repair it
Multi-table delete that is optimized with QUICK_RANGE reports table
corruption.
DELETE statement must not use KEYREAD optimization, and sets
table->no_keyread to 1. This was ignored in QUICK_RANGE optimization.
With this fix QUICK_RANGE optimization honors table->no_keyread
value and does not enable KEYREAD when it is requested.
When only one row was present, the subtraction of nearly the same number
resulted in catastropic cancellation, introducing an error in the
VARIANCE calculation near 1e-15. That was sqrt()ed to get STDDEV, the
error was escallated to near 1e-8.
The simple fix of testing for a row count of 1 and forcing that to yield
0.0 is insufficient, as two rows of the same value should also have a
variance of 0.0, yet the error would be about the same.
So, this patch changes the formula that computes the VARIANCE to be one
that is not subject to catastrophic cancellation.
In addition, it now uses only (faster-than-decimal) floating point numbers
to calculate, and renders that to other types on demand.
from log):
When row-based logging is used, the CREATE-SELECT is written as two
parts: as a CREATE TABLE statement and as the rows for the table. For
both transactional and non-transactional tables, the CREATE TABLE
statement was written to the transaction cache, as were the rows, and
on statement end, the entire transaction cache was written to the binary
log if the table was non-transactional. For transactional tables, the
events were kept in the transaction cache until end of transaction (or
statement that were not part of a transaction).
For the case when AUTOCOMMIT=0 and we are creating a transactional table
using a create select, we would then keep the CREATE TABLE statement and
the rows for the CREATE-SELECT, while executing the following statements.
On a rollback, the transaction cache would then be cleared, which would
also remove the CREATE TABLE statement. Hence no table would be created
on the slave, while there is an empty table on the master.
This relates to BUG#22865 where the table being created exists on the
master, but not on the slave during insertion of rows into the newly
created table. This occurs since the CREATE TABLE statement were still
in the transaction cache until the statement finished executing, and
possibly longer if the table was transactional.
This patch changes the behaviour of the CREATE-SELECT statement by
adding an implicit commit at the end of the statement when creating
non-temporary tables. Hence, non-temporary tables will be written to the
binary log on completion, and in the even of AUTOCOMMIT=0, a new
transaction will be started. Temporary tables do not commit an ongoing
transaction: neither as a pre- not a post-commit.
The events for both transactional and non-transactional tables are
saved in the transaction cache, and written to the binary log at end
of the statement.
An update that used a join of a table to itself and modified the
table on one side of the join reported the table as crashed or
updated wrong rows.
Fixed by creating temporary table for self-joined multi update statement.
