revno: 4559
committer: Marc Alff <marc.alff@oracle.com>
branch nick: mysql-5.6-bug14741537-v4
timestamp: Thu 2012-11-08 22:40:31 +0100
message:
Bug#14741537 - MYSQL 5.6, GTID AND PERFORMANCE_SCHEMA
Before this fix, statements using performance_schema tables:
- were marked as unsafe for replication,
- did cause warnings during execution,
- were written to the binlog, either in STATEMENT or ROW format.
When using replication with the new GTID feature,
unsafe warnings are elevated to errors,
which prevents to use both the performance_schema and GTID together.
The root cause of the problem is not related to raising warnings/errors
in some special cases, but deeper: statements involving the performance
schema should not even be written to the binary log in the first place,
because the content of the performance schema tables is 'local' to a server
instance, and may differ greatly between nodes in a replication
topology.
In particular, the DBA should be able to configure (INSERT, UPDATE, DELETE)
or flush (TRUNCATE) performance schema tables on one node,
without affecting other nodes.
This fix introduces the concept of a 'non-replicated' or 'local' table,
and adjusts the replication logic to ignore tables that are not replicated
when deciding if or how to log a statement to the binlog.
Note that while this issue was detected using the performance_schema,
other tables are also affected by the same problem.
This fix define as 'local' the following tables, which are then never
replicated:
- performance_schema.*
- mysql.general_log
- mysql.slow_log
- mysql.slave_relay_log_info
- mysql.slave_master_info
- mysql.slave_worker_info
Existing behavior for information_schema.* is unchanged by this fix,
to limit the scope of changes.
Coding wise, this fix implements the following changes:
1)
Performance schema tables are not using any replication flags,
since performance schema tables are not replicated.
2)
In open_table_from_share(),
tables with no replication capabilities (performance_schema.*),
tables with TABLE_CATEGORY_LOG (logs)
and tables with TABLE_CATEGORY_RPL_INFO (replication)
are marked as non replicated, with TABLE::no_replicate
3)
A new THD member, THD::m_binlog_filter_state,
indicate if the current statement is written to the binlog
(normal cases for most statements), or is to be discarded
(because the statements affects non replicated tables).
4)
In THD::decide_logging_format(), the replication logic
is changed to take into account non replicated tables.
Statements that affect only non replicated tables are
executed normally (no warning or errors), but not written
to the binlog.
Statements that affect (i.e., write to) a replicated table
while also using (i.e., reading from or writing to) a non replicated table
are executed normally in MIXED and ROW binlog format,
and cause a new error in STATEMENT binlog format.
THD::decide_logging_format() uses THD::m_binlog_filter_state
to indicate if a statement is to be ignored, when writing to
the binlog.
5)
In THD::binlog_query(), statements marked as ignored
are not written to the binary log.
6)
For row based replication, the existing test for 'table->no_replicate',
has been moved from binlog_log_row() to check_table_binlog_row_based().
mysql-test/r/mdl_sync.result:
Full merge with 5.6
mysql-test/t/mdl_sync.test:
Full merge with 5.6
sql/debug_sync.cc:
Full merge with 5.6
sql/debug_sync.h:
Full merge with 5.6
sql/mdl.cc:
Full merge with 5.6
sql/sql_base.cc:
Removed code not in 5.6 anymore
Implement discovery of table non-existence, and related changes:
1. Split GTS_FORCE_DISCOVERY (that was meaning two different things in
two different functions) into GTS_FORCE_DISCOVERY and GTS_USE_DISCOVERY.
2. Move GTS_FORCE_DISCOVERY implementation into open_table_def().
3. In recover_from_failed_open() clear old errors *before* discovery,
not after successful discovery. The final error should come
from the discovery.
4. On forced discovery delete table .frm first. Discovery will write
a new one, if desired.
5. If the frm file exists, but not the table in the engine, force
rediscovery if the engine supports it.
1. default db type for partitions was stored as 1-byte DB_TYPE code,
which doesn't work for dynamically generated codes.
2. storage engine plugin for default db type wasn't locked at all,
which could trivially crash for dynamic plugins.
Now the storage engine name is stored in the extra2 section,
and the plugin is correctly locked.
Partitioning didn't store the name of default storage engine for partitions
in the frm file - it only store the typecode. Typecodes aren't stable, and
might vary depending on the order in which storage engines are loaded (can
be changed even from my.cnf without recompilation).
As a temporary workaround for 5.5, we hard-code Aria's typecode, to make sure it
never changes.
Merge of 10.0-mdev26 feature tree into 10.0-base.
Global transaction ID is prepended to each event group in the binlog.
Slave connect can request to start from GTID position instead of specifying
file name/offset of master binlog. This facilitates easy switch to a new
master.
Slave GTID state is stored in a table mysql.rpl_slave_state, which can be
InnoDB to get crash-safe slave state.
GTID includes a replication domain ID, allowing to keep track of distinct
positions for each of multiple masters.
* persistent table versions in the extra2
* ha_archive::frm_compare using TABLE_SHARE::tabledef_version
* distinguish between "important" and "optional" extra2 frm values
* write engine-defined attributes (aka "table options") to extra2, not to extra,
but still read from the old location, if they're found there.
* comments
* cosmetic changes, *(ptr+5) -> ptr[5]
* a couple of trivial functions -> inline
* remove unused argument from pack_header()
* create_frm() no longer creates frm file (the function used to prepare and
fill a memory buffer and call my_create at the end. Now it only prepares
a memory buffer). Renamed accordingly.
* don't call pack_screen twice, go for a smaller screen area in the first attempt
* remove useless calls to check_duplicate_warning()
* don't write unireg screens to .frm files
* remove make_new_entry(), it's basically dead code, always calculating
and writing into frm the same string value. replace the function call
with the constant string.
* print "table doesn't exist in engine" when a table doesn't exist in the engine,
instead of "file not found" (if no file was involved)
* print a complete filename that cannot be found ('t1.MYI', not 't1')
* it's not an error for a DROP if a table doesn't exist in the engine (or some table
files cannot be found) - if the DROP succeeded regardless
and let open_binary_frm to parse it from the buffer, not a file.
this avoids jumping back in forth in the frm file, and doing
intermediate buffer mallocs.
