Comment 1 for bug 1190580

It seems (have to confirm still) that Facebook solved this in day zero code import [1] by adding the following to ha_commit_trans(). For some reason we didn't fix it then.

    else if (is_real_trans && !trans->no_2pc && (rw_ha_count == 1))
    {
      /* When innodb_fake_changes is enabled for a transaction and that transaction
      changes no InnoDB rows then the InnoDB handlerton is marked as read-only and
      does not participate in 2 phase commit above. So InnoDB won't be able to
      raise an error during commit and the transaction would be written to the
      binlog. This code prevents that. */

      Ha_trx_info *ha;
      for (ha= ha_info_orig; ha; ha= ha->next())
      {
        handlerton *ht= ha->ht();
        if (ht->is_fake_change && ht->is_fake_change(ht, thd))
        {
          thd_reset_diagnostics(thd); /* avoid debug assertion */
          ha_rollback_trans(thd, all);
          my_error(ER_ERROR_DURING_COMMIT, MYF(0), HA_ERR_WRONG_COMMAND);
          error= 1;
          goto end;
        }
      }
    }

[1]

facebook-5.1$ bzr log -r 3763
------------------------------------------------------------
revno: 3763
committer: Mark Callaghan <email address hidden>
branch nick: 51fb
timestamp: Tue 2012-01-03 10:03:20 -0800
message:
  Port innodb_fake_changes.patch from Percona

  This adds a session parameter, innodb_fake_changes, that prevents
  rows from being locked and modified when set. It can be used to prefetch
  transactions. When set COMMIT and auto-commit statements will return an
  error. Non auto-commit statements are not foced to fail with this set.
  Binlog events are not written for a session when this is set for changes
  to InnoDB tables. All bets are off if a session mixes changes to MyISAM
  and InnoDB.

  The patch is at
  http://bazaar.launchpad.net/~percona-dev/percona-server/5.1/view/head:/patches/innodb_fake_changes.patch