Event Notification

Event Notification

Motivation

• Users want to be updated ASAP (e.g. security patches)
• Current software systems
  • Update is done by connection to a main/secondary server
  • Update is requested by the local machine at predefined times or by user request
  • Polling is inefficient
    • Too many attempts for little gain (Do I have new mail? Do I have new mail? Do I have new mail?)
    • Increases load on central servers
    • Saving bandwidth by longer polling cycles causes the updates to take longer to arrive
• Push-based notification incentives
  • Most of users are now connected 24-7 using high-bandwidth connections
  • Distribute notifications using a P2P network
    • Minimum load on central servers
    • Events are stored in more than one location (redundancy)
    • Low latency in Update/Errata information delivered to user
  • No need to retransmit the same event again (ideally)

Requirements

• Scalable to millions of users
• Resilient (in case of a massive peers failure)
• Efficient
• Can be applied on top of "standard" P2P overlay network
• Channel-based pub/sub system
  • Channel hierarchy
  • Dynamic (channels can be created with ease)
  • Integrated with a persistent storage design/implementation
    • Stores previous notifications for offline/disconnected users
    • Enables user to dig up older versions (of single packages and entire distributions)
  • Permissions mechanism for publication in a channel

Directions

State-of-the-art

• Scribe
• Bayeux
• Gryphon
• Siena

Open Issues / Decisions

• Decision - We chose "fixed" Channel-based over Content-based
  • The topology of the peers in a channel model is more stable, and requires less overhead to maintain
  • Channel model is less flexible than the content-based model
  • The channel model has simpler and more efficient implementations
• Decision - No logging of events
  • It is sufficient for a user to get the latest version of a distribution/package
  • Users who are interested in historic versions will "dig" them up using persistent storage
  • Under this assumption, current event notification systems (see above) are sufficient