removed the code from an earlier patch to keep the table order in
join->best_ref in the original order. This is not needed anymore as we
are now sorting the tables for each best_extension_by_limited_search()
call.
1. For INSERT..SELECT statements: don't include table/view the data
is inserted into in the list of leaf tables
2. Remove duplicated and dead code related to table_count
optimize_semi_joins() calls update_sj_state() to update semi-join
optimization state in the JOIN class.
greedy_search() algorithm considers different join prefixes,
and then picks one table to put into the join prefix.
Most of the semi-join optimization state is in the table's entry
in the join->positions[cur_prefix_size].
However, it also needs to call update_sj_state() to update the
semi-join optimization state in the JOIN class.
There is one exception, which is the cause of this bug: when we're
inside optimize_semi_join_nests() and are optimizing a subquery,
optimize_semi_joins() does nothing, it doesn't call update_sj_state().
greedy_search() must not do that either.
The bug was that build_notnull_conds_for_range_scans() did not take into
account the join_tab is not yet sorted with constant tables first.
Fixed the bug by testing explicitely if a table is a const table.
Part of:
MDEV-28073 Slow query performance in MariaDB when using many tables
s->key_dependent has a list of tables that are compared with key fields
in the current table. However it does not take into account if a key
field could be resolved by another table.
This is because MariaDB expands 'join_tab->keyuse' to include all generated
comparisons.
For example:
SELECT * from t1,t2,t3 where t1.key=t2.key and t2.key=t3.key
In this case keyuse for t1 includes t2.key and t3.key and key_dependent
contains 't2.map | t3.map'
If we in best_extension_by_limited_search() consider t2,t1 then t1's
key is fully defined, but we cannot do any prune of plans as
s->key_dependent indicates that t3 is still needed.
Fixed by calculating in best_access_patch the current key_dependent map
of tables that is needed to satisfy all keys. This allows us to prune
more bad plans earlier as soon as all keys can be used.
We also set key_dependent to 0 if we found an EQ_REF key, as this an
optimal key for the table and there is no reason to check more keys.
best_extension_by_limited_search() assumes that tables should be sorted
according to size to be able to quickly disregard bad plans. However the
current usage of swap_variables() will change the table order to a not
sorted one for the next recursive call. This breaks the assumtion and
causes performance issues when using many tables (we have to examine
many more plans).
This patch fixes this by ensuring that the original table order is kept
for the not yet used tables when best_extension_by_limited_search() is
called.
This was done by always calling swap_variables() for each table and
restoring the original table order at exit.
Some test changed:
- In a majority of the test the change was that two "identical tables"
where swapped and the optimzer is now using the first/smaller table
- In few test the table order was changed. The new plan looks identical
or slighly better than the original.
(Try 2)
The code that updates semi-join optimization state for a join order prefix
had several bugs. The visible effect was bad optimization for FirstMatch or
LooseScan strategies: they either weren't considered when they should have
been, or considered when they shouldn't have been.
In order to hit the bug, the optimizer needs to consider several different
join prefixes in a certain order. Queries with "obvious" query plans which
prune all join orders except one are not affected.
Internally, the bugs in updates of semi-join state were:
1. restore_prev_sj_state() assumed that
"we assume remaining_tables doesnt contain @tab"
which wasn't true.
2. Another bug in this function: it did remove bits from
join->cur_sj_inner_tables but never added them.
3. greedy_search() adds tables into the join prefix but neglects to update
the semi-join optimization state. (It does update nested outer join
state, see this call:
check_interleaving_with_nj(best_table)
but there's no matching call to update the semi-join state.
(This wasn't visible because most of the state is in the POSITION
structure which is updated. But there is also state in JOIN, too)
The patch:
- Fixes all of the above
- Adds JOIN::dbug_verify_sj_inner_tables() which is used to verify the
state is correct at every step.
- Renames advance_sj_state() to optimize_semi_joins().
= Introduces update_sj_state() which ideally should have been called
"advance_sj_state" but I didn't reuse the name to not create confusion.
Main fix was replacing read_time+= with read_time
I also did updated the 'identical' code in optimize_straight_join) and
best_extension_by_limited_search() to make them eaiser to compare.
Reviewer: Sergei Petrunia <sergey@mariadb.com>
(Try 2) (Cherry-pick back into 10.3)
The code that updates semi-join optimization state for a join order prefix
had several bugs. The visible effect was bad optimization for FirstMatch or
LooseScan strategies: they either weren't considered when they should have
been, or considered when they shouldn't have been.
In order to hit the bug, the optimizer needs to consider several different
join prefixes in a certain order. Queries with "obvious" query plans which
prune all join orders except one are not affected.
Internally, the bugs in updates of semi-join state were:
1. restore_prev_sj_state() assumed that
"we assume remaining_tables doesnt contain @tab"
which wasn't true.
2. Another bug in this function: it did remove bits from
join->cur_sj_inner_tables but never added them.
3. greedy_search() adds tables into the join prefix but neglects to update
the semi-join optimization state. (It does update nested outer join
state, see this call:
check_interleaving_with_nj(best_table)
but there's no matching call to update the semi-join state.
(This wasn't visible because most of the state is in the POSITION
structure which is updated. But there is also state in JOIN, too)
The patch:
- Fixes all of the above
- Adds JOIN::dbug_verify_sj_inner_tables() which is used to verify the
state is correct at every step.
- Renames advance_sj_state() to optimize_semi_joins().
= Introduces update_sj_state() which ideally should have been called
"advance_sj_state" but I didn't reuse the name to not create confusion.
