The Bayou Architecture: Support for Data Sharing Among Mobile Users

One-line summary: From-the-ground-up design of a weakly consistent, highly available, replicated shared database using peer-to-peer anti-entropy, per-write merge and update functions, and selectable consistency/availability tradeoff at the clients.

Overview/Main Points

• High availability for reads and writes: use read-any/write-any weakly consistent replication.
• Eventual consistency with minimal assumptions about communication characteristics: Peer-peer anti-entropy to propagate updates.
• Detect and resolve update conflicts: dependency checks on each write, application-specific merge proc passed w/each write.
• Commit to stable value ASAP: primary server sets commit ordering.
• Disconnected operation while allowing groups to see their own updates: allow client to read "tentative" data with expectation that it will be committed with the same effect if possible.
• Client's view should be consistent with its own updates: range of selectable session guarantees:
  • Read your own writes
  • Monotic reads (successive reads reflect non-decreasing set of writes)
  • Writes follow reads: writes propagated after the reads on which they depend
  • Monotonic writes: writes propagated after writes that logically precede them
  • Age parameter allowed on reads
  • Choice of tentative or committed on reads
• Flexible control over placement and use of DB's: number and locations of DB's can vary over time, as can its primary server. Scales OK since anti-entropy is peer-to-peer.

Relevance

Flaws