March 02, 2010

• Attendees: Dies, Anissa, Jennifer, Ken, Nazrul, Bill, Byron, Paul, ???
• Today's topic was cluster support for GF V3.1.
  During the meeting we referred to the following document (which will
  continuously be updated):
  V3.1Clustering
  For GF V3.1 we're planning cluster support without a Node Agent (NA).
  At the later stage we could consider adding the Node Agent again.
  Without a Node Agent, it will not be possible to start remote instances.
  There was an idea to use ssh to replace that functionality from the NA,
  but this will probably not be done in GFv3.1.
  One of the main reasons to not including a NA is that people complained
  about it. Not only because of issues they had and because it turned out
  not to be as light-weight as it was initially intended to be, but users
  say they prefer to register their instances with their OS's services
  functionality (Windows service, etc.). This would give them behaviour
  they are familiar with, and provide extra functionality, such as setting
  the maximum number of times an instance should be restarted in case of a
  failure.
  Most of the discussion was about the way remote instances should
  synchronise their files (and state) with the DAS.
  In GF v2.x, the deploy process runs on the DAS and the remote instances
  download the post-deployment (including generated) artefacts.
  In GF v3.1, we want the remote instances to do the operations
  themselves, so there would be less issues with file synchronization.
  A potential issue with this approach is when deploying of applications
  that require a lot of resources for deployment. For instance, JSPs can
  be precompiled during deployment, and CMP applications access the DB for
  schema generation. This will need to be investigated further
  (performance comparison tests, etc.).
  We can also consider optimizations like doing the deploy per DAS instead
  of per
  instance when the cluster's instances are on the same machine, and share
  the files instead of each instance having its own copy.
  Someone pointed out that this could create issues when redeploying an
  application that is already running on a cluster instance. (Classes that
  are in use would be overwritten).
  For the file synchronization we would use the files' timestamps. For
  complete subdirectories, only the subdirectories' filestamps would be
  compared.
  Discussion on this topic will be continued next week.

  ---