WL#2735: Refactor Replication

Affects: Server-7.1   —   Status: On-Hold   —   Priority: Medium

This WL contains various ideas for refactoring replication.
We'll probably do them piece by piece and not all of them
might be done.

SOME IDEAS (More below)

1. The new file master.cc should only contain code for the master.
   All slave code should move to slave.cc

2. Move log_loaded_block into sql_load.cc
   Guilhem wrote:
   > [G] This log_loaded_block() was in sql_repl.cc but it had nothing to do
   > [G] there. Maybe in a further reorg we could move it to where it belongs:
   > [G] binlogging of LOAD DATA on the master side, so maybe sql_load.cc

Should be very easy to fix after WL#1697 is pushed.
Copyright (c) 2001-2007 by MySQL AB. All rights reserved.



   - sql_slave.* (was sql_repl.*)
     Interface from MySQL server to slave functionality
     (roughly one call per SQL statement)

   - sql_master.* (was sql_repl.*)
     Interface from MySQL server to master functionality
     (roughly one call per SQL statement)

   - binlog.* (to be created, move code from sql_base,sql_class
     and away from thd object)
     Interface from MySQL server to binlogging.  The idea is that
     the server does not know how things are logged and does not
     know if there are multiple binlogs below this interface.
     The future handler binlog interface will be below this interface
     (there could, in far future, be multiple handlers storing the
     binlog here).  I need to look at this in some more detail
     to figure our if there are any obstacles to do this.

   - rpl_mi.* (already created)
     Master info functionality (will, in far future, no longer be
     directly accessible from MySQL server)

   - rpl_rli.* (already created)
     Relay log info functionality (will, in far future, no longer be
     directly accessible from MySQL server)

     repl_failsafe.* to be moved into sql_slave.* and sql_master.* (the
     stuff that actually are interface functions).  Some code to be



I think log_event.h/cc needs some improvements.  Frequently we get
questions about the binlog format, and it usually takes hours to
understand things that should only take seconds to look up if the code
was simpler.  I think it

- takes hrs to try to understand what the binary encoding of an
  event is,
- is hard to test the binary format of the events (would be super with
  readers and writers of events),
- functionality of events, e.g. header_len is spread into
  multiple classes making it hard to read,
- versioning info is in multiple places making it hard to
  understand what each version needs and makes it hard to deprecate
  code and make sure the code when the master and slave has different
  versions, and
- both client and server code is mixed in same file making
  the code filled with #ifdef making it harder to read.

I'm thinking of changing it in some way like this:

1. Separate client-server code (example: Query_log_event):
   - Let all event classes (e.g. Log_event_query) only contain things
     that the *client* needs.
   - Create a server class for each event (e.g. Server_event_query)
     that inherits from the client class.
   Negative: the server will then have print functions compiled into it.
   Positive: the client object can be used as a binlog API.

   Mats explands this into: Move common items into a common base
   class: let the client inherit from the common parts and add the print
   member function, which will be based on the common parts. Let the
   server class inherit from the common base class and add whatever it
   needs. It is common to use a POD as the root base class, and maybe add
   a layer of common functionality in a subclass.  That way, there is an
   easy and efficient way to process the raw data, if necessary (for
   serialization purposes, for example).

2. Move headlen info into the events.  The event itself should know its
   header length dependent on the binlog version.  (Idea is that one
   should only need to look at the event class to see the binary

   Mats: Agree. That will also allow it to be used with template
   function, if the need arises. In general, it's a good idea to put all
   information associated with an event *in* the event class. That allows
   us to build generic methods for handling events that are *far* more
   efficient and easy to maintain, and also with good locality, which
   makes cached architectures work more efficiently.

3. Institute a policy that binlog version X and event type Y might
   correpond to a different class than binlog version X' with
   event type Y.  This so that we can move into future versions
   in two ways:
   - Adding new event types for new events.
   - If the number of event types exceed 256, we can increase the
     binlog version to get a whole new range of types.
   (This is thoughts for the far future though.)

4. Move some of the code into:
   log_event_rbr.h/cc   : RBR events
   log_event_load.h/cc  : LOAD DATA INFILE (3 different implementations!)
   so that events that belong together are in same file, but that not
   all events are in the same huge file.

5. Factor out decode/encode (Mats idea)

   Factor out the parts that handle the decoding/encoding of log
   events into a separate class. That way the events does not have to
   carry around a lot of representational information, making them faster
   to build and destroy. It would also be easy to look up encodings,
   since they will be written separately and probably in a condensed
   format (just to give an idea, something along the lines):

   class Query_log_event_code : public Log_event_code {
     void encode(Query_log_event const& ev, Output& out) {
       Log_event_code::encode(ev, out);

   This would also make it easy to switch encoding, or keep two
   different encodings for the same events (essentially creating an
   Abstract Factory and implementing different concrete factories for
   each encoding).

6. (Mats idea) Develop a little language that can take a format
   description and generate an encoding and decoding function. Changing
   or adding fields to the log events is then done in an instant. With a
   generic tool, we can do the same for other parts of the server. That
   way, data format documentation and coding is one, and the code can be
   organized radically different. If the format description format is
   (X)HTML, or something similar, generation of documentation is

   Lars thinks this is cool but might be over-kill.

7. Problem: query log event contains a log of state information
   and this makes the events fat, i.e. many bytes.
   Refactor into:
   1) master start by writing its global state in log (possibly in format desc
   2) if master session != master global, add info to event, 
   3) if global change, then this must be propagated to all slaves (same as item
   4) slave needs to remember the global state of the master (for the log it is