"Federared Transactions Failure"
Bug occurs when the user performs an operation which inserts more than
one row into the federated table and the federated table references a
remote table stored within a transactional storage engine. When the
insert operation for any one row in the statement fails due to
constraint violation, the federated engine is unable to perform
statement rollback and so the remote table contains a partial commit.
The user would expect a statement to perform the same so a statement
rollback is expected.
This bug was fixed by implementing bulk-insert handling into the
federated storage engine. This will relieve the bug for most common
situations by enabling the generation of a multi-row insert into the
remote table and thus permitting the remote table to perform
statement rollback when neccessary.
The multi-row insert is limited to the maximum packet size between
servers and should the size overflow, more than one insert statement
will be sent and this bug will reappear. Multi-row insert is disabled
when an "INSERT...ON DUPLICATE KEY UPDATE" is being performed.
The bulk-insert handling will offer a significant performance boost
when inserting a large number of small rows.
This patch builds on Bug29019 and Bug25511
"Federated INSERT failures"
Federated does not correctly handle "INSERT...ON DUPLICATE KEY UPDATE"
However, implementing such support is not reasonably possible without
increasing complexity of the storage engine: checking that constraints
on remote server match local server and parsing error messages.
This patch causes 'ON DUPLICATE KEY' to fail with ER_DUP_KEY message
if a conflict occurs and not to fail silently.
"REPLACE/INSERT IGNORE/UPDATE IGNORE doesn't work"
Federated does not record neccessary HA_EXTRA flags in order to
support REPLACE/INSERT IGNORE/UPDATE IGNORE.
Implement ::extra() to capture flags neccessary for functionality.
New function append_ident() to better escape identifiers consistantly.
the loose scan optimization for grouping queries was applied returned
a wrong result set when the query was used with the SQL_BIG_RESULT
option.
The SQL_BIG_RESULT option forces to use sorting algorithm for grouping
queries instead of employing a suitable index. The current loose scan
optimization is applied only for one table queries when the suitable
index is covering. It does not make sense to use sort algorithm in this
case. However the create_sort_index function does not take into account
the possible choice of the loose scan to implement the DISTINCT operator
which makes sorting unnecessary. Moreover the current implementation of
the loose scan for queries with distinct assumes that sorting will
never happen. Thus in this case create_sort_index should not call
the function filesort.
INSERT into table from SELECT from the same table
with ORDER BY and LIMIT was inserting other data
than sole SELECT ... ORDER BY ... LIMIT returns.
One part of the patch for bug #9676 improperly pushed
LIMIT to temporary table in the presence of the ORDER BY
clause.
That part has been removed.
long shared-memory-base-names could overflow a static internal buffer
and thus crash mysqld and various clients. change both to dynamic
buffers, show everything but overflowing those buffers still works.
The test case for this would pretty much amount to
mysqld --shared-memory-base-name=HeyMrBaseNameXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX --shared-memory=1 &
mysqladmin --no-defaults --shared-memory-base-name=HeyMrBaseNameXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX shutdown
Unfortunately, we can't just use an .opt file for the
server. The .opt file is used at start-up, before any
include in the actual test can tell mysqltest to skip
this one on non-Windows. As a result, such a test would
break on unices.
Fixing mysql-test-run.pl to export full path for master
and slave would enable us to start a server from within
the test which is ugly and, what's more, doesn't work as
the server blocks (mysqltest offers no fire-and-forget
fork-and-exec), and mysqladmin never gets run.
Making the test rpl_windows_shm or some such so we can
is beyond ugly. As is introducing another file-name based
special case (run "win*.test" only when on Windows). As is
(yuck) coding half the test into mtr (as in, having it
hand out a customized environment conductive to the shm-
thing on Win only).
Situation is exacerbated by the fact that .sh is not
necessary run as expected on Win.
In short, it's just not worth it. No test-case until we
have a new-and-improved test framework.
Occasionally mysqlbinlog --hexdump failed with error:
ERROR 1064 (42000) at line ...: You have an error in your
SQL syntax; check the manual that corresponds to your MySQL
server version for the right syntax to use near
'Query thread_id=... exec_time=... error_code=...
When the length of hexadecimal dump of binlog header was
divisible by 16, commentary sign '#' after header was lost.
The Log_event::print_header function has been modified to always
finish hexadecimal binlog header with "\n# ".
The abort happened when a query contained a conjunctive predicate
of the form 'view column = constant' in the WHERE condition and
the grouping list also contained a reference to a view column yet
a different one.
Removed the failing assertion as invalid in a general case.
Also fixed a bug that prevented applying some optimization for grouping
queries using views. If the WHERE condition of such a query contains
a conjunctive condition of the form 'view column = constant' and
this view column is used in the grouping list then grouping by this
column can be eliminated. The bug blocked performing this elimination.
and replicated):
A DROP USER statement with a non-existing user was correctly written to
the binary log (there might be users that were removed, but not all),
but the error code was not set, which caused the slave to stop with an
error.
The error reporting code was moved to before the statement was logged
to ensure that the error information for the thread was correctly set
up. This works since my_error() will set the fields net.last_errno and
net.last_error for the thread that is reporting the error, and this
will then be picked up when the Query_log_event is created and written
to the binary log.
slave_sql thread calls thd->clear_error() to force error to be ignored,
though this method didn't clear thd->killed state, what causes
slave_sql thread to stop.
clear thd->killed state if we ignore an error
For a join query with GROUP BY and/or ORDER BY and a view reference
in the FROM list the metadata erroneously showed empty table aliases
and database names for the view columns.
Changed code to enforce that SQL_CACHE only in the first SELECT is used to turn on caching(as documented), but any SQL_NO_CACHE will turn off caching (not documented, but a useful behaviour, especially for machine generated queries). Added test cases to explicitly test the documented caching behaviour and test cases for the reported bug.
represented by an expression of the type UNSIGNED INT and this
expression was evaluated to 0 then the function erroneously returned
the value of the first argument instead of an empty string.
This problem was introduced by the patch for bug 10963.
The problem has been resolved by a proper modification of the code of
Item_func_substr::val_str.
DECIMAL column was used instead of BIGINT for the minimal possible
BIGINT (-9223372036854775808).
The Item_func_neg::fix_length_and_dec has been adjusted to
to inherit the type of the argument in the case when it's an
Item_int object whose value is equal to LONGLONG_MIN.
Problem: Temporary buffer which is used for quoting and escaping
was initialized to character set utf8, and thus didn't allow
to store data in other character sets.
Fix: changing character set of the buffer to be able to
store any arbitrary sequence of bytes.
of its arguments was evaluated to NULL, while the predicate
LOCATE(str,NULL) IS NULL erroneously was evaluated to FALSE.
This happened because the Item_func_locate::fix_length_and_dec
method by mistake set the value of the maybe_null flag for
the function item to 0. In consequence of this the function
was considered as the one that could not ever return NULL.
In many cases, binaries can no longer dump core after calling setuid().
Where the PR_SET_DUMPABLE macro is set, use the prctl() system call
to tell the kernel that it's allowed to dump the core of the server.
