DELETE query against memory table with btree index may remove
not all matching rows. This happens only when DELETE uses
index read method to find matching rows. E.g. for queries
like DELETE FROM t1 WHERE a=1.
Fixed by reverting fix for BUG9719 and applying proper solution.
Using a MEMORY table BTREE index for scanning for updatable rows
could lead to an infinite loop.
Everytime a key was inserted into a btree index, the position
in the index scan was cleared. The search started from the
beginning and found the same key again.
Now we do not clear the position on key insert an more.
server
The problem was that when memory was exhausted HEAP engine could crash
(GROUP BY uses HEAP TABLE). Alternatively, if SET was used, it could
report an error "You may only use constant expressions with SET" instead
of "Out of memory (Needed NNNNNN bytes)".
The solution is:
- pass MY_WME to (some) calls to my_malloc() to get correct message.
- fix heap_write() so that the first key is skipped during cleanup
on ENOMEM because it wasn't inserted and doesn't have to be
deleted.
No test case is provided because we can't test out-of-memory behaviour
in our current test framework.
Two handler objects were present, one was used for an insert and the other for a select
The state of the statistics was local to the handler object and thus the other handler
object didn't notice the insert.
Fix included:
1) Add a new variable key_stat_version added to whenever statistics was considered in need
of update (previously key_stats_ok= FALSE in those places)
2) Add a new handler variable key_stat_version assigned whenever key_stats_ok= TRUE was set
previously
3) Fix records_in_range to return records if records <= 1
4) Fix records_in_range to add 2 to rec_per_key to ensure we don't specify 0 or 1 when it isn't
and thus invoking incorrect optimisations.
5) Fix unique key handling for HEAP table in records_in_range
On 64 bit platforms the changed statement doesn't work
so making sure that the larger value is first and using
subtraction is a quick and backwards-compatible fix of this
line.
New records_in_range() interface (similar to read_range())
Macros for faster bitmap handling
Simplify read_range() code (#WL1786)
New general key_cmp() function to compare keys
HEAP: Copies the key count to a backup variable and sets the key count to zero.
That way, no HEAP function will ever try to touch any index.
Re-enabling is done by copying back the backup variable.
To avoid memory leak at table close, disable deletes all index trees.
Re-enabling must be done with empty indexes and data anyway. Otherwise,
the indexes would need to be repaired, wich HEAP is not capable of.
MyISAM: Only the key_map is cleared and set.
Re-enabling must be done with empty indexes and data. Otherwise, repair needs
to be done which will enable all keys persistently.
The former implementation disabled only non-unique keys and maked this persistent.
The new implementation additionally can disable all keys, but only without
making this persistent. Re-enabling all keys can be done without repair,
if data file and indexes are empty.
Don't read character set files if we are using only the default charset. In most cases the user will not anymore get a warning about missing character set files
Compare strings with space extend instead of space strip. Now the following comparisons holds: "a" == "a " and "a\t" < "a". (Bug #3152).
Note: Because of the above fix, one has to do a REPAIR on any table that has an ascii character < 32 last in a CHAR/VARCHAR/TEXT columns.
Added more DBUG statements
Ensure that we are comparing end space with BINARY strings
Use 'any_db' instead of '' to mean any database. (For HANDLER command)
Only strip ' ' when comparing CHAR, not other space-like characters (like \t)
when we open the HEAP table for the first time since server restart,
in hp_open(), we set a flag to propagate this info to the handler level
which then writes a DELETE FROM this_heap_table to the binlog.
It is not a perfect solution for the bug, because between the server start and
the first open of the table, the slave still had old data in his table so
a SELECT on the slave may show wrong content. But if there is a --init-file
to populate the HEAP table on master as startup, then this is a safe fix
(I'll put a note about init-file in the HEAP section of the manual).
Signed auto_increment keys for HASH tables (like for MyISAM tables in 4.0)
nitialize system_charset_info() early. Fixes core dump when starting windows service
CREATE TABLE t1 (s1 CHAR(5) CHARACTER SET latin1 COLLATE latin1_german2_ci);
INSERT INTO t1 VALUES ('Ü');
INSERT INTO t1 VALUES ('ue');
SELECT DISTINCT s1 FROM t1;
The above returned two rows in error.
Now it returns one row, in latin1_german2_ci: Ü == ue
