2617.31.12, 2617.31.15, 2617.31.15, 2617.31.16, 2617.43.1
- initial changeset that introduced the fix for
Bug#989 and follow up fixes for all test suite failures
introduced in the initial changeset.
------------------------------------------------------------
revno: 2617.31.1
committer: Davi Arnaut <Davi.Arnaut@Sun.COM>
branch nick: 4284-6.0
timestamp: Fri 2009-03-06 19:17:00 -0300
message:
Bug#989: If DROP TABLE while there's an active transaction, wrong binlog order
WL#4284: Transactional DDL locking
Currently the MySQL server does not keep metadata locks on
schema objects for the duration of a transaction, thus failing
to guarantee the integrity of the schema objects being used
during the transaction and to protect then from concurrent
DDL operations. This also poses a problem for replication as
a DDL operation might be replicated even thought there are
active transactions using the object being modified.
The solution is to defer the release of metadata locks until
a active transaction is either committed or rolled back. This
prevents other statements from modifying the table for the
entire duration of the transaction. This provides commitment
ordering for guaranteeing serializability across multiple
transactions.
- Incompatible change:
If MySQL's metadata locking system encounters a lock conflict,
the usual schema is to use the try and back-off technique to
avoid deadlocks -- this schema consists in releasing all locks
and trying to acquire them all in one go.
But in a transactional context this algorithm can't be utilized
as its not possible to release locks acquired during the course
of the transaction without breaking the transaction commitments.
To avoid deadlocks in this case, the ER_LOCK_DEADLOCK will be
returned if a lock conflict is encountered during a transaction.
Let's consider an example:
A transaction has two statements that modify table t1, then table
t2, and then commits. The first statement of the transaction will
acquire a shared metadata lock on table t1, and it will be kept
utill COMMIT to ensure serializability.
At the moment when the second statement attempts to acquire a
shared metadata lock on t2, a concurrent ALTER or DROP statement
might have locked t2 exclusively. The prescription of the current
locking protocol is that the acquirer of the shared lock backs off
-- gives up all his current locks and retries. This implies that
the entire multi-statement transaction has to be rolled back.
- Incompatible change:
FLUSH commands such as FLUSH PRIVILEGES and FLUSH TABLES WITH READ
LOCK won't cause locked tables to be implicitly unlocked anymore.
One statement that have more than one different tables to update with
autoinc columns just was marked as unsafe in mixed mode, so the unsafe
warning can't be produced in statement mode.
To fix the problem, mark the statement as unsafe in statement mode too.
binlog, replication aborts
In SBR or MBR, the schema name is not being written to the binlog
when executing a LOAD DATA statement. This becomes a problem when
the current database (lets call it db1) is different from the
table's schema (lets call it db2). For instance, take the
following statements:
use db1;
load data local infile 'infile.txt' into table db2.t
Should this statement be logged without t's schema (db2), when
replaying it, one can get db1.t populated instead of db2.t (if
db1.t exists). On the other hand, if there is no db1.t at all,
replication will stop.
We fix this by always logging the table (in load file) with fully
qualified name when its schema is different from the current
database or when no default database was selected.
The BINLOG statement was sharing too much code with the slave SQL thread, introduced with
the patch for Bug#32407. This caused statements to be logged with the wrong server_id, the
id stored inside the events of the BINLOG statement rather than the id of the running
server.
Fix by rearranging code a bit so that only relevant parts of the code are executed by
the BINLOG statement, and the server_id of the server executing the statements will
not be overrided by the server_id stored in the 'format description BINLOG statement'.
Let
- T be a transactional table and N non-transactional table.
- B be begin, C commit and R rollback.
- N be a statement that accesses and changes only N-tables.
- T be a statement that accesses and changes only T-tables.
In RBR, changes to N-tables that happen early in a transaction are not immediately flushed
upon committing a statement. This behavior may, however, break consistency in the presence
of concurrency since changes done to N-tables become immediately visible to other
connections. To fix this problem, we do the following:
. B N N T C would log - B N C B N C B T C.
. B N N T R would log - B N C B N C B T R.
Note that we are not preserving history from the master as we are introducing a commit that
never happened. However, this seems to be more acceptable than the possibility of breaking
consistency in the presence of concurrency.
Let
- T be a transactional table and N non-transactional table.
- B be begin, C commit and R rollback.
- M be a mixed statement, i.e. a statement that updates both T and N.
- M* be a mixed statement that fails while updating either T or N.
This patch restore the behavior presented in 5.1.37 for rows either produced in
the RBR or MIXED modes, when a M* statement that happened early in a transaction
had their changes written to the binary log outside the boundaries of the
transaction and wrapped in a BEGIN/ROLLBACK. This was done to keep the slave
consistent with with the master as the rollback would keep the changes on N and
undo them on T. In particular, we do what follows:
. B M* T C would log - B M* R B T C.
Note that, we are not preserving history from the master as we are introducing a
rollback that never happened. However, this seems to be more acceptable than
making the slave diverge. We do not fix the following case:
. B T M* C would log B T M* C.
The slave will diverge as the changes on T tables that originated from the M
statement are rolled back on the master but not on the slave. Unfortunately, we
cannot simply rollback the transaction as this would undo any uncommitted
changes on T tables.
SBR is not considered in this patch because a failing statement is written to
the binary along with the error code and a slave executes and then rolls back
the statement when it has an associated error code, thus undoing the effects
on T. In RBR and MBR, a full-fledged fix will be pushed after the WL 2687.
The problem is that there is only one autoinc value associated with
the query when binlogging. If more than one autoinc values are used
in the query, the autoinc values after the first one can be inserted
wrongly on slave. So these autoinc values can become inconsistent on
master and slave.
The problem is resolved by marking all the statements that invoke
a trigger or call a function that updated autoinc fields as unsafe,
and will switch to row-format in Mixed mode. Actually, the statement
is safe if just one autoinc value is used in sub-statement, but it's
impossible to check how many autoinc values are used in sub-statement.)
vs not null
NOTE: Backporting the patch to next-mr.
The replication was generating corrupted data, warning messages on Valgrind
and aborting on debug mode while replicating a "null" to "not null" field.
Specifically the unpack_row routine, was considering the slave's table
definition and trying to retrieve a field value, where there was nothing to be
retrieved, ignoring the fact that the value was defined as "null" by the master.
To fix the problem, we proceed as follows:
1 - If it is not STRICT sql_mode, implicit default values are used, regardless
if it is multi-row or single-row statement.
2 - However, if it is STRICT mode, then a we do what follows:
2.1 If it is a transactional engine, we do a rollback on the first NULL that is
to be set into a NOT NULL column and return an error.
2.2 If it is a non-transactional engine and it is the first row to be inserted
with multi-row, we also return the error. Otherwise, we proceed with the
execution, use implicit default values and print out warning messages.
Unfortunately, the current patch cannot mimic the behavior showed by the master
for updates on multi-tables and multi-row inserts. This happens because such
statements are unfolded in different row events. For instance, considering the
following updates and strict mode:
(master)
create table t1 (a int);
create table t2 (a int not null);
insert into t1 values (1);
insert into t2 values (2);
update t1, t2 SET t1.a=10, t2.a=NULL;
t1 would have (10) and t2 would have (0) as this would be handled as a
multi-row update. On the other hand, if we had the following updates:
(master)
create table t1 (a int);
create table t2 (a int);
(slave)
create table t1 (a int);
create table t2 (a int not null);
(master)
insert into t1 values (1);
insert into t2 values (2);
update t1, t2 SET t1.a=10, t2.a=NULL;
On the master t1 would have (10) and t2 would have (NULL). On
the slave, t1 would have (10) but the update on t1 would fail.
NOTE: Backporting the patch to next-mr.
The reason of the bug was incompatibile with the master side behaviour.
INSERT query on the master is allowed to insert into a table without specifying
values of DEFAULT-less fields if sql_mode is not strict.
Fixed with checking sql_mode by the sql thread to decide how to react.
Non-strict sql_mode should allow Write_rows event to complete.
todo: warnings can be shown via show slave status, still this is a
separate rather general issue how to show warnings for the slave threads.
files
NOTE: this is the backport to next-mr.
SHOW BINLOG EVENTS does not work with relay log files. If issuing
"SHOW BINLOG EVENTS IN 'relay-log.000001'" in a non-empty relay
log file (relay-log.000001), mysql reports empty set.
This patch addresses this issue by extending the SHOW command
with RELAYLOG. Events in relay log files can now be inspected by
issuing SHOW RELAYLOG EVENTS [IN 'log_name'] [FROM pos] [LIMIT
[offset,] row_count].
In RBR, 'DROP TEMPORARY TABLE IF EXISTS...' statement is binlogged when the table
does not exist.
In fact, 'DROP TEMPORARY TABLE ...' statement should never be binlogged in RBR
no matter if the table exists or not.
This patch addresses this by checking whether we are dropping a
temporary table or not, when building the custom drop statement.
failure to cleanup of table
The test case was not dropping a table before exiting (ie, it was
not cleaning itself after execution). In this case, the warning
message stating that the test did not do a proper cleanup was
deterministic (which can be annoying).
I have found other tests cases on which mtr sporadically reports
that they have not cleaned up after execution:
- rpl_ndb_circular
- rpl_failed_optimize
In this case, the master was dropping a table but there was no
synchronization between the slave and the master.
This patch addresses the rpl_ndb_circular_simplex case by adding
the missing DROP table. The other cases are fixed by deploying
the missing sync_slave_with_master instruction.
Network error happened here, but it can be caused by CR_CONNECTION_ERROR,
CR_CONN_HOST_ERROR, CR_SERVER_GONE_ERROR, CR_SERVER_LOST, ER_CON_COUNT_ERROR,
and ER_SERVER_SHUTDOWN. We just check CR_SERVER_LOST here, so the test fails.
To fix the problem, check all errors that can be cause by the master shutdown.
When setting AUTOCOMMIT=1 after starting a transaction, the binary log
did not commit the outstanding transaction. The reason was that the binary
log commit function saw the values of the new settings, deciding that there
were nothing to commit.
Fixed the problem by moving the implicit commit to before the thread option
flags were changed, so that the binary log sees the old values of the flags
instead of the values they will take after the statement.
In RBR, There is an inconsistency between slaves and master.
When INSERT statement which includes an auto_increment field is executed,
Store engine of master will check the value of the auto_increment field.
It will generate a sequence number and then replace the value, if its value is NULL or empty.
if the field's value is 0, the store engine will do like encountering the NULL values
unless NO_AUTO_VALUE_ON_ZERO is set into SQL_MODE.
In contrast, if the field's value is 0, Store engine of slave always generates a new sequence number
whether or not NO_AUTO_VALUE_ON_ZERO is set into SQL_MODE.
SQL MODE of slave sql thread is always consistency with master's.
Another variable is related to this bug.
If generateing a sequence number is decided by the values of
table->auto_increment_field_not_null and SQL_MODE(if includes MODE_NO_AUTO_VALUE_ON_ZERO)
The table->auto_increment_is_not_null is FALSE, which causes this bug to appear. ..
Essentially, Bug#45574 results in this bug. The 'CREATE DATABASE IF NOT EXISTS' statement was not
binlogged, when the database has existed.
Sometimes, the master and slaves become inconsistent. The "CREATE DATABASE
IF NOT EXISTS mysqltest1" statement is not binlogged
if the db 'mysqltest1' existed before the test case is executed.
So the db 'mysqltest1' can't be created on slave.
Patch of Bug#45574 has resolved this problem.
But I think it is better to replace 'mysqltest1' by default db 'test'.
Slave does not correctly handle "expected errors" leading to inconsistencies
between the mater and slave. Specifically, when a statement changes both
transactional and non-transactional tables, the transactional changes are
automatically rolled back on the master but the slave ignores the error and
does not roll them back thus leading to inconsistencies.
To fix the problem, we automatically roll back a statement that fails on
the slave but note that the transaction is not rolled back unless a "rollback"
command is in the relay log file.
binlog
Mixing transactional (T) and non-transactional (N) tables on behalf of a
transaction may lead to inconsistencies among masters and slaves in STATEMENT
mode. The problem stems from the fact that although modifications done to
non-transactional tables on behalf of a transaction become immediately visible
to other connections they do not immediately get to the binary log and therefore
consistency is broken. Although there may be issues in mixing T and M tables in
STATEMENT mode, there are safe combinations that clients find useful.
In this bug, we fix the following issue. Mixing N and T tables in multi-level
(e.g. a statement that fires a trigger) or multi-table table statements (e.g.
update t1, t2...) were not handled correctly. In such cases, it was not possible
to distinguish when a T table was updated if the sequence of changes was N and T.
In a nutshell, just the flag "modified_non_trans_table" was not enough to reflect
that both a N and T tables were changed. To circumvent this issue, we check if an
engine is registered in the handler's list and changed something which means that
a T table was modified.
Check WL 2687 for a full-fledged patch that will make the use of either the MIXED or
ROW modes completely safe.
The "get_master_version_and_clock(...)" function in sql/slave.cc ignores
error and passes directly when queries fail, or queries succeed
but the result retrieved is empty.
The "get_master_version_and_clock(...)" function should try to reconnect master
if queries fail because of transient network problems, and fail otherwise.
The I/O thread should print a warning if the some system variables do not
exist on master (very old master)
"create as select" (innodb table)
Problem: code constructing "CREATE TABLE..." statement
doesn't take into account that current database is not set
in some cases. That may lead to a server crash.
Fix: check if current database is set.
The test case added failed sporadically on PB. This is due to the
fact that the user thread in some cases is waiting for slave IO
to stop and then check the error number. Thence, sometimes the
user thread would race for the error number with IO thread.
This post push fix addresses this by replacing the wait for slave
io to stop with a wait for slave io error (as it seems it was
added in 6.0 also after patch on which this is based was
pushed). This implied backporting wait_for_slave_io_error.inc
from 6.0 also.
The server was not cleaning the last IO error and error number when
resetting slave.
This patch addresses this issue by backporting into 5.1 part of the
patch in BUG 34654. A fix for this issue had already been pushed into
6.0 as part of the aforementioned bug, however the patch also included
some refactoring. The fix for 5.1 does not take into account the
refactoring part.
1. Test case was rewritten completely.
2. Test covers 3 cases:
a) do deadlock on slave, wait retries of transaction, unlock slave before lock
timeout;
b) do deadlock on slave and wait error 'lock timeout exceed' on slave;
c) same as b) but if of max relay log size = 0;
3. Added comments inline.
4. Updated result file.
LOAD_FILE
LOAD_FILE is not safe to replicate in STATEMENT mode, because it
depends on a file (which is loaded on master and may not exist in
slave(s)). This leads to scenarios on which the slave replicates the
statement with 'load_file' and it will try to load the file from local
file system. Given that the file may not exist in the slave filesystem
the operation will not succeed (probably returning NULL), causing
master and slave(s) to diverge. However, when using MIXED mode
replication, this can be made to work, if the statement including
LOAD_FILE is marked as unsafe, triggering a switch to ROW mode,
meaning that the contents of the file are written to binlog as row
events. Consequently, the contents from the file in the master will
reach the slave via the binlog.
This patch addresses this bug by marking the load_file function as
unsafe. When in mixed mode and when LOAD_FILE is issued, there will be
a switch to row mode. Furthermore, when in statement mode, the
LOAD_FILE will raise a warning that the statement is unsafe in that
mode.
TRUNCATE TABLE fails to replicate when stmt-based binlogging is not supported.
There were two separate problems with the code, both of which are fixed with
this patch:
1. An error was printed by InnoDB for TRUNCATE TABLE in statement mode when
the in isolation levels READ COMMITTED and READ UNCOMMITTED since InnoDB
does permit statement-based replication for DML statements. However,
the TRUNCATE TABLE is not transactional, but is a DDL, and should therefore
be allowed to be replicated as a statement.
2. The statement was not logged in mixed mode because of the error above, but
the error was not reported to the client.
This patch fixes the problem by treating TRUNCATE TABLE a DDL, that is, it is
always logged as a statement and not reporting an error from InnoDB for TRUNCATE
TABLE.
Documented behaviour was broken by the patch for bug 33699
that actually is not a bug.
This fix reverts patch for bug 33699 and reverts the
UPDATE of NOT NULL field with NULL query to old
behavior.
Remove size of binlog file from SHOW BINARY LOGS.
Changing size of binlog file is an affect of adding or removing events to/from
binlog and it can be checked in next command of test: SHOW BINLOG EVENTS.
For SHOW BINARY LOGS statement enough to show the list of file names.
conflicts:
Text conflict in client/mysqltest.cc
Text conflict in mysql-test/include/wait_until_connected_again.inc
Text conflict in mysql-test/lib/mtr_report.pm
Text conflict in mysql-test/mysql-test-run.pl
Text conflict in mysql-test/r/events_bugs.result
Text conflict in mysql-test/r/log_state.result
Text conflict in mysql-test/r/myisam_data_pointer_size_func.result
Text conflict in mysql-test/r/mysqlcheck.result
Text conflict in mysql-test/r/query_cache.result
Text conflict in mysql-test/r/status.result
Text conflict in mysql-test/suite/binlog/r/binlog_index.result
Text conflict in mysql-test/suite/binlog/r/binlog_innodb.result
Text conflict in mysql-test/suite/rpl/r/rpl_packet.result
Text conflict in mysql-test/suite/rpl/t/rpl_packet.test
Text conflict in mysql-test/t/disabled.def
Text conflict in mysql-test/t/events_bugs.test
Text conflict in mysql-test/t/log_state.test
Text conflict in mysql-test/t/myisam_data_pointer_size_func.test
Text conflict in mysql-test/t/mysqlcheck.test
Text conflict in mysql-test/t/query_cache.test
Text conflict in mysql-test/t/rpl_init_slave_func.test
Text conflict in mysql-test/t/status.test
