Skip to content
TensorDB

Replication

Replication Model

TensorDB’s append-only architecture simplifies replication:

  • No conflict resolution needed: Facts are immutable and identified by (key, commit_ts)
  • Simple synchronization: Replicate the WAL log entries
  • Temporal consistency: commit_ts ordering is preserved across replicas

Planned Features

Synchronous Replication

Writes are committed only after replication to a majority of nodes:

Client → Leader → Replicate to Followers → Majority ACK → Commit → Response

Asynchronous Replication

For lower-latency writes with eventual consistency:

Client → Leader → Commit → Response (async replicate to followers)

Read Replicas

Dedicated read replicas for scaling read-heavy workloads:

  • Receive streamed WAL updates from the leader
  • Serve read queries with bounded staleness
  • No write capability

SSTable-Level Replication

Because SSTables are immutable, they can be replicated at the file level:

  • New SSTables are transmitted to followers after compaction
  • Followers apply the same manifest updates
  • No need to replay individual writes