BUG#54872 MBR: replication failure caused by using tmp table inside transaction
Changed criteria to classify a statement as unsafe in order to reduce the
number of spurious warnings. So a statement is classified as unsafe when
there is on-going transaction at any point of the execution if:
1. The mixed statement is about to update a transactional table and
a non-transactional table.
2. The mixed statement is about to update a temporary transactional
table and a non-transactional table.
3. The mixed statement is about to update a transactional table and
read from a non-transactional table.
4. The mixed statement is about to update a temporary transactional
table and read from a non-transactional table.
5. The mixed statement is about to update a non-transactional table
and read from a transactional table when the isolation level is
lower than repeatable read.
After updating a transactional table if:
6. The mixed statement is about to update a non-transactional table
and read from a temporary transactional table.
7. The mixed statement is about to update a non-transactional table
and read from a temporary transactional table.
8. The mixed statement is about to update a non-transactionala table
and read from a temporary non-transactional table.
9. The mixed statement is about to update a temporary non-transactional
table and update a non-transactional table.
10. The mixed statement is about to update a temporary non-transactional
table and read from a non-transactional table.
11. A statement is about to update a non-transactional table and the
option variables.binlog_direct_non_trans_update is OFF.
The reason for this is that locks acquired may not protected a concurrent
transaction of interfering in the current execution and by consequence in
the result. So the patch reduced the number of spurious unsafe warnings.
Besides we fixed a regression caused by BUG#51894, which makes temporary
tables to go into the trx-cache if there is an on-going transaction. In
MIXED mode, the patch for BUG#51894 ignores that the trx-cache may have
updates to temporary non-transactional tables that must be written to the
binary log while rolling back the transaction.
So we fix this problem by writing the content of the trx-cache to the
binary log while rolling back a transaction if a non-transactional
temporary table was updated and the binary logging format is MIXED.
errors
In the fix of BUG#39934 in 5.1-rep+3, errors are generated when
binlog_format=row and a statement modifies a table restricted to
statement-logging (ER_BINLOG_ROW_MODE_AND_STMT_ENGINE); or if
binlog_format=statement and a statement modifies a table restricted to
row-logging (ER_BINLOG_STMT_MODE_AND_ROW_ENGINE).
However, some DDL statements that lock tables (e.g. ALTER TABLE,
CREATE INDEX and CREATE TRIGGER) were causing spurious errors,
although no row might be inserted into the binary log.
To fix the problem, we tagged statements that may generate
rows into the binary log and thence the warning messages are
only printed out when the appropriate conditions hold and rows
might be changed.
Clarified error messages related to unsafe statements:
- avoid the internal technical term "row injection"
- use 'binary log' instead of 'binlog'
- avoid the word 'unsafeness'
transaction
BUG#52616 Temp table prevents switch binlog format from STATEMENT to ROW
Before the WL#2687 and BUG#46364, every non-transactional change that happened
after a transactional change was written to trx-cache and flushed upon
committing the transaction. WL#2687 and BUG#46364 changed this behavior and
non-transactional changes are now written to the binary log upon committing
the statement.
A binary log event is identified as transactional or non-transactional through
a flag in the Log_event which is set taking into account the underlie storage
engine on what it is stems from. In the current bug, this flag was not being
set properly when the DROP TEMPORARY TABLE was executed.
However, while fixing this bug we figured out that changes to temporary tables
should be always written to the trx-cache if there is an on-going transaction.
Otherwise, binlog events in the reversed order would be produced.
Regarding concurrency, keeping changes to temporary tables in the trx-cache is
also safe as temporary tables are only visible to the owner connection.
In this patch, we classify the following statements as unsafe:
1 - INSERT INTO t_myisam SELECT * FROM t_myisam_temp
2 - INSERT INTO t_myisam_temp SELECT * FROM t_myisam
3 - CREATE TEMPORARY TABLE t_myisam_temp SELECT * FROM t_myisam
On the other hand, the following statements are classified as safe:
1 - INSERT INTO t_innodb SELECT * FROM t_myisam_temp
2 - INSERT INTO t_myisam_temp SELECT * FROM t_innodb
The patch also guarantees that transactions that have a DROP TEMPORARY are
always written to the binary log regardless of the mode and the outcome:
commit or rollback. In particular, the DROP TEMPORARY is extended with the
IF EXISTS clause when the current statement logging format is set to row.
Finally, the patch allows to switch from STATEMENT to MIXED/ROW when there
are temporary tables but the contrary is not possible.
BUG#46364 introduced the flag binlog_direct_non_transactional_updates which
would make N-changes to be written to the binary log upon committing the
statement when "ON". On the other hand, when "OFF" the option was supposed
to mimic the behavior in 5.1. However, the implementation was not mimicking
the behavior correctly and the following bugs popped up:
Case #1: N-changes executed within a transaction would go into
the S-cache. When later in the same transaction a
T-change occurs, N-changes following it were written
to the T-cache instead of the S-cache. In some cases,
this raises problems. For example, a
Table_map_log_event being written initially into the
S-cache, together with the initial N-changes, would be
absent from the T-cache. This would log N-changes
orphaned from a Table_map_log_event (thence discarded
at the slave). (MIXED and ROW)
Case #2: When rolling back a transaction, the N-changes that
might be in the T-cache were disregarded and
truncated along with the T-changes. (MIXED and ROW)
Case #3: When a MIXED statement (TN) is ahead of any other
T-changes in the transaction and it fails, it is kept
in the T-cache until the transaction ends. This is
not the case in 5.1 or Betony (5.5.2). In these, the
failed TN statement would be written to the binlog at
the same instant it had failed and not deferred until
transaction end. (SBR)
To fix these problems, we have decided to do what follows:
For Case #1 and #2, we circumvent them:
1. by not letting binlog_direct_non_transactional_updates
affect MIXED and RBR. These modes will keep the behavior
provided by WL#2687. Although this will make Celosia to
behave differently from 5.1, an execution will be always
safe under such modes in the sense that slaves will never
go out sync. In 5.1, using either MIXED or ROW while
mixing N-statements and T-statements was not safe.
For Case #3, we don't actually fix it. We:
1. keep it and make all MIXED statements whether they end
up failing or not or whether they are up front in the
transaction or after some transactional change to always
be stored in the T-cache. This means that it is written
to the binary log on transaction commit/rollback only.
2. We make the warning message even more specific about the
MIXED statement and SBR.
Conflicts:
- mysql-test/r/mysqld--help-win.result
- sql/sys_vars.cc
Original revsion (in next-mr-bugfixing):
------------------------------------------------------------
revno: 2971 [merge]
revision-id: alfranio.correia@sun.com-20100121210527-rbuheu5rnsmcakh1
committer: Alfranio Correia <alfranio.correia@sun.com>
branch nick: mysql-next-mr-bugfixing
timestamp: Thu 2010-01-21 21:05:27 +0000
message:
BUG#46364 MyISAM transbuffer problems (NTM problem)
It is well-known that due to concurrency issues, a slave can become
inconsistent when a transaction contains updates to both transaction and
non-transactional tables.
In a nutshell, the current code-base tries to preserve causality among the
statements by writing non-transactional statements to the txn-cache which
is flushed upon commit. However, modifications done to non-transactional
tables on behalf of a transaction become immediately visible to other
connections but may not immediately get into the binary log and therefore
consistency may be broken.
In general, it is impossible to automatically detect causality/dependency
among statements by just analyzing the statements sent to the server. This
happen because dependency may be hidden in the application code and it is
necessary to know a priori all the statements processed in the context of
a transaction such as in a procedure. Moreover, even for the few cases that
we could automatically address in the server, the computation effort
required could make the approach infeasible.
So, in this patch we introduce the option
- "--binlog-direct-non-transactional-updates" that can be used to bypass
the current behavior in order to write directly to binary log statements
that change non-transactional tables.
Besides, it is used to enable the WL#2687 which is disabled by default.
------------------------------------------------------------
revno: 2970.1.1
revision-id: alfranio.correia@sun.com-20100121131034-183r4qdyld7an5a0
parent: alik@sun.com-20100121083914-r9rz2myto3tkdya0
committer: Alfranio Correia <alfranio.correia@sun.com>
branch nick: mysql-next-mr-bugfixing
timestamp: Thu 2010-01-21 13:10:34 +0000
message:
BUG#46364 MyISAM transbuffer problems (NTM problem)
It is well-known that due to concurrency issues, a slave can become
inconsistent when a transaction contains updates to both transaction and
non-transactional tables.
In a nutshell, the current code-base tries to preserve causality among the
statements by writing non-transactional statements to the txn-cache which
is flushed upon commit. However, modifications done to non-transactional
tables on behalf of a transaction become immediately visible to other
connections but may not immediately get into the binary log and therefore
consistency may be broken.
In general, it is impossible to automatically detect causality/dependency
among statements by just analyzing the statements sent to the server. This
happen because dependency may be hidden in the application code and it is
necessary to know a priori all the statements processed in the context of
a transaction such as in a procedure. Moreover, even for the few cases that
we could automatically address in the server, the computation effort
required could make the approach infeasible.
So, in this patch we introduce the option
- "--binlog-direct-non-transactional-updates" that can be used to bypass
the current behavior in order to write directly to binary log statements
that change non-transactional tables.
Besides, it is used to enable the WL#2687 which is disabled by default.
Non-transactional updates that take place inside a transaction present problems
for logging because they are visible to other clients before the transaction
is committed, and they are not rolled back even if the transaction is rolled
back. It is not always possible to log correctly in statement format when both
transactional and non-transactional tables are used in the same transaction.
In the current patch, we ensure that such scenario is completely safe under the
ROW and MIXED modes.
General overview:
The logic for switching to row format when binlog_format=MIXED had
numerous flaws. The underlying problem was the lack of a consistent
architecture.
General purpose of this changeset:
This changeset introduces an architecture for switching to row format
when binlog_format=MIXED. It enforces the architecture where it has
to. It leaves some bugs to be fixed later. It adds extensive tests to
verify that unsafe statements work as expected and that appropriate
errors are produced by problems with the selection of binlog format.
It was not practical to split this into smaller pieces of work.
Problem 1:
To determine the logging mode, the code has to take several parameters
into account (namely: (1) the value of binlog_format; (2) the
capabilities of the engines; (3) the type of the current statement:
normal, unsafe, or row injection). These parameters may conflict in
several ways, namely:
- binlog_format=STATEMENT for a row injection
- binlog_format=STATEMENT for an unsafe statement
- binlog_format=STATEMENT for an engine only supporting row logging
- binlog_format=ROW for an engine only supporting statement logging
- statement is unsafe and engine does not support row logging
- row injection in a table that does not support statement logging
- statement modifies one table that does not support row logging and
one that does not support statement logging
Several of these conflicts were not detected, or were detected with
an inappropriate error message. The problem of BUG#39934 was that no
appropriate error message was written for the case when an engine
only supporting row logging executed a row injection with
binlog_format=ROW. However, all above cases must be handled.
Fix 1:
Introduce new error codes (sql/share/errmsg.txt). Ensure that all
conditions are detected and handled in decide_logging_format()
Problem 2:
The binlog format shall be determined once per statement, in
decide_logging_format(). It shall not be changed before or after that.
Before decide_logging_format() is called, all information necessary to
determine the logging format must be available. This principle ensures
that all unsafe statements are handled in a consistent way.
However, this principle is not followed:
thd->set_current_stmt_binlog_row_based_if_mixed() is called in several
places, including from code executing UPDATE..LIMIT,
INSERT..SELECT..LIMIT, DELETE..LIMIT, INSERT DELAYED, and
SET @@binlog_format. After Problem 1 was fixed, that caused
inconsistencies where these unsafe statements would not print the
appropriate warnings or errors for some of the conflicts.
Fix 2:
Remove calls to THD::set_current_stmt_binlog_row_based_if_mixed() from
code executed after decide_logging_format(). Compensate by calling the
set_current_stmt_unsafe() at parse time. This way, all unsafe statements
are detected by decide_logging_format().
Problem 3:
INSERT DELAYED is not unsafe: it is logged in statement format even if
binlog_format=MIXED, and no warning is printed even if
binlog_format=STATEMENT. This is BUG#45825.
Fix 3:
Made INSERT DELAYED set itself to unsafe at parse time. This allows
decide_logging_format() to detect that a warning should be printed or
the binlog_format changed.
Problem 4:
LIMIT clause were not marked as unsafe when executed inside stored
functions/triggers/views/prepared statements. This is
BUG#45785.
Fix 4:
Make statements containing the LIMIT clause marked as unsafe at
parse time, instead of at execution time. This allows propagating
unsafe-ness to the view.
The Blackhole engine did not support row-based replication
since the delete_row(), update_row(), and the index and range
searching functions were not implemented.
This patch adds row-based replication support for the
Blackhole engine by implementing the two functions mentioned
above, and making the engine pretend that it has found the
correct row to delete or update when executed from the slave
SQL thread by implementing index and range searching functions.
It is necessary to only pretend this for the SQL thread, since
a SELECT executed on the Blackhole engine will otherwise never
return EOF, causing a livelock.
The bug allow multiple executing transactions working with non-transactional
to interfere with each others by interleaving the events of different trans-
actions.
Bug is fixed by writing non-transactional events to the transaction cache and
flushing the cache to the binary log at statement commit. To mimic the behavior
of normal statement-based replication, we flush the transaction cache in row-
based mode when there is no committed statements in the transaction cache,
which means we are committing the first one. This means that it will be written
to the binary log as a "mini-transaction" with just the rows for the statement.
Note that the changes here does not take effect when building the server with
HAVE_TRANSACTIONS set to false, but it is not clear if this was possible before
this patch either.
For row-based logging, we also have that when AUTOCOMMIT=1, the code now always
generates a BEGIN/COMMIT pair for single statements, or BEGIN/ROLLBACK pair in the
case of non-transactional changes in a statement that was rolled back. Note that
for the case where changes to a non-transactional table causes a rollback due
to error, the statement will now be logged with a BEGIN/ROLLBACK pair, even
though some changes has been committed to the non-transactional table.
Now, every transaction (including autocommit transactions) starts with
a BEGIN and ends with a COMMIT/ROLLBACK in the binlog.
Added a test case, and updated lots of test case result files.
