Overview

Each of these is a short architecture decision record: the choice made, the alternative(s) seriously considered, and why the tradeoff went the way it did. They're written from the actual design specs and implementation plans in the repo, not reconstructed after the fact.

• Raft over Paxos — why consensus is Raft, not Paxos or a Paxos variant.
• LSM over B-tree — why the storage engine is log-structured, and why the Raft log store is a separate component rather than another LSM consumer.
• Atomic flush via temp+rename — how a memtable flush or compaction avoids ever exposing a partially-written SSTable to a reader.
• Seqno recovery — why sequence numbers must recover from SSTables, not just the WAL, on restart.