A query with a group by and having clauses could return a wrong
result set if the having condition contained a constant conjunct
evaluated to FALSE.
It happened because the pushdown condition for table with
grouping columns lost its constant conjuncts.
Pushdown conditions are always built by the function make_cond_for_table
that ignores constant conjuncts. This is apparently not correct when
constant false conjuncts are present.
The bug was as follows: When merge_key_fields() encounters "t.key=X OR t.key=Y" it will
try to join them into ref_or_null access via "t.key=X OR NULL". In order to make this
inference it checks if Y<=>NULL, ignoring the fact that value of Y may be not yet known.
The fix is that the check if Y<=>NULL is made only if value of Y is known (i.e. it is a
constant).
TODO: When merging to 5.0, replace used_tables() with const_item() everywhere in merge_key_fields().
The reason of the bug is in that `get_var_with_binlog' performs missed
assingment of
the variables as side-effect. Doing that it eventually calls
`free_underlaid_joins' to pass as an argument `thd->lex->select_lex' of the lex
which belongs to the user query, not
to one which is emulated i.e SET @var1:=NULL.
`get_var_with_binlog' is refined to supply a temporary lex to sql_set_variables's stack.
mysqldump / SHOW CREATE TABLE will show the NEXT available value for
the PK, rather than the *first* one that was available (that named in
the original CREATE TABLE ... AUTO_INCREMENT = ... statement).
This should produce correct and robust behaviour for the obvious use
cases -- when no data were inserted, then we'll produce a statement
featuring the same value the original CREATE TABLE had; if we dump
with values, INSERTing the values on the target machine should set the
correct next_ID anyway (and if not, we'll still have our AUTO_INCREMENT =
... to do that). Lastly, just the CREATE statement (with no data) for
a table that saw inserts would still result in a table that new values
could safely be inserted to).
There seems to be no robust way however to see whether the next_ID
field is > 1 because it was set to something else with CREATE TABLE
... AUTO_INCREMENT = ..., or because there is an AUTO_INCREMENT column
in the table (but no initial value was set with AUTO_INCREMENT = ...)
and then one or more rows were INSERTed, counting up next_ID. This
means that in both cases, we'll generate an AUTO_INCREMENT =
... clause in SHOW CREATE TABLE / mysqldump. As we also show info on,
say, charsets even if the user did not explicitly give that info in
their own CREATE TABLE, this shouldn't be an issue.
As per above, the next_ID will be affected by any INSERTs that have
taken place, though. This /should/ result in correct and robust
behaviour, but it may look non-intuitive to some users if they CREATE
TABLE ... AUTO_INCREMENT = 1000 and later (after some INSERTs) have
SHOW CREATE TABLE give them a different value (say, CREATE TABLE
... AUTO_INCREMENT = 1006), so the docs should possibly feature a
caveat to that effect.
It's not very intuitive the way it works now (with the fix), but it's
*correct*. We're not storing the original value anyway, if we wanted
that, we'd have to change on-disk representation?
If we do dump/load cycles with empty DBs, nothing will change. This
changeset includes an additional test case that proves that tables
with rows will create the same next_ID for AUTO_INCREMENT = ... across
dump/restore cycles.
Confirmed by support as likely solution for client's problem.
In the code that converts IN predicates to EXISTS predicates it is changing
the select list elements to constant 1. Example :
SELECT ... FROM ... WHERE a IN (SELECT c FROM ...)
is transformed to :
SELECT ... FROM ... WHERE EXISTS (SELECT 1 FROM ... HAVING a = c)
However there can be no FROM clause in the IN subquery and it may not be
a simple select : SELECT ... FROM ... WHERE a IN (SELECT f(..) AS
c UNION SELECT ...) This query is transformed to : SELECT ... FROM ...
WHERE EXISTS (SELECT 1 FROM (SELECT f(..) AS c UNION SELECT ...)
x HAVING a = c) In the above query c in the HAVING clause is made to be
an Item_null_helper (a subclass of Item_ref) pointing to the real
Item_field (which is not referenced anywhere else in the query anymore).
This is done because Item_ref_null_helper collects information whether
there are NULL values in the result. This is OK for directly executed
statements, because the Item_field pointed by the Item_null_helper is
already fixed when the transformation is done. But when executed as
a prepared statement all the Item instances are "un-fixed" before the
recompilation of the prepared statement. So when the Item_null_helper
gets fixed it discovers that the Item_field it points to is not fixed
and issues an error. The remedy is to keep the original select list
references when there are no tables in the FROM clause. So the above
becomes : SELECT ... FROM ... WHERE EXISTS (SELECT c FROM (SELECT f(..)
AS c UNION SELECT ...) x HAVING a = c) In this way c is referenced
directly in the select list as well as by reference in the HAVING
clause. So it gets correctly fixed even with prepared statements. And
since the Item_null_helper subclass of Item_ref_null_helper is not used
anywhere else it's taken out.
Backporting a changeset made for 5.0. Comments from there:
The fix refines the algorithm of generating DROPs for binlog.
Temp tables with common pseudo_thread_id are clustered into one query.
Consequently one replication event per pseudo_thread_id is generated.
The bug caused wrong result sets for union constructs of the form
(SELECT ... ORDER BY order_list1 [LIMIT n]) ORDER BY order_list2.
For such queries order lists were concatenated and limit clause was
completely neglected.
The bug caused a reported index corruption in the cases when
key_cache_block_size was not a multiple of myisam_block_size,
e.g. when key_cache_block_size=1536 while myisam_block_size=1024.
