Comment 19 for bug 1327763

commit d4fa27b020df425a6421ae12a18de9306cf5c90d
Merge: 9d30f12 f06d17f
Author: Krunal Bauskar <email address hidden>
Date: Thu Feb 18 06:35:17 2016 +0530

    Merge pull request #90 from kbauskar/5.6-pxc-497

    - PXC#497: Warning: "BF applier failed to open_and_lock_tables"

commit f06d17fd75c2aee6dbdbecd81e0795be1017284a
Author: Krunal Bauskar <email address hidden>
Date: Mon Feb 15 22:15:47 2016 +0530

    - PXC#497: Warning: "BF applier failed to open_and_lock_tables"

      Issue:
      -----

      Let me explain it with an example. Would be simpler.

      Let's assume there are 2 pxc nodes and each node and following
      workload is being

      node-1: executes already prepare stmt and about to commit
      (update i > 2 and i < 5)

      node-2: executes some stmt that would conflict with node-1 (insert (3))
      inserting in gap.

      As per the protocol local trx is selected for rollback and so
      when node-1 get node-2 trx which is applier trx node-1 local trx that
      is prepare stmt execution is rolled back and re-tried using replay trx.

      Note: retried using replay trx effectively means only write-set
      is applied this makes it bit faster.
      Application of write-set uses the apply-log-event path which
      open and locks table so there is no need to keep track of
      table versioning change which is normally done (in normal path)
      by installing a reprepare observer.

      Solution:
      ---------

      * Avoid installing observer when prepare trx is being replayed
        as trx will open and lock the tables.