Support index condition pushdown within partitioned tables.
- ha_partition will pass the pushed index condition into all of the used
partitions.
- We require that all of the partitions to handle the pushed index
condition in the same way.
- When using ICP, one may read rows (e.g. call h->index_read_map(buf, ...)
only to buf= table->record[0], for two reasons:
* Pushed index condition's Item_field objects point into record[0]
* InnoDB requires this: it calls offset() which assumes record[0].
So, when using ICP, ha_partition will read partition records to
table->record[0] and then will copy record away if it needs it to be
elsewhere.
- Use log2() insted of log()
- Added missing ''+' when calculating rowid setup cost
- Adjusted ROWID_FILTER_PER_ELEMENT_MODIFIER (from 3 to 1)
Other things:
- Adjusted cost for index_merge where rows_out < 1.0
The effects of the changes:
- rowid filter will have higher setup cost
- rowid filter will have slightly less costs per row
This can be seen in mtr where some tests, with 'small tables or
that uses rowid filters with many rows, will not use rowid filter anymore.
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)
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)
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.
- multi_range_read_info_const now uses the new records_in_range interface
- Added handler::avg_io_cost()
- Don't calculate avg_io_cost() in get_sweep_read_cost if avg_io_cost is
not 1.0. In this case we trust the avg_io_cost() from the handler.
- Changed test_quick_select to use TIME_FOR_COMPARE instead of
TIME_FOR_COMPARE_IDX to align this with the rest of the code.
- Fixed bug when using test_if_cheaper_ordering where we didn't use
keyread if index was changed
- Fixed a bug where we didn't use index only read when using order-by-index
- Added keyread_time() to HEAP.
The default keyread_time() was optimized for blocks and not suitable for
HEAP. The effect was the HEAP prefered table scans over ranges for btree
indexes.
- Fixed get_sweep_read_cost() for HEAP tables
- Ensure that range and ref have same cost for simple ranges
Added a small cost (MULTI_RANGE_READ_SETUP_COST) to ranges to ensure
we favior ref for range for simple queries.
- Fixed that matching_candidates_in_table() uses same number of records
as the rest of the optimizer
- Added avg_io_cost() to JT_EQ_REF cost. This helps calculate the cost for
HEAP and temporary tables better. A few tests changed because of this.
- heap::read_time() and heap::keyread_time() adjusted to not add +1.
This was to ensure that handler::keyread_time() doesn't give
higher cost for heap tables than for normal tables. One effect of
this is that heap and derived tables stored in heap will prefer
key access as this is now regarded as cheap.
- Changed cost for index read in sql_select.cc to match
multi_range_read_info_const(). All index cost calculation is now
done trough one function.
- 'ref' will now use quick_cost for keys if it exists. This is done
so that for '=' ranges, 'ref' is prefered over 'range'.
- scan_time() now takes avg_io_costs() into account
- get_delayed_table_estimates() uses block_size and avg_io_cost()
- Removed default argument to test_if_order_by_key(); simplifies code
This patch contains a full implementation of the optimization
that allows to use in-memory rowid / primary filters built for range
conditions over indexes. In many cases usage of such filters reduce
the number of disk seeks spent for fetching table rows.
In this implementation the choice of what possible filter to be applied
(if any) is made purely on cost-based considerations.
This implementation re-achitectured the partial implementation of
the feature pushed by Galina Shalygina in the commit
8d5a11122c.
Besides this patch contains a better implementation of the generic
handler function handler::multi_range_read_info_const() that
takes into account gaps between ranges when calculating the cost of
range index scans. It also contains some corrections of the
implementation of the handler function records_in_range() for MyISAM.
This patch supports the feature for InnoDB and MyISAM.
