evento-accord 2.0.0-alpha.38

Accord consensus protocol (CEP-15) executor for evento — a leaderless, strictly-serializable, highly-available replicated event store
docs.rs failed to build evento-accord-2.0.0-alpha.38
Please check the build logs for more information.
See Builds for ideas on how to fix a failed build, or Metadata for how to configure docs.rs builds.
If you believe this is docs.rs' fault, open an issue.
Visit the last successful build: evento-accord-2.0.0-alpha.37

evento-accord

A new evento Executor backed by the Accord consensus protocol (Apache Cassandra CEP-15) — a leaderless, strictly-serializable, highly-available replicated event store. Events are replicated across N = 2f + 1 nodes with a single global serial order and no elected leader, tolerating f failures.

Status: alpha. A well-tested reference implementation, not yet production-hardened (pending independent review + real-cluster soak). See DESIGN.md for the architecture and milestone plan, and OPERATIONS.md for the deployment/ops runbook.

Part of the evento event sourcing toolkit; see examples/bank-axum-accord for a runnable 1- or 3-node demo (make accord / make accord.cluster).

What works

  • AccordExecutor — a drop-in evento_core::Executor. Writes are coordinated through the cluster (replicated, strictly serializable); reads, subscriptions, and snapshots are served from a local evento backend (Fjall/Sql) kept current by applying committed transactions.
  • Multi-shard, atomic transactions — keys partition across disjoint shards, each with its own replica set and quorum. A write spanning aggregates commits all-or-nothing via a Read → Apply phase split (the basis for cross-aggregate transactions like a money transfer).
  • Fault tolerance — quorum-based progress survives f failures per shard; a live node recovers a transaction whose coordinator died (ballot-fenced).
  • Real transport — length-delimited framed TCP behind the MessageSink trait, plus an in-memory transport for deterministic tests.

Layering

AccordExecutor (evento_core::Executor)
        │
   Node (coordinator + replica) ── Replica (consensus state machine)
        │
   api traits: MessageSink · Topology · Clock · Journal · DataStore

The protocol depends only on the api traits; each has an in-memory implementation for the test harness and a production one (framed TCP, static membership, a Fjall/SQL-backed store).

Also supported

  • Elastic membership — epoch-versioned topology; nodes join (bootstrapping their key range) and leave; epoch changes are decided by Paxos over the current members, so they survive a coordinator crash and keep working after the founders have left; keys re-shard onto new owners.
  • Multi-shard reads — a read for a key a node doesn't own is routed to an owner.

Not yet

Garbage-collecting a moved key range from its old owners.