This patch changes how old rows in mysql.gtid_slave_pos* tables are deleted.
Instead of doing it as part of every replicated transaction in
record_gtid(), it is done periodically (every @@gtid_cleanup_batch_size
transaction) in the slave background thread.
This removes the deletion step from the replication process in SQL or worker
threads, which could speed up replication with many small transactions. It
also decreases contention on the global mutex LOCK_slave_state. And it
simplifies the logic, eg. when a replicated transaction fails after having
deleted old rows.
With this patch, the deletion of old GTID rows happens asynchroneously and
slightly non-deterministic. Thus the number of old rows in
mysql.gtid_slave_pos can temporarily exceed @@gtid_cleanup_batch_size. But
all old rows will be deleted eventually after sufficiently many new GTIDs
have been replicated.
sorting them by usernames first, and then by get_sort() value.
Search functions now use binary search to find the the first entry with
given name. Then, linear search is done, until the first match.
The previous patch 269da4bf19 was
actually for MDEV-8894 (not for MDEV-5377). It was erroneously
pushed with a wrong title.
This patch is a small cleanup for MDEV-8894.
CREATE TABLE is now not a part of binary logs recorded with MySQL,
only INSERT statements are. This will allow to reuse the same binary
logs in combinations with different CREATE TABLE statements,
to tests different data types.
- This is a regression of commit b26e603aeb. While dropping
the incompletely created table, InnoDB shouldn't consider that operation as non-atomic one.
When the with clause of a query contains a recursive CTE that is not used
then processing of EXPLAIN for this query does not require optimization
of the unit specifying this CTE. In this case if 'derived' is the
TABLE_LIST object created for this CTE then derived->derived_result is NULL
and any assignment to derived->derived_result->table causes a crash.
After fixing this problem in the code of st_select_lex_unit::prepare()
EXPLAIN for such a query worked without crashes. Yet an execution
plan for the recursive CTE appeared there. The cause of this problem was
an incorrect condition used in JOIN::save_explain_data_intern() that
determined whether CTE was to be optimized or not. A similar condition was
used in select_describe() and this patch has corrected it as well.
In 10.3, rec_is_metadata() takes a pointer, while in 10.4 it
takes a reference as a parameter. I ported this patch from
10.4 to 10.3, and then only ran a release build, not debug build.
The special flag REC_INFO_MIN_REC_FLAG used to be only set on the
first record in the leftmost node pointer page of each level of the tree.
It was never set on leaf pages.
MDEV-11369 Instant ADD COLUMN in MariaDB Server 10.3 repurposed the flag
to identify a hidden metadata record, which is stored in the first record
on the leftmost leaf page.
If the adaptive hash index points to records in the leftmost leaf page
after instant ALTER TABLE, we would have such a metadata record in the
table, an assertion could fail when trying to validate the index record.
In a release build, we might wrongly qualify the hidden metadata record
and thus return garbage results.
cmp_dtuple_rec_with_match_bytes(): If the REC_INFO_MIN_REC_FLAG is
set on the record, assert that this is the first record on the
leftmost page and that the record is a metadata record, and finally
return 1, because by definition, anything is greater than the
minimum record.
Also, related to MDEV-15522, MDEV-17304, MDEV-17835,
remove the Galera xtrabackup tests, because xtrabackup never worked
with MariaDB Server 10.3 due to InnoDB redo log format changes.
dict_create_add_foreigns_to_dictionary(): Do not commit the transaction.
The operation can still fail in dict_load_foreigns(), and we want
to be able to roll back the transaction.
create_table_info_t::create_table(): Never reset m_drop_before_rollback,
and never commit the transaction. We use a single point of rollback
in ha_innobase::create(). Merge the logic from
row_table_add_foreign_constraints().
This is a regression due to MDEV-17816.
When creating a table fails, we must roll back the dictionary
transaction. Because the rollback may rename tables, and because
InnoDB lacks proper undo logging for CREATE operations, we must
drop the incompletely created table before rolling back the
transaction, which could include a RENAME operation.
But, we must not blindly drop the table by name; after all,
the operation could have failed because another table by the
same name already existed.
create_table_info_t::m_drop_before_rollback: A flag that is set
if the table needs to be dropped before transaction rollback.
create_table_info_t::create_table(): Remove some duplicated
error handling.
ha_innobase::create(): On error, only drop the table if it was
actually created.
Added support for usual agreggate UDF (UDAF)
Added remove() call support for more efficient window function processing
Added example of aggregate UDF with efficient windows function support
fil_space_t::add(): Replaces fil_node_create(), fil_node_create_low().
Let the caller pass fil_node_t::handle, to avoid having to close and
re-open files.
fil_node_t::read_page0(): Refactored from fil_node_open_file().
Read the first page of a data file.
fil_node_open_file(): Open the file only once.
srv_undo_tablespace_open(): Set the file handle for the opened
undo tablespace. This should ensure that ut_ad(file->is_open())
no longer fails in recv_add_trim().
xtrabackup_backup_func(): Remove some dead code.
xb_fil_cur_open(): Open files only if needed. Undo tablespaces
should already have been opened.