List:Commits« Previous MessageNext Message »
From:Andrei Elkin Date:November 18 2010 2:51pm
Subject:bzr push into mysql-next-mr.crash-safe branch (andrei.elkin:3211 to 3212)
WL#5569 WL#5599
View as plain text  
 3212 Andrei Elkin	2010-11-18
      wl#5569 wl#5599
      Recovery related. Prototyping the worker RLI instantiation, to be elaborated on.

 3211 Andrei Elkin	2010-11-18
      wl#5569 MTS
      Extending the wl#5563 prototype gradually.
      This commit addresses:
      1. recovery interface (a new Worker rli plus rli->gaq and pseudo-code for checkpoint
         to update GAQ and the central RLI recovery table.
         Wrt rli, C and W execute do_apply_event(c_rli) where c_rli is the central
         instance. C executes update_pos(c_rli), but W update_pos(w_rli).
      - decreased processing time for rpl_parallel, serial.
     @ sql/
        Enhance Log_event::get_slave_worker_id() to classify events by set of parallelization properties;
         Presence of a property in an event forces some actions both on C and W side.
        en_queue etc are prepared to turn into circular_buffer_queue methods.
        Pseudo-coded numerious todo:s wrt to low-level-design implementation.
        Deployed changes due to Worker private rli.
        Annotated on Deferred Array for B,p,r property events.
        delete ev is moved from C to W which is fault-prone but it could not be kept
        any longer as a part of de_queue() that transits into cir_buf_queue class.
     @ sql/log_event.h
        removed `soiled' that was used to make delete ev run safely.
        Added Log_event methods identifying the parallelization properties, incl
         - contains_partition_info() to identify events containing 
           info to be processed by the partition hash func
         - starts,ends_group()
         - also updated the list of only_serial().
     @ sql/
        Only Coordinator can destroy Workers dynarray;
        Relay_log_info::get_current_worker() turned out to become more complicated, see comments;
        Reminder to migrate rli->future... into  ev-> future_event_relay_log_pos
        which would make Worker to find the value out the event's context;
        Prototyped //  w->flush_info() in stmt_done;
     @ sql/rpl_rli.h
        The worker RLI has `this_worker' pointing to the actual worker instance.
     @ sql/
        Annotated with fine details APH etc implementation.
     @ sql/rpl_rli_pdb.h
        Trasformed earlier queue struct into a family of classes.
        Recovery interface: last_group_done_index of Slave_worker to be filled in with an index
        of GAQ queue by W. To poll the value by C at checkpoint.
        Added CGEP to W context (sim to CGAP of C).
     @ sql/
        Simplified the Worker poll.
        Deployed worker rli initialization.
        Recovery: rli->gaq is instantiated by C at worker poll activization.
        Recovery: pseudo-code for checkpoint in next_event().
     @ sql/
        editted help lines for slave_max_pending_jobs.

=== modified file 'sql/'
--- a/sql/	2010-11-18 14:00:52 +0000
+++ b/sql/	2010-11-18 14:50:54 +0000
@@ -3524,7 +3524,13 @@ pthread_handler_t handle_slave_worker(vo
   w->tables_to_lock_count= 0;
   // fixme: experimenting to make Workers to run ev->update_pos(w->w_rli)
-  w->w_rli= Rpl_info_factory::create_rli(RLI_REPOSITORY_FILE, FALSE);
+  // fixme: a real hack! part of Rpl_info_factory::create_rli(RLI_REPOSITORY_FILE, FALSE);
+  w->w_rli= new Relay_log_info(FALSE,
+                               &key_relay_log_info_run_lock,
+                               &key_relay_log_info_data_lock,
+                               &key_relay_log_info_data_cond,
+                               &key_relay_log_info_start_cond,
+                               &key_relay_log_info_stop_cond);
   w->w_rli->info_thd= thd;
   w->w_rli->workers= rli->workers; // shallow copying is sufficient
   w->w_rli->this_worker= w;

Attachment: [text/bzr-bundle] bzr/
bzr push into mysql-next-mr.crash-safe branch (andrei.elkin:3211 to 3212)WL#5569 WL#5599Andrei Elkin19 Nov