The following problems are also known and fixing them is also a high priority:
MySQL Cluster fails to recover from an out-of-disk failure when using disk data. (Bug#17614)
Subquery optimization for IN
is not as
effective as for =
.
Even if you use
lower_case_table_names=2
(which enables
MySQL to remember the case used for databases and table
names), MySQL does not remember the case used for database
names for the function
DATABASE()
or within the
various logs (on case-insensitive systems).
Dropping a FOREIGN KEY
constraint
doesn't work in replication because the constraint may
have another name on the slave.
REPLACE
(and
LOAD DATA
with the
REPLACE
option) does not
trigger ON DELETE CASCADE
.
DISTINCT
with ORDER
BY
doesn't work inside
GROUP_CONCAT()
if you don't
use all and only those columns that are in the
DISTINCT
list.
If one user has a long-running transaction and another
user drops a table that is updated in the transaction,
there is small chance that the binary log may contain the
DROP TABLE
command before
the table is used in the transaction itself. We plan to
fix this by having the DROP
TABLE
command wait until the table is not being
used in any transaction.
When inserting a big integer value (between 263 and 264–1) into a decimal or string column, it is inserted as a negative value because the number is evaluated in a signed integer context.
FLUSH TABLES WITH
READ LOCK
does not block
COMMIT
if the server is
running without binary logging, which may cause a problem
(of consistency between tables) when doing a full backup.
ANALYZE TABLE
,
OPTIMIZE TABLE
, and
REPAIR TABLE
may cause
problems on tables for which you are using
INSERT DELAYED
.
Performing LOCK TABLE ...
and
FLUSH TABLES ...
doesn't guarantee that
there isn't a half-finished transaction in progress on the
table.
Replication uses query-level logging: The master writes the executed queries to the binary log. This is a very fast, compact, and efficient logging method that works perfectly in most cases.
It is possible for the data on the master and slave to become different if a query is designed in such a way that the data modification is nondeterministic (generally not a recommended practice, even outside of replication).
For example:
CREATE
TABLE ... SELECT
or
INSERT
... SELECT
statements that insert zero or
NULL
values into an
AUTO_INCREMENT
column.
DELETE
if you are
deleting rows from a table that has foreign keys with
ON DELETE CASCADE
properties.
REPLACE ...
SELECT
, INSERT IGNORE ...
SELECT
if you have duplicate key values in
the inserted data.
If and only if the preceding queries
have no ORDER BY
clause guaranteeing a
deterministic order.
For example, for
INSERT ...
SELECT
with no ORDER BY
, the
SELECT
may return rows in a
different order (which results in a row having different
ranks, hence getting a different number in the
AUTO_INCREMENT
column), depending on
the choices made by the optimizers on the master and
slave.
A query is optimized differently on the master and slave only if:
The table is stored using a different storage engine
on the master than on the slave. (It is possible to
use different storage engines on the master and slave.
For example, you can use InnoDB
on
the master, but MyISAM
on the slave
if the slave has less available disk space.)
MySQL buffer sizes
(key_buffer_size
, and
so on) are different on the master and slave.
The master and slave run different MySQL versions, and the optimizer code differs between these versions.
This problem may also affect database restoration using mysqlbinlog|mysql.
The easiest way to avoid this problem is to add an
ORDER BY
clause to the aforementioned
nondeterministic queries to ensure that the rows are
always stored or modified in the same order.
In future MySQL versions, we will automatically add an
ORDER BY
clause when needed.
The following issues are known and will be fixed in due time:
Log file names are based on the server host name (if you
don't specify a file name with the startup option). You
have to use options such as
--log-bin=
if you change your host name to something else. Another
option is to rename the old files to reflect your host
name change (if these are binary logs, you need to edit
the binary log index file and fix the binary log file
names there as well). See
Section 5.1.2, “Server Command Options”.
old_host_name
-bin
mysqlbinlog does not delete temporary
files left after a
LOAD DATA
INFILE
command. See
Section 4.6.7, “mysqlbinlog — Utility for Processing Binary Log Files”.
RENAME
doesn't work with
TEMPORARY
tables or tables used in a
MERGE
table.
Due to the way table format (.frm
)
files are stored, you cannot use character 255
(CHAR(255)
) in table names, column
names, or enumerations. This is scheduled to be fixed in
version 5.1 when we implement new table definition format
files.
When using SET CHARACTER SET
, you can't
use translated characters in database, table, and column
names.
You can't use “_
” or
“%
” with
ESCAPE
in
LIKE ...
ESCAPE
.
You cannot build the server in another directory when using MIT-pthreads. Because this requires changes to MIT-pthreads, we are not likely to fix this. See Section 2.3.5, “MIT-pthreads Notes”.
BLOB
and
TEXT
values can't reliably
be used in GROUP BY
, ORDER
BY
or DISTINCT
. Only the
first max_sort_length
bytes are used when comparing
BLOB
values in these cases.
The default value of
max_sort_length
is 1024
and can be changed at server startup time or at runtime.
Numeric calculations are done with
BIGINT
or
DOUBLE
(both are normally
64 bits long). Which precision you get depends on the
function. The general rule is that bit functions are
performed with BIGINT
precision, IF()
and
ELT()
with
BIGINT
or
DOUBLE
precision, and the
rest with DOUBLE
precision.
You should try to avoid using unsigned long long values if
they resolve to be larger than 63 bits
(9223372036854775807) for anything other than bit fields.
In MIN()
,
MAX()
, and other aggregate
functions, MySQL currently compares
ENUM
and
SET
columns by their string
value rather than by the string's relative position in the
set.
mysqld_safe redirects all messages from
mysqld to the mysqld
log. One problem with this is that if you execute
mysqladmin refresh to close and reopen
the log, stdout
and
stderr
are still redirected to the old
log. If you use the general query log extensively, you
should edit mysqld_safe to log to
instead of
host_name
.err
so that you can easily reclaim the space for the old log
by deleting it and executing mysqladmin
refresh.
host_name
.log
In an UPDATE
statement,
columns are updated from left to right. If you refer to an
updated column, you get the updated value instead of the
original value. For example, the following statement
increments KEY
by 2
,
not 1
:
mysql> UPDATE tbl_name
SET KEY=KEY+1,KEY=KEY+1;
You can refer to multiple temporary tables in the same query, but you cannot refer to any given temporary table more than once. For example, the following doesn't work:
mysql> SELECT * FROM temp_table, temp_table AS t2;
ERROR 1137: Can't reopen table: 'temp_table'
The optimizer may handle DISTINCT
differently when you are using “hidden”
columns in a join than when you are not. In a join, hidden
columns are counted as part of the result (even if they
are not shown), whereas in normal queries, hidden columns
don't participate in the DISTINCT
comparison. We will probably change this in the future to
never compare the hidden columns when executing
DISTINCT
.
An example of this is:
SELECT DISTINCT mp3id FROM band_downloads WHERE userid = 9 ORDER BY id DESC;
and
SELECT DISTINCT band_downloads.mp3id FROM band_downloads,band_mp3 WHERE band_downloads.userid = 9 AND band_mp3.id = band_downloads.mp3id ORDER BY band_downloads.id DESC;
In the second case, using MySQL Server 3.23.x, you may get
two identical rows in the result set (because the values
in the hidden id
column may differ).
Note that this happens only for queries where that do not
have the ORDER BY
columns in the
result.
If you execute a PROCEDURE
on a query
that returns an empty set, in some cases the
PROCEDURE
does not transform the
columns.
Creation of a table of type MERGE
doesn't check whether the underlying tables are compatible
types.
If you use ALTER TABLE
to
add a UNIQUE
index to a table used in a
MERGE
table and then add a normal index
on the MERGE
table, the key order is
different for the tables if there was an old,
non-UNIQUE
key in the table. This is
because ALTER TABLE
puts
UNIQUE
indexes before normal indexes to
be able to detect duplicate keys as early as possible.
User Comments
Add your own comment.