Heap tables are allocated blocks to store rows according to
my_default_record_cache (mapped to the server global variable
read_buffer_size).
This causes performance issues when the record length is big
(> 1000 bytes) and the my_default_record_cache is small.
Changed to instead split the default heap allocation to 1/16 of the
allowed space and not use my_default_record_cache anymore when creating
the heap. The allocation is also aligned to be just under a power of 2.
For some test that I have been running, which was using record length=633,
the speed of the query doubled thanks to this change.
Other things:
- Fixed calculation of max_records passed to hp_create() to take
into account padding between records.
- Updated calculation of memory needed by heap tables. Before we
did not take into account internal structures needed to access rows.
- Changed block sized for memory_table from 1 to 16384 to get less
fragmentation. This also avoids a problem where we need 1K
to manage index and row storage which was not counted for before.
- Moved heap memory usage to a separate test for 32 bit.
- Allocate all data blocks in heap in powers of 2. Change reported
memory usage for heap to reflect this.
Reviewed-by: Sergei Golubchik <serg@mariadb.org>
When parsing statements like (SELECT .. FROM ..) ORDER BY <expr>,
there is a step LEX::add_tail_to_query_expression_body_ext_parens()
which calls LEX::wrap_unit_into_derived(). After that the statement
looks like SELECT * FROM (SELECT .. FROM ..), and parser's
Lex_order_limit_lock structure (ORDER BY <expr>) is assigned to
the new SELECT. But what is missing here is that Items in
Lex_order_limit_lock are left with their original name resolution
contexts, and fix_fields() later resolves the names incorrectly.
For example, when processing
(SELECT * FROM t1 JOIN t2 ON a=b) ORDER BY a
Item_field 'a' in the ORDER BY clause is left with the name resolution
context of the derived table (first_name_resolution_table='t1'), so
it is resolved to 't1.a', which is incorrect.
After LEX::wrap_unit_into_derived() the statement looks like
SELECT * FROM (SELECT * FROM t1 JOIN t2 ON a=b) AS '__2' ORDER BY a,
and the name resolution context for Item_field 'a' in the ORDER BY
must be set to the wrapping SELECT's one.
This commit fixes the issue by changing context for Items in
Lex_order_limit_lock after LEX::wrap_unit_into_derived().
test_if_skip_sort_order() should catch the join types JT_EQ_REF,
JT_CONST and JT_SYSTEM and skip sort order for these.
Such join types imply retrieving of a single row of data, and sorting
of a single row can always be skipped.
A subquery in form "(SELECT not_null_value LIMIT 1 OFFSET 1)" will
produce no rows which will translate into scalar SQL NULL value.
The code in Item_singlerow_subselect::fix_length_and_dec() failed to
take the LIMIT/OFFSET clause into account and used to set
item_subselect->maybe_null=0, despite that SQL NULL will be produced.
If such subselect was used in ORDER BY, this would cause a crash in
filesort() code when it would get a NULL value for a not-nullable item.
also made subselect_engine::no_tables() const function.
ANALYZE FORMAT=JSON output now includes table.r_engine_stats which
has the engine statistics. Only non-zero members are printed.
Internally: EXPLAIN data structures Explain_table_acccess and
Explain_update now have handler* handler_for_stats pointer.
It is used to read statistics from handler_for_stats->handler_stats.
The following applies only to 10.9+, backport doesn't use it:
Explain data structures exist after the tables are closed. We avoid
walking invalid pointers using this:
- SQL layer calls Explain_query::notify_tables_are_closed() before
closing tables.
- After that call, printing of JSON output is disabled. Non-JSON output
can be printed but we don't access handler_for_stats when doing that.
If a select query contained an ORDER BY clause that followed a LIMIT clause
or an ORDER BY clause or ORDER BY with LIMIT the EXPLAIN output for the
query showed an execution plan different from that was actually executed.
Approved by Roman Nozdrin <roman.nozdrin@mariadb.com>
For a correlated subquery filesort is executed multiple times.
During each execution, sortlength() computed total sort key length in
Sort_keys::sort_length, without resetting it first.
Eventually Sort_keys::sort_length got larger than @@sort_buffer_size, which
caused filesort() to be aborted with error.
Fixed by making sortlength() to compute lengths only during the first
invocation. Subsequent invocations return pre-computed values.
and
MDEV-23414 Assertion `res->charset() == item->collation.collation' failed in Type_handler_string_result::make_packed_sort_key_part
pack_sort_string() *must* take a collation from the Item, not from the
String value. Because when casting a string to _binary the original
String is not copied for performance reasons, it's reused but its
collation does not match Item's collation anymore.
Note, that String's collation cannot be simply changed to _binary,
because for an Item_string literal the original String must stay
unchanged for the duration of the query.
this partially reverts 61c15ebe32
An overflow was happening with LONGTEXT columns, when the length was converted to the length
in the strxfrm form (mem-comparable keys).
Introduced a function to truncate the length to the max_sort_length before calculating
the length of the strxfrm form.
For the case when the SJM scan table is the first table in the join order,
then if we want to do the sorting on the SJM scan table, then we need to
make sure that we unpack the values to base table fields in two cases:
1) Reading the SJM table and writing the sort-keys inside the sort-buffer
2) Reading the sorted data from the sort file
In the merge_buffers phase for sorting, the sort buffer size is divided between the number of chunks.
The chunks have a start and end position (m_buffer_start and m_buffer_end).
Then we read the as many records that fit in this buffer for a chunk of the file.
The issue here was we were resetting the end of buffer(m_buffer_end) to the number of bytes that was
read, this was causing a problem because with dynamic size of sort keys it is possible that later
we would not be able to accommodate even one key inside a chunk of file.
So the fix was to not reset the end of buffer for a chunk of file.
The issue here is charset for Sort_param::tmp_buffer is cleared when bzero is done for Sort_param.
Make sure to set the charset explicitly in the constructor for tmp_buffer.
