04_concurrent_writes

Concurrent Writes

Concurrent writes happen when two clients write to a database at the same time, unaware of each other’s write. This can cause inconsistencies in the replicas.

[Write 1]   [Write 2]   [Write 3]
   \            |            /
    \           |           /
     \          |          /
      \         |         /
       \        |        /
       +------------------+
       |     Database     |
       +------------------+

Detecting Concurrent Writes

• Both multi-leader and leaderless replication systems must detect and resolve concurrent writes.
• Events may arrive in different orders at different nodes; it’s important for nodes to reach a consistent state eventually.

Last Write Wins

• A common approach for concurrent writes is the “last write wins” strategy.
• Writes are given an arbitrary order, and the last write is chosen as the correct one.
• Easy to implement, but not durable; some writes are discarded.
• Timestamps for ordering writes can be problematic due to clock skew.

Immutable Keys

• An alternative is to allow each key to be written only once, making it immutable.
• No writes are lost, but may not suit all use cases.

Version Numbers

• Use version numbers to track a key’s state.
• Clients pass the last version number they’ve seen to the server, which assigns a new version number and returns it.
• The server determines which data the client had access to at the time of writing.
• If the client writes without relying on existing data, the server can return both versions of the data on subsequent reads.
• The application can handle these two values, either merging them automatically or prompting the user to do so.

Version Vectors

• Version vectors track the state of data for multiple replicas.
• A version vector includes a version number for each replica and key.
• Each replica keeps track of its own version number and those of other replicas to decide which values to use.