The problem was an assignment in test_quick_select() that flagged empty
tables with "Impossible where". This test was however wrong as it
didn't work correctly for left join.
Removed the test, but added checking of empty tables in DELETE and UPDATE
to get similar EXPLAIN as before.
The new tests is a bit more strict (better) than before as it catches all
cases of empty tables in single table DELETE/UPDATE.
The main difference in code path between EQ_REF and REF is that for
REF we have to do an extra read_next on the index to check that there
is no more matching rows.
Before this patch we added a preference of EQ_REF by ensuring that REF
would always estimate to find at least 2 rows.
This patch adds the cost of the extra key read_next to REF access and
removes the code that limited REF to at least 2 rows. For some queries
this can have a big effect as the total estimated rows will be halved
for each REF table with 1 rows.
multi_range cost calculations are also changed to take into account
the difference between EQ_REF and REF.
The effect of the patch to the test suite:
- About 80 test case changed
- Almost all changes where for EXPLAIN where estimated rows for REF
where changed from 2 to 1.
- A few test cases using explain extended had a change of 'filtered'.
This is because of the estimated rows are now closer to the
calculated selectivity.
- A very few test had a change of table order.
This is because the change of estimated rows from 2 to 1 or the small
cost change for REF
(main.subselect_sj_jcl6, main.group_by, main.dervied_cond_pushdown,
main.distinct, main.join_nested, main.order_by, main.join_cache)
- No key statistics and the estimated rows are now smaller which cased
estimated filtering to be lower.
(main.subselect_sj_mat)
- The number of total rows are halved.
(main.derived_cond_pushdown)
- Plans with 1 row changed to use RANGE instead of REF.
(main.group_min_max)
- ALL changed to REF
(main.key_diff)
- Key changed from ref + index_only to PRIMARY key for InnoDB, as
OPTIMIZER_ROW_LOOKUP_COST + OPTIMIZER_ROW_NEXT_FIND_COST is smaller than
OPTIMIZER_KEY_LOOKUP_COST + OPTIMIZER_KEY_NEXT_FIND_COST.
(main.join_outer_innodb)
- Cost changes printouts
(main.opt_trace*)
- Result order change
(innodb_gis.rtree)
The reason for this is that we call file->index_flags(index, 0, 1)
multiple times in best_access_patch()when optimizing a table.
For example, in InnoDB, the calls is not trivial (4 if's and 2 assignments)
Now the function is inlined and is just a memory reference.
Other things:
- handler::is_clustering_key() and pk_is_clustering_key() are now inline.
- Added TABLE::can_use_rowid_filter() to simplify some code.
- Test if we should use a rowid_filter only if can_use_rowid_filter() is
true.
- Added TABLE::is_clustering_key() to avoid a memory reference.
- Simplify some code using the fact that HA_KEYREAD_ONLY is true implies
that HA_CLUSTERED_INDEX is false.
- Added DBUG_ASSERT to TABLE::best_range_rowid_filter() to ensure we
do not call it with a clustering key.
- Reorginized elements in struct st_key to get better memory alignment.
- Updated ha_innobase::index_flags() to not have
HA_DO_RANGE_FILTER_PUSHDOWN for clustered index
This was done after discussions with Igor, Sanja and Bar.
The main reason for removing the deprication was to ensure that MariaDB
is always backward compatible whenever possible.
Other things:
- Added statistics counters, mainly for the feedback plugin.
- INTO OUTFILE
- INTO variable
- If INTO is using the old syntax (end of query)
In the case one has an old Aria log file that ands with a Aria checkpoint
and the server restarts after next recovery, just after created a
new Aria log file (of 8K), the Aria recovery code would abort.
If one would try to delete all Aria log files after this (but not the
aria_control_file), the server would crash during recovery.
The problem was that translog_get_last_page_addr() would regard a log file
of exactly 8K as illegal and the rest of the code could not handle this
case.
Another issue was that if there was a crash directly after the log file
head was written to the next page, the code in translog_get_next_chunk()
would crash.
This patch fixes most of the issues, but not all. For Sanja to look at!
Things fixed:
- Added code to ignore 8K log files.
- Removed ASSERT in translog_get_next_chunk() that checks if page only
contains the log page header.
I spent 4 hours on work and 12 hours of testing to try to find
the reason for aria crashing in recovery when starting a new test,
in which case the 'data directory' should be a copy of "install.db",
but aria_log.00000001 content was not correct.
The following changes are mostly done to make it a bit easier to find out
more in case of future similar crashes:
- Mark last_checkpoint_lsn volatile (safety).
- Write checkpoint message to aria_recovery.trace
- When compling with DBUG and with HAVE_DBUG_TRANSLOG_SRC,
use checksum's for Aria log pages. We cannot have it on by default
for DBUG servers yet as there is bugs when changing CRC between
restarts.
- Added a message to mtr --verbose when copying the data directory.
- Removed extra linefeed in Aria recovery message (cleanup)
This includes:
- cleanup and optimization of filtering and pushdown engine code.
- Adjusted costs for rowid filters (based on extensive testing
and profiling).
This made a small two changes to the handler_rowid_filter_is_active()
API:
- One should not call it with a zero pointer!
- One does not need to call handler_rowid_filter_is_active() for every
row anymore. It is enough to check if filter is active by calling it
call it during index_init() or when handler::rowid_filter_changed()
is called
The changes was to avoid unnecessary function calls and checks if
pushdown conditions and rowid_filter is not used.
Updated costs for rowid_filter_lookup() to be closer to reality.
The old cost was based only on rowid_compare_cost. This is now
changed to take into account the overhead in checking the rowid.
Changed the Range_rowid_filter class to use DYNAMIC_ARRAY directly
instead of Dynamic_array<>. This was done to be able to use the new
append_dynamic() functions which gives a notable speed improvment
compared to the old code. Removing the abstraction also makes
the code easier to understand.
The cost of filtering is now slightly lower than before, which
is reflected in some test cases that is now using rowid filters.
This includes all test changes from
"Changing all cost calculation to be given in milliseconds"
and forwards.
Some of the things that caused changes in the result files:
- As part of fixing tests, I added 'echo' to some comments to be able to
easier find out where things where wrong.
- MATERIALIZED has now a higher cost compared to X than before. Because
of this some MATERIALIZED types have changed to DEPENDEND SUBQUERY.
- Some test cases that required MATERIALIZED to repeat a bug was
changed by adding more rows to force MATERIALIZED to happen.
- 'Filtered' in SHOW EXPLAIN has in many case changed from 100.00 to
something smaller. This is because now filtered also takes into
account the smallest possible ref access and filters, even if they
where not used. Another reason for 'Filtered' being smaller is that
we now also take into account implicit filtering done for subqueries
using FIRSTMATCH.
(main.subselect_no_exists_to_in)
This is caluculated in best_access_path() and stored in records_out.
- Table orders has changed because more accurate costs.
- 'index' and 'ALL' for small tables has changed to use 'range' or
'ref' because of optimizer_scan_setup_cost.
- index can be changed to 'range' as 'range' optimizer assumes we don't
have to read the blocks from disk that range optimizer has already read.
This can be confusing in the case where there is no obvious where clause
but instead there is a hidden 'key_column > NULL' added by the optimizer.
(main.subselect_no_exists_to_in)
- Scan on primary clustered key does not report 'Using Index' anymore
(It's a table scan, not an index scan).
- For derived tables, the number of rows is now 100 instead of 2,
which can be seen in EXPLAIN.
- More tests have "Using index for group by" as the cost of this
optimization is now more correct (lower).
- A primary key could be preferred for a normal key, even if it would
access more rows, as it's faster to do 1 lokoup and 3 'index_next' on a
clustered primary key than one lookup trough a secondary.
(main.stat_tables_innodb)
Notes:
- There was a 4.7% more calls to best_extension_by_limited_search() in
the main.greedy_optimizer test. However examining the test results
it looked that the plans where slightly better (eq_ref where more
chained together) so I assume this is ok.
- I have verified a few test cases where there was notable/unexpected
changes in the plan and in all cases the new optimizer plans where
faster. (main.greedy_optimizer and some others)
InnoDB FTS scan was used by a subquery. A subquery execution may start
a table read and continue until it finds the first matching record
combination. This can happen before the table read returns EOF.
The next time the subquery is executed, it will start another table read.
InnoDB FTS table read fails to re-initialize its data structures in this
scenario and will try to continue the scan started at the first execution.
Fixed by ha_innobase::ft_init() to stop the FTS scan if there is one.
Author: Sergei Petrunia <sergey@mariadb.com>
Reviewer: Monty
This solves the current problem in the optimizer
- SELECT FROM big_table
- SELECT from small_table where small_table.eq_ref_key=big_table.id
The old code assumed that each eq_ref access will cause an IO.
As the cost of IO is high, this dominated the cost for the later table
which caused the optimizer to prefer table scans + join cache over
index reads.
This patch fixes this issue by limit the number of expected IO calls,
for rows and index separately, to the size of the table or index or
the number of accesses that we except in a range for the index.
The major changes are:
- Adding a new structure ALL_READ_COST that is mainly used in
best_access_path() to hold the costs parts of the cost we are
calculating. This allows us to limit the number of IO when multiplying
the cost with the previous row combinations.
- All storage engine cost functions are changed to return IO_AND_CPU_COST.
The virtual cost functions should now return in IO_AND_CPU_COST.io
the number of disk blocks that will be accessed instead of the cost
of the access.
- We are not limiting the io_blocks for table or index scans as we
assume that engines may not store these in the 'hot' part of the
cache. Table and index scan also uses much less IO blocks than
key accesses, so the original issue is not as critical with scans.
Other things:
OPT_RANGE now holds a 'Cost_estimate cost' instead a lot of different
costs. All the old costs, like index_only_read, can be extracted
from 'cost'.
- Added to the start of some functions 'handler *file= table->file'
to shorten the code that is using the handler.
- handler->cost() is used to change a ALL_READ_COST or IO_AND_CPU_COST
to 'cost in milliseconds'
- New functions: handler::index_blocks() and handler::row_blocks()
which are used to limit the IO.
- Added index_cost and row_cost to Cost_estimate and removed all not
needed members.
- Removed cost coefficients from Cost_estimate as these don't make sense
when costs (except IO_BLOCKS) are in milliseconds.
- Removed handler::avg_io_cost() and replaced it with DISK_READ_COST.
- Renamed best_range_rowid_filter_for_partial_join() to
best_range_rowid_filter() as using the old name made rows too long.
- Changed all SJ_MATERIALIZATION_INFO 'Cost_estimate' variables to
'double' as Cost_estimate power was not used for these and thus
just caused storage and performance overhead.
- Changed cost_for_index_read() to use 'worst_seeks' to only limit
IO, not number of table accesses. With this patch worst_seeks is
probably not needed anymore, but I kept it around just in case.
- Applying cost for filter got to be much shorter and easier thanks
to the API changes.
- Adjusted cost for fulltext keys in collaboration with Sergei Golubchik.
- Most test changes caused by this patch is that table scans are changed
to use indexes.
- Added ha_seq::keyread_time() and ha_seq::key_scan_time() to get
make checking number of potential IO blocks easier during debugging.
The issue was that when limit is used,
SQL_SELECT::test_quick_select would set the cost of table scan to be
unreasonable high to force a range to be used.
The problem with this approach was that range was used even when the
cost of range, when it would only read 'limit rows' would be higher
than the cost of a table scan.
This patch fixes it by not accepting ranges when the range can never
have a lower cost than a table scan, even if every row would match the
WHERE clause.
This makes it easier to compare different costs and also allows
the optimizer to optimizer different storage engines more reliably.
- Added tests/check_costs.pl, a tool to verify optimizer cost calculations.
- Most engine costs has been found with this program. All steps to
calculate the new costs are documented in Docs/optimizer_costs.txt
- User optimizer_cost variables are given in microseconds (as individual
costs can be very small). Internally they are stored in ms.
- Changed DISK_READ_COST (was DISK_SEEK_BASE_COST) from a hard disk cost
(9 ms) to common SSD cost (400MB/sec).
- Removed cost calculations for hard disks (rotation etc).
- Changed the following handler functions to return IO_AND_CPU_COST.
This makes it easy to apply different cost modifiers in ha_..time()
functions for io and cpu costs.
- scan_time()
- rnd_pos_time() & rnd_pos_call_time()
- keyread_time()
- Enhanched keyread_time() to calculate the full cost of reading of a set
of keys with a given number of ranges and optional number of blocks that
need to be accessed.
- Removed read_time() as keyread_time() + rnd_pos_time() can do the same
thing and more.
- Tuned cost for: heap, myisam, Aria, InnoDB, archive and MyRocks.
Used heap table costs for json_table. The rest are using default engine
costs.
- Added the following new optimizer variables:
- optimizer_disk_read_ratio
- optimizer_disk_read_cost
- optimizer_key_lookup_cost
- optimizer_row_lookup_cost
- optimizer_row_next_find_cost
- optimizer_scan_cost
- Moved all engine specific cost to OPTIMIZER_COSTS structure.
- Changed costs to use 'records_out' instead of 'records_read' when
recalculating costs.
- Split optimizer_costs.h to optimizer_costs.h and optimizer_defaults.h.
This allows one to change costs without having to compile a lot of
files.
- Updated costs for filter lookup.
- Use a better cost estimate in best_extension_by_limited_search()
for the sorting cost.
- Fixed previous issues with 'filtered' explain column as we are now
using 'records_out' (min rows seen for table) to calculate filtering.
This greatly simplifies the filtering code in
JOIN_TAB::save_explain_data().
This change caused a lot of queries to be optimized differently than
before, which exposed different issues in the optimizer that needs to
be fixed. These fixes are in the following commits. To not have to
change the same test case over and over again, the changes in the test
cases are done in a single commit after all the critical change sets
are done.
InnoDB changes:
- Updated InnoDB to not divide big range cost with 2.
- Added cost for InnoDB (innobase_update_optimizer_costs()).
- Don't mark clustered primary key with HA_KEYREAD_ONLY. This will
prevent that the optimizer is trying to use index-only scans on
the clustered key.
- Disabled ha_innobase::scan_time() and ha_innobase::read_time() and
ha_innobase::rnd_pos_time() as the default engine cost functions now
works good for InnoDB.
Other things:
- Added --show-query-costs (\Q) option to mysql.cc to show the query
cost after each query (good when working with query costs).
- Extended my_getopt with GET_ADJUSTED_VALUE which allows one to adjust
the value that user is given. This is used to change cost from
microseconds (user input) to milliseconds (what the server is
internally using).
- Added include/my_tracker.h ; Useful include file to quickly test
costs of a function.
- Use handler::set_table() in all places instead of 'table= arg'.
- Added SHOW_OPTIMIZER_COSTS to sys variables. These are input and
shown in microseconds for the user but stored as milliseconds.
This is to make the numbers easier to read for the user (less
pre-zeros). Implemented in 'Sys_var_optimizer_cost' class.
- In test_quick_select() do not use index scans if 'no_keyread' is set
for the table. This is what we do in other places of the server.
- Added THD parameter to Unique::get_use_cost() and
check_index_intersect_extension() and similar functions to be able
to provide costs to called functions.
- Changed 'records' to 'rows' in optimizer_trace.
- Write more information to optimizer_trace.
- Added INDEX_BLOCK_FILL_FACTOR_MUL (4) and INDEX_BLOCK_FILL_FACTOR_DIV (3)
to calculate usage space of keys in b-trees. (Before we used numeric
constants).
- Removed code that assumed that b-trees has similar costs as binary
trees. Replaced with engine calls that returns the cost.
- Added Bitmap::find_first_bit()
- Added timings to join_cache for ANALYZE table (patch by Sergei Petrunia).
- Added records_init and records_after_filter to POSITION to remember
more of what best_access_patch() calculates.
- table_after_join_selectivity() changed to recalculate 'records_out'
based on the new fields from best_access_patch()
Bug fixes:
- Some queries did not update last_query_cost (was 0). Fixed by moving
setting thd->...last_query_cost in JOIN::optimize().
- Write '0' as number of rows for const tables with a matching row.
Some internals:
- Engine cost are stored in OPTIMIZER_COSTS structure. When a
handlerton is created, we also created a new cost variable for the
handlerton. We also create a new variable if the user changes a
optimizer cost for a not yet loaded handlerton either with command
line arguments or with SET
@@global.engine.optimizer_cost_variable=xx.
- There are 3 global OPTIMIZER_COSTS variables:
default_optimizer_costs The default costs + changes from the
command line without an engine specifier.
heap_optimizer_costs Heap table costs, used for temporary tables
tmp_table_optimizer_costs The cost for the default on disk internal
temporary table (MyISAM or Aria)
- The engine cost for a table is stored in table_share. To speed up
accesses the handler has a pointer to this. The cost is copied
to the table on first access. If one wants to change the cost one
must first update the global engine cost and then do a FLUSH TABLES.
This was done to be able to access the costs for an open table
without any locks.
- When a handlerton is created, the cost are updated the following way:
See sql/keycaches.cc for details:
- Use 'default_optimizer_costs' as a base
- Call hton->update_optimizer_costs() to override with the engines
default costs.
- Override the costs that the user has specified for the engine.
- One handler open, copy the engine cost from handlerton to TABLE_SHARE.
- Call handler::update_optimizer_costs() to allow the engine to update
cost for this particular table.
- There are two costs stored in THD. These are copied to the handler
when the table is used in a query:
- optimizer_where_cost
- optimizer_scan_setup_cost
- Simply code in best_access_path() by storing all cost result in a
structure. (Idea/Suggestion by Igor)
Before this patch, when calculating the cost of fetching and using a
row/key from the engine, we took into account the cost of finding a
row or key from the engine, but did not consistently take into account
index only accessed, clustered key or covered keys for all access
paths.
The cost of the WHERE clause (TIME_FOR_COMPARE) was not consistently
considered in best_access_path(). TIME_FOR_COMPARE was used in
calculation in other places, like greedy_search(), but was in some
cases (like scans) done an a different number of rows than was
accessed.
The cost calculation of row and index scans didn't take into account
the number of rows that where accessed, only the number of accepted
rows.
When using a filter, the cost of index_only_reads and cost of
accessing and disregarding 'filtered rows' where not taken into
account, which made filters cost less than there actually where.
To remedy the above, the following key & row fetch related costs
has been added:
- The cost of fetching and using a row is now split into different costs:
- key + Row fetch cost (as before) but multiplied with the variable
'optimizer_cache_cost' (default to 0.5). This allows the user to
tell the optimizer the likehood of finding the key and row in the
engine cache.
- ROW_COPY_COST, The cost copying a row from the engine to the
sql layer or creating a row from the join_cache to the record
buffer. Mostly affects table scan costs.
- ROW_LOOKUP_COST, the cost of fetching a row by rowid.
- KEY_COPY_COST the cost of finding the next key and copying it from
the engine to the SQL layer. This is used when we calculate the cost
index only reads. It makes index scans more expensive than before if
they cover a lot of rows. (main.index_merge_myisam)
- KEY_LOOKUP_COST, the cost of finding the first key in a range.
This replaces the old define IDX_LOOKUP_COST, but with a higher cost.
- KEY_NEXT_FIND_COST, the cost of finding the next key (and rowid).
when doing a index scan and comparing the rowid to the filter.
Before this cost was assumed to be 0.
All of the above constants/variables are now tuned to be somewhat in
proportion of executing complexity to each other. There is tuning
need for these in the future, but that can wait until the above are
made user variables as that will make tuning much easier.
To make the usage of the above easy, there are new (not virtual)
cost calclation functions in handler:
- ha_read_time(), like read_time(), but take optimizer_cache_cost into
account.
- ha_read_and_copy_time(), like ha_read_time() but take into account
ROW_COPY_TIME
- ha_read_and_compare_time(), like ha_read_and_copy_time() but take
TIME_FOR_COMPARE into account.
- ha_rnd_pos_time(). Read row with row id, taking ROW_COPY_COST
into account. This is used with filesort where we don't need
to execute the WHERE clause again.
- ha_keyread_time(), like keyread_time() but take
optimizer_cache_cost into account.
- ha_keyread_and_copy_time(), like ha_keyread_time(), but add
KEY_COPY_COST.
- ha_key_scan_time(), like key_scan_time() but take
optimizer_cache_cost nto account.
- ha_key_scan_and_compare_time(), like ha_key_scan_time(), but add
KEY_COPY_COST & TIME_FOR_COMPARE.
I also added some setup costs for doing different types of scans and
creating temporary tables (on disk and in memory). This encourages
the optimizer to not use these for simple 'a few row' lookups if
there are adequate key lookup strategies.
- TABLE_SCAN_SETUP_COST, cost of starting a table scan.
- INDEX_SCAN_SETUP_COST, cost of starting an index scan.
- HEAP_TEMPTABLE_CREATE_COST, cost of creating in memory
temporary table.
- DISK_TEMPTABLE_CREATE_COST, cost of creating an on disk temporary
table.
When calculating cost of fetching ranges, we had a cost of
IDX_LOOKUP_COST (0.125) for doing a key div for a new range. This is
now replaced with 'io_cost * KEY_LOOKUP_COST (1.0) *
optimizer_cache_cost', which matches the cost we use for 'ref' and
other key lookups. The effect is that the cost is now a bit higher
when we have many ranges for a key.
Allmost all calculation with TIME_FOR_COMPARE is now done in
best_access_path(). 'JOIN::read_time' now includes the full
cost for finding the rows in the table.
In the result files, many of the changes are now again close to what
they where before the "Update cost for hash and cached joins" commit,
as that commit didn't fix the filter cost (too complex to do
everything in one commit).
The above changes showed a lot of a lot of inconsistencies in
optimizer cost calculation. The main objective with the other changes
was to do calculation as similar (and accurate) as possible and to make
different plans more comparable.
Detailed list of changes:
- Calculate index_only_cost consistently and correctly for all scan
and ref accesses. The row fetch_cost and index_only_cost now
takes into account clustered keys, covered keys and index
only accesses.
- cost_for_index_read now returns both full cost and index_only_cost
- Fixed cost calculation of get_sweep_read_cost() to match other
similar costs. This is bases on the assumption that data is more
often stored on SSD than a hard disk.
- Replaced constant 2.0 with new define TABLE_SCAN_SETUP_COST.
- Some scan cost estimates did not take into account
TIME_FOR_COMPARE. Now all scan costs takes this into
account. (main.show_explain)
- Added session variable optimizer_cache_hit_ratio (default 50%). By
adjusting this on can reduce or increase the cost of index or direct
record lookups. The effect of the default is that key lookups is now
a bit cheaper than before. See usage of 'optimizer_cache_cost' in
handler.h.
- JOIN_TAB::scan_time() did not take into account index only scans,
which produced a wrong cost when index scan was used. Changed
JOIN_TAB:::scan_time() to take into consideration clustered and
covered keys. The values are now cached and we only have to call
this function once. Other calls are changed to use the cached
values. Function renamed to JOIN_TAB::estimate_scan_time().
- Fixed that most index cost calculations are done the same way and
more close to 'range' calculations. The cost is now lower than
before for small data sets and higher for large data sets as we take
into account how many keys are read (main.opt_trace_selectivity,
main.limit_rows_examined).
- Ensured that index_scan_cost() ==
range(scan_of_all_rows_in_table_using_one_range) +
MULTI_RANGE_READ_INFO_CONST. One effect of this is that if there
is choice of doing a full index scan and a range-index scan over
almost the whole table then index scan will be preferred (no
range-read setup cost). (innodb.innodb, main.show_explain,
main.range)
- Fixed the EQ_REF and REF takes into account clustered and covered
keys. This changes some plans to use covered or clustered indexes
as these are much cheaper. (main.subselect_mat_cost,
main.state_tables_innodb, main.limit_rows_examined)
- Rowid filter setup cost and filter compare cost now takes into
account fetching and checking the rowid (KEY_NEXT_FIND_COST).
(main.partition_pruning heap.heap_btree main.log_state)
- Added KEY_NEXT_FIND_COST to
Range_rowid_filter_cost_info::lookup_cost to account of the time
to find and check the next key value against the container
- Introduced ha_keyread_time(rows) that takes into account finding
the next row and copying the key value to 'record'
(KEY_COPY_COST).
- Introduced ha_key_scan_time() for calculating an index scan over
all rows.
- Added IDX_LOOKUP_COST to keyread_time() as a startup cost.
- Added index_only_fetch_cost() as a convenience function to
OPT_RANGE.
- keyread_time() cost is slightly reduced to prefer shorter keys.
(main.index_merge_myisam)
- All of the above caused some index_merge combinations to be
rejected because of cost (main.index_intersect). In some cases
'ref' where replaced with index_merge because of the low
cost calculation of get_sweep_read_cost().
- Some index usage moved from PRIMARY to a covering index.
(main.subselect_innodb)
- Changed cost calculation of filter to take KEY_LOOKUP_COST and
TIME_FOR_COMPARE into account. See sql_select.cc::apply_filter().
filter parameters and costs are now written to optimizer_trace.
- Don't use matchings_records_in_range() to try to estimate the number
of filtered rows for ranges. The reason is that we want to ensure
that 'range' is calculated similar to 'ref'. There is also more work
needed to calculate the selectivity when using ranges and ranges and
filtering. This causes filtering column in EXPLAIN EXTENDED to be
100.00 for some cases where range cannot use filtering.
(main.rowid_filter)
- Introduced ha_scan_time() that takes into account the CPU cost of
finding the next row and copying the row from the engine to
'record'. This causes costs of table scan to slightly increase and
some test to changed their plan from ALL to RANGE or ALL to ref.
(innodb.innodb_mysql, main.select_pkeycache)
In a few cases where scan time of very small tables have lower cost
than a ref or range, things changed from ref/range to ALL.
(main.myisam, main.func_group, main.limit_rows_examined,
main.subselect2)
- Introduced ha_scan_and_compare_time() which is like ha_scan_time()
but also adds the cost of the where clause (TIME_FOR_COMPARE).
- Added small cost for creating temporary table for
materialization. This causes some very small tables to use scan
instead of materialization.
- Added checking of the WHERE clause (TIME_FOR_COMPARE) of the
accepted rows to ROR costs in get_best_ror_intersect()
- Removed '- 0.001' from 'join->best_read' and optimize_straight_join()
to ensure that the 'Last_query_cost' status variable contains the
same value as the one that was calculated by the optimizer.
- Take avg_io_cost() into account in handler::keyread_time() and
handler::read_time(). This should have no effect as it's 1.0 by
default, except for heap that overrides these functions.
- Some 'ref_or_null' accesses changed to 'range' because of cost
adjustments (main.order_by)
- Added scan type "scan_with_join_cache" for optimizer_trace. This is
just to show in the trace what kind of scan was used.
- When using 'scan_with_join_cache' take into account number of
preceding tables (as have to restore all fields for all previous
table combination when checking the where clause)
The new cost added is:
(row_combinations * ROW_COPY_COST * number_of_cached_tables).
This increases the cost of join buffering in proportion of the
number of tables in the join buffer. One effect is that full scans
are now done earlier as the cost is then smaller.
(main.join_outer_innodb, main.greedy_optimizer)
- Removed the usage of 'worst_seeks' in cost_for_index_read as it
caused wrong plans to be created; It prefered JT_EQ_REF even if it
would be much more expensive than a full table scan. A related
issue was that worst_seeks only applied to full lookup, not to
clustered or index only lookups, which is not consistent. This
caused some plans to use index scan instead of eq_ref (main.union)
- Changed federated block size from 4096 to 1500, which is the
typical size of an IO packet.
- Added costs for reading rows to Federated. Needed as there is no
caching of rows in the federated engine.
- Added ha_innobase::rnd_pos_time() cost function.
- A lot of extra things added to optimizer trace
- More costs, especially for materialization and index_merge.
- Make lables more uniform
- Fixed a lot of minor bugs
- Added 'trace_started()' around a lot of trace blocks.
- When calculating ORDER BY with LIMIT cost for using an index
the cost did not take into account the number of row retrivals
that has to be done or the cost of comparing the rows with the
WHERE clause. The cost calculated would be just a fraction of
the real cost. Now we calculate the cost as we do for ranges
and 'ref'.
- 'Using index for group-by' is used a bit more than before as
now take into account the WHERE clause cost when comparing
with 'ref' and prefer the method with fewer row combinations.
(main.group_min_max).
Bugs fixed:
- Fixed that we don't calculate TIME_FOR_COMPARE twice for some plans,
like in optimize_straight_join() and greedy_search()
- Fixed bug in save_explain_data where we could test for the wrong
index when displaying 'Using index'. This caused some old plans to
show 'Using index'. (main.subselect_innodb, main.subselect2)
- Fixed bug in get_best_ror_intersect() where 'min_cost' was not
updated, and the cost we compared with was not the one that was
used.
- Fixed very wrong cost calculation for priority queues in
check_if_pq_applicable(). (main.order_by now correctly uses priority
queue)
- When calculating cost of EQ_REF or REF, we added the cost of
comparing the WHERE clause with the found rows, not all row
combinations. This made ref and eq_ref to be regarded way to cheap
compared to other access methods.
- FORCE INDEX cost calculation didn't take into account clustered or
covered indexes.
- JT_EQ_REF cost was estimated as avg_io_cost(), which is half the
cost of a JT_REF key. This may be true for InnoDB primary key, but
not for other unique keys or other engines. Now we use handler
function to calculate the cost, which allows us to handle
consistently clustered, covered keys and not covered keys.
- ha_start_keyread() didn't call extra_opt() if keyread was already
enabled but still changed the 'keyread' variable (which is wrong).
Fixed by not doing anything if keyread is already enabled.
- multi_range_read_info_cost() didn't take into account io_cost when
calculating the cost of ranges.
- fix_semijoin_strategies_for_picked_join_order() used the wrong
record_count when calling best_access_path() for SJ_OPT_FIRST_MATCH
and SJ_OPT_LOOSE_SCAN.
- Hash joins didn't provide correct best_cost to the upper level, which
means that the cost for hash_joins more expensive than calculated
in best_access_path (a difference of 10x * TIME_OF_COMPARE).
This is fixed in the new code thanks to that we now include
TIME_OF_COMPARE cost in 'read_time'.
Other things:
- Added some 'if (thd->trace_started())' to speed up code
- Removed not used function Cost_estimate::is_zero()
- Simplified testing of HA_POS_ERROR in get_best_ror_intersect().
(No cost changes)
- Moved ha_start_keyread() from join_read_const_table() to join_read_const()
to enable keyread for all types of JT_CONST tables.
- Made a few very short functions inline in handler.h
Notes:
- In main.rowid_filter the join order of order and lineitem is swapped.
This is because the cost of doing a range fetch of lineitem(98 rows) is
almost as big as the whole join of order,lineitem. The filtering will
also ensure that we only have to do very small key fetches of the rows
in lineitem.
- main.index_merge_myisam had a few changes where we are now using
less keys for index_merge. This is because index scans are now more
expensive than before.
- handler->optimizer_cache_cost is updated in ha_external_lock().
This ensures that it is up to date per statements.
Not an optimal solution (for locked tables), but should be ok for now.
- 'DELETE FROM t1 WHERE t1.a > 0 ORDER BY t1.a' does not take cost of
filesort into consideration when table scan is chosen.
(main.myisam_explain_non_select_all)
- perfschema.table_aggregate_global_* has changed because an update
on a table with 1 row will now use table scan instead of key lookup.
TODO in upcomming commits:
- Fix selectivity calculation for ranges with and without filtering and
when there is a ref access but scan is chosen.
For this we have to store the lowest known value for
'accepted_records' in the OPT_RANGE structure.
- Change that records_read does not include filtered rows.
- test_if_cheaper_ordering() needs to be updated to properly calculate
costs. This will fix tests like main.order_by_innodb,
main.single_delete_update
- Extend get_range_limit_read_cost() to take into considering
cost_for_index_read() if there where no quick keys. This will reduce
the computed cost for ORDER BY with LIMIT in some cases.
(main.innodb_ext_key)
- Fix that we take into account selectivity when counting the number
of rows we have to read when considering using a index table scan to
resolve ORDER BY.
- Add new calculation for rnd_pos_time() where we take into account the
benefit of reading multiple rows from the same page.
The old code did not't correctly add TIME_FOR_COMPARE to rows that are
part of the scan that will be compared with the attached where clause.
Now the cost calculation for hash join and full join cache join are
identical except for HASH_FANOUT (10%)
The cost for a join with keys is now also uniform.
The total cost for a using a key for lookup is calculated in one place as:
(cost_of_finding_rows_through_key(records) + records/TIME_FOR_COMPARE)*
record_count_of_previous_row_combinations + startup_cost
startup_cost is the cost of a creating a temporary table (if needed)
Best_cost now includes the cost of comparing all WHERE clauses and also
cost of joining with previous row combinations.
Other things:
- Optimizer trace is now printing the total costs, including testing the
WHERE clause (TIME_FOR_COMPARE) and comparing with all previous rows.
- In optimizer trace, include also total cost of query together with the
final join order. This makes it easier to find out where the cost was
calculated.
- Old code used filter even if the cost for it was higher than not using a
filter. This is not corrected.
- When rebasing on 10.11, I noticed some changes to access_cost_factor
calculation. These changes was not picked as the coming changes
to filtering will make that code obsolete.
Fixed also that the 'with_found_constraint parameter' to
matching_candidates_in_table() is as documented: It is now true only
if there is a reference to a previous table in the WHERE condition for
the current examined table (as it was originally documented)
Changes in test results:
- Filtered was 25% smaller for some queries (expected).
- Some join order changed (probably because the tables had very few rows).
- Some more table scans, probably because there would be fewer returned
rows.
- Some tests exposes a bug that if there is more filtered rows, then the
cost for table scan will be higher. This will be fixed in a later commit.
Other changes:
- In test_quick_select(), assume that if table->used_stats_records is 0
then the table has 0 rows.
- Fixed prepare_simple_select() to populate table->used_stat_records
- Enusre that set_statistics_for_tables() doesn't cause used_stats_records
to be 0 when using stat_tables.
- To get blackhole to work with replication, set stats.records to 2 so
that test_quick_select() doesn't assume the table is empty.
The Boolean flag mlog_init_t::init::created was only needed by
mark_ibuf_exist(), which commit f27e9c8947
removed. We only need to store the page initialization LSN in the map.
buf_LRU_get_free_block(): Replace the Boolean parameter with a
ternary parameter, so that have_no_mutex_soft can be specified
reduce the chances of initiating page eviction flushing in read-ahead.
buf_read_acquire(): Invoke buf_LRU_get_free_block(have_no_mutex_soft)
and check in each caller for a nullptr return value.
buf_pool_t::page_hash_contains(): Check if a page is cached.
buf_read_ahead_random(), buf_read_page_background(),
buf_read_ahead_linear(): Before invoking buf_read_page_low(),
preallocate a buffer page for the read request.
buf_read_page(), buf_page_init_for_read(), buf_read_page_low():
Add a parameter for the buf_pool.page_hash chain, to avoid duplicated
computations.
buf_page_t::read_complete(): Only attempt recovery if an uncompressed
page frame has been allocated.
buf_page_init_for_read(): Before trying to acquire buf_pool.mutex, acquire
an exclusive buf_pool.page_hash latch and check if the page is already
located in the buffer pool. If the buf_pool.mutex is not immediately
available, release both latches and acquire them in the correct order,
and then recheck if the page is already in the buffer pool. This should
hopefully reduce some contention on buf_pool.mutex.
buf_page_init_for_read(), buf_read_page_low(): Input the "recovery needed"
flag in the least significant bit of zip_size.
buf_read_acquire(), buf_read_release(): Interface for allocating and
freeing buffer pages for reading.
buf_read_recv_pages(): Set the flag that recovery is needed.
Other ROW_FORMAT=COMPRESSED reads during recovery
will not need any recovery.
This also fixes part of MDEV-29835 Partial server freeze
which is caused by violations of the latching order that was
defined in https://dev.mysql.com/worklog/task/?id=6326
(WL#6326: InnoDB: fix index->lock contention). Unless the
current thread is holding an exclusive dict_index_t::lock,
it must acquire page latches in a strict parent-to-child,
left-to-right order. Not all cases of MDEV-29835 are fixed yet.
Failure to follow the correct latching order will cause deadlocks
of threads due to lock order inversion.
As part of these changes, the BTR_MODIFY_TREE mode is modified
so that an Update latch (U a.k.a. SX) will be acquired on the
root page, and eXclusive latches (X) will be acquired on all pages
leading to the leaf page, as well as any left and right siblings
of the pages along the path. The DEBUG_SYNC test innodb.innodb_wl6326
will be removed, because at the time the DEBUG_SYNC point is hit,
the thread is actually holding several page latches that will be
blocking a concurrent SELECT statement.
We also remove double bookkeeping that was caused due to excessive
information hiding in mtr_t::m_memo. We simply let mtr_t::m_memo
store information of latched pages, and ensure that
mtr_memo_slot_t::object is never a null pointer.
The tree_blocks[] and tree_savepoints[] were redundant.
buf_page_get_low(): If innodb_change_buffering_debug=1, to avoid
a hang, do not try to evict blocks if we are holding a latch on
a modified page. The test innodb.innodb-change-buffer-recovery
will be removed, because change buffering may no longer be forced
by debug injection when the change buffer comprises multiple pages.
Remove a debug assertion that could fail when
innodb_change_buffering_debug=1 fails to evict a page.
For other cases, the assertion is redundant, because we already
checked that right after the got_block: label. The test
innodb.innodb-change-buffering-recovery will be removed, because
due to this change, we will be unable to evict the desired page.
mtr_t::lock_register(): Register a change of a page latch
on an unmodified buffer-fixed block.
mtr_t::x_latch_at_savepoint(), mtr_t::sx_latch_at_savepoint():
Replaced by the use of mtr_t::upgrade_buffer_fix(), which now
also handles RW_S_LATCH.
mtr_t::set_modified(): For temporary tables, invoke
buf_page_t::set_modified() here and not in mtr_t::commit().
We will never set the MTR_MEMO_MODIFY flag on other than
persistent data pages, nor set mtr_t::m_modifications when
temporary data pages are modified.
mtr_t::commit(): Only invoke the buf_flush_note_modification() loop
if persistent data pages were modified.
mtr_t::get_already_latched(): Look up a latched page in mtr_t::m_memo.
This avoids many redundant entries in mtr_t::m_memo, as well as
redundant calls to buf_page_get_gen() for blocks that had already
been looked up in a mini-transaction.
btr_get_latched_root(): Return a pointer to an already latched root page.
This replaces btr_root_block_get() in cases where the mini-transaction
has already latched the root page.
btr_page_get_parent(): Fetch a parent page that was already latched
in BTR_MODIFY_TREE, by invoking mtr_t::get_already_latched().
If needed, upgrade the root page U latch to X.
This avoids bloating mtr_t::m_memo as well as performing redundant
buf_pool.page_hash lookups. For non-QUICK CHECK TABLE as well as for
B-tree defragmentation, we will invoke btr_cur_search_to_nth_level().
btr_cur_search_to_nth_level(): This will only be used for non-leaf
(level>0) B-tree searches that were formerly named BTR_CONT_SEARCH_TREE
or BTR_CONT_MODIFY_TREE. In MDEV-29835, this function could be
removed altogether, or retained for the case of
CHECK TABLE without QUICK.
btr_cur_t::left_block: Remove. btr_pcur_move_backward_from_page()
can retrieve the left sibling from the end of mtr_t::m_memo.
btr_cur_t::open_leaf(): Some clean-up.
btr_cur_t::search_leaf(): Replaces btr_cur_search_to_nth_level()
for searches to level=0 (the leaf level). We will never release
parent page latches before acquiring leaf page latches. If we need to
temporarily release the level=1 page latch in the BTR_SEARCH_PREV or
BTR_MODIFY_PREV latch_mode, we will reposition the cursor on the
child node pointer so that we will land on the correct leaf page.
btr_cur_t::pessimistic_search_leaf(): Implement new BTR_MODIFY_TREE
latching logic in the case that page splits or merges will be needed.
The parent pages (and their siblings) should already be latched on
the first dive to the leaf and be present in mtr_t::m_memo; there
should be no need for BTR_CONT_MODIFY_TREE. This pre-latching almost
suffices; it must be revised in MDEV-29835 and work-arounds removed
for cases where mtr_t::get_already_latched() fails to find a block.
rtr_search_to_nth_level(): A SPATIAL INDEX version of
btr_search_to_nth_level() that can search to any level
(including the leaf level).
rtr_search_leaf(), rtr_insert_leaf(): Wrappers for
rtr_search_to_nth_level().
rtr_search(): Replaces rtr_pcur_open().
rtr_latch_leaves(): Replaces btr_cur_latch_leaves(). Note that unlike
in the B-tree code, there is no error handling in case the sibling
pages are corrupted.
rtr_cur_restore_position(): Remove an unused constant parameter.
btr_pcur_open_on_user_rec(): Remove the constant parameter
mode=PAGE_CUR_GE.
row_ins_clust_index_entry_low(): Use a new
mode=BTR_MODIFY_ROOT_AND_LEAF to gain access to the root page
when mode!=BTR_MODIFY_TREE, to write the PAGE_ROOT_AUTO_INC.
BTR_SEARCH_TREE, BTR_CONT_SEARCH_TREE: Remove.
BTR_CONT_MODIFY_TREE: Note that this is only used by
rtr_search_to_nth_level().
btr_pcur_optimistic_latch_leaves(): Replaces
btr_cur_optimistic_latch_leaves().
ibuf_delete_rec(): Acquire exclusive ibuf.index->lock in order
to avoid a deadlock with ibuf_insert_low(BTR_MODIFY_PREV).
btr_blob_log_check_t(): Acquire a U latch on the root page,
so that btr_page_alloc() in btr_store_big_rec_extern_fields()
will avoid a deadlock.
btr_store_big_rec_extern_fields(): Assert that the root page latch
is being held.
Tested by: Matthias Leich
Reviewed by: Vladislav Lesin
- introduce table key construction function in wsrep service interface
- don't add row keys when replicating bulk insert
- don't start bulk insert on applier or when transaction is not active
- don't start bulk insert on system versioned tables
- implement actual bulk insert table-level key replication
Reviewed-by: Jan Lindström <jan.lindstrom@mariadb.com>
Problem:
========
- InnoDB fails to remove the newly created table or index from
data dictionary and table cache if the alter fails in commit phase
Solution:
========
- InnoDB should restart the transaction to remove the newly
created table and index when it fails in commit phase of an alter
operation. innodb_fts.misc_debug tests the scenario with the
help of debug point "stats_lock_fail"
In commit 24648768b4, some use of
O_DIRECT was added without proper #ifdef guard. That broke the
compilation in environments that do not define O_DIRECT, such as
OpenBSD.
- InnoDB fails to clear the freed ranges during truncation of innodb
undo log tablespace. During shutdown, InnoDB flushes the freed page
ranges and throws the out of bound error.
mtr_t::commit_shrink(): clear the freed ranges while doing undo
tablespace truncation
This also fixes part of MDEV-29835 Partial server freeze
which is caused by violations of the latching order that was
defined in https://dev.mysql.com/worklog/task/?id=6326
(WL#6326: InnoDB: fix index->lock contention). Unless the
current thread is holding an exclusive dict_index_t::lock,
it must acquire page latches in a strict parent-to-child,
left-to-right order. Not all cases are fixed yet. Failure to
follow the correct latching order will cause deadlocks of threads
due to lock order inversion.
As part of these changes, the BTR_MODIFY_TREE mode is modified
so that an Update latch (U a.k.a. SX) will be acquired on the
root page, and eXclusive latches (X) will be acquired on all pages
leading to the leaf page, as well as any left and right siblings
of the pages along the path. The test innodb.innodb_wl6326
will be removed, because at the time the DEBUG_SYNC point is hit,
the thread is actually holding several page latches that will be
blocking a concurrent SELECT statement.
We also remove double bookkeeping that was caused due to excessive
information hiding in mtr_t::m_memo. We simply let mtr_t::m_memo
store information of latched pages, and ensure that
mtr_memo_slot_t::object is never a null pointer.
The tree_blocks[] and tree_savepoints[] were redundant.
mtr_t::get_already_latched(): Look up a latched page in mtr_t::m_memo.
This avoids many redundant entries in mtr_t::m_memo, as well as
redundant calls to buf_page_get_gen() for blocks that had already
been looked up in a mini-transaction.
btr_get_latched_root(): Return a pointer to an already latched root page.
This replaces btr_root_block_get() in cases where the mini-transaction
has already latched the root page.
btr_page_get_parent(): Fetch a parent page that was already latched
in BTR_MODIFY_TREE, by invoking mtr_t::get_already_latched().
If needed, upgrade the root page U latch to X.
This avoids bloating mtr_t::m_memo as well as redundant
buf_pool.page_hash lookups. For non-QUICK CHECK TABLE as well as for
B-tree defragmentation, we will invoke btr_cur_search_to_nth_level().
btr_cur_search_to_nth_level(): This will only be used for non-leaf
(level>0) B-tree searches that were formerly named BTR_CONT_SEARCH_TREE
or BTR_CONT_MODIFY_TREE. In MDEV-29835, this function could be
removed altogether, or retained for the case of
CHECK TABLE without QUICK.
btr_cur_t::search_leaf(): Replaces btr_cur_search_to_nth_level()
for searches to level=0 (the leaf level).
btr_cur_t::pessimistic_search_leaf(): Implement the new
BTR_MODIFY_TREE latching logic in the case that page splits
or merges will be needed. The parent pages (and their siblings)
should already be latched on the first dive to the leaf and be
present in mtr_t::m_memo; there should be no need for
BTR_CONT_MODIFY_TREE. This pre-latching almost suffices;
MDEV-29835 will have to revise it and remove work-arounds where
mtr_t::get_already_latched() fails to find a block.
rtr_search_to_nth_level(): A SPATIAL INDEX version of
btr_search_to_nth_level() that can search to any level
(including the leaf level).
rtr_search_leaf(), rtr_insert_leaf(): Wrappers for
rtr_search_to_nth_level().
rtr_search(): Replaces rtr_pcur_open().
rtr_cur_restore_position(): Remove an unused constant parameter.
btr_pcur_open_on_user_rec(): Remove the constant parameter
mode=PAGE_CUR_GE.
btr_cur_latch_leaves(): Update a pre-existing mtr_t::m_memo entry
for the current leaf page.
row_ins_clust_index_entry_low(): Use a new
mode=BTR_MODIFY_ROOT_AND_LEAF to gain access to the root page
when mode!=BTR_MODIFY_TREE, to write the PAGE_ROOT_AUTO_INC.
btr_cur_t::open_leaf(): Some clean-up.
mtr_t::lock_register(): Register a page latch on a buffer-fixed block.
BTR_SEARCH_TREE, BTR_CONT_SEARCH_TREE: Remove.
BTR_CONT_MODIFY_TREE: Note that this is only used by
rtr_search_to_nth_level().
btr_pcur_optimistic_latch_leaves(): Replaces
btr_cur_optimistic_latch_leaves().
ibuf_delete_rec(): Acquire ibuf.index->lock.u_lock() in order
to avoid a deadlock with ibuf_insert_low(BTR_MODIFY_PREV).
Tested by: Matthias Leich
To avoid heap memory allocation overhead for mtr_t::m_memo,
we will allocate a small number of elements statically in
mtr_t::m_memo::small. Only if that preallocated data is
insufficient, we will invoke my_alloc() or my_realloc() for
more storage. The implementation of the data structure is
inspired by llvm::SmallVector.
The condition is freed in sp_head::execute, after calling
ha_spider::reset. This commit partially reverts the change in commit
e954d9de88, so that the condition is
always freed regardless of the wide_handler->sql_command, which will
prevent access to the freed condition later.
Signed-off-by: Yuchen Pei <yuchen.pei@mariadb.com>
