Bug #31609 Not all RBR slave errors reported as errors

Description:
Due to idempotency requirements, RBR does not report error
on some duplicate key/key does not exists failures.

How to repeat:
Code inspection.

Suggested fix:
RBR needs to, per default, report these as errors and have an option
to turn these checks off.

See also BUG#31552.

I had a talk with Tomas and we think the simplest solution is a
dynamic system variable:

  slave-rbr-mode=[idempotent | strict]  

with default 'strict' setting.

Benefits:

+ One can use this also for other storage engines when one wants idempotency.
+ Easy to test
+ Having a "mode" makes it possible for us to set a family of internal 
  options at once
+ It is east to use
+ Adding new modes is easy

Drawback:

- Telco/cluster customers need to set the option when upgrading
  (But since it is dynamic, it should be easy to do).

For the future (later versions), Tomas and I think that cluster might
want to add a flag for transaction that they should be run
idempotently.  Not all epochs need to be executed idempotently for
cluster.

These two things need to be fixed before closing this bug.

Documentation
=============
We need to clarify in manual the conditions when idempotency of log is
attainable and can be used:

1. Primary keys are needed to get idempotency.

2. Primary keys can't be changed

Example 1:
----------
Consider a binlog that changes the primary key of a row in this way:
(2->3, 1->2) and that the initial data is a record 1.

Then running binlog once gives one record 2.  Running it twice gives
another record 3.

Example 2:
----------
Consider a binlog that changes the primary key of a row in this way:
(2->3, 1->2) and that the initial data is (1,2).

Then running binlog once gives (2,3).  Running it twice gives 
(3,3) = (3), i.e only one record is left.

Cluster 
=======
We need to have verified that these conditions are fulfilled in NDB,
since NDB will have the idempotency flag enabled.

Tomas has confirmed that primary key change is always logged as a
DELETE plus an INSERT.

Taking the 2nd example above, we then get:

First apply of binlog
---------------------
Assume two initial records: (1,2).

Assume two updates: (2->3, 1->2).

Then in binlog this becomes:

1. Delete 2.
2. Insert 3.
3. Delete 1.
4. Insert 2.

The result is then two records: (2,3).

Second apply of binlog
----------------------
Running this binlog on (2,3) gives:

1. Delete 2.  Ok.
2. Insert 3.  "Error: Duplicate key".  Ignored.
3. Delete 1.  "Error: Can't find record".  Ignored.
4. Insert 2.  Ok.

Result (2,3).

So all should work fine.  This should always work in the context we
work in, but anyone having a formal proof that this will always work,
is encouraged to provide such a proof. :)

The patch is within changeset for Bug #31552.

Pushed into 5.1.23-rc

Pushed into 6.0.4-alpha

D1/I2

Pushed into 5.1.24-rc

Pushed into 6.0.5-alpha

Documented bugfix in 5.1.24/6.0.5 changelogs; documented new slave_exec_mode system variable; see http://lists.mysql.com/commits/41771 for details.

Also documented for 5.1.23-ndb-6.2.14.

Also documented for 5.1.23-ndb-6.3.11.