Repository navigation
Yes, Numax is still alive, and so am I!
It’s been almost two months since the last release. I haven’t disappeared: I’ve been working on this release on my own, and it turned into one of the most demanding versions I’ve tackled so far. It took a lot of iterations across membership, replication, recovery, testing, and documentation. But it was worth it: I’m genuinely happy with how these pieces now fit together.
v0.1.6 takes peer discovery beyond finding other nodes. It builds a path from discovering a candidate, to tracking its membership and health, to delivering and recovering CRDT state.
Peer discovery, membership, and gossip
- Bring discovered and manually registered peers into a bounded candidate registry, while keeping discovery separate from trust: candidates still have to pass the normal connection, identity, and protocol checks.
- Add SWIM/Lifeguard-style membership with versioned node epochs and incarnations, direct and indirect probes, suspicion, self-refutation, and
Alive,Suspect,Dead, andLeftstates. - Keep membership control exchanges separate from the data-delivery queue, so congestion in gossip delivery does not directly block membership control traffic.
- Add configurable bounded K-fanout gossip, rotating neighbor sets, and recovery contacts to help sparse topologies and partitioned nodes reconnect.
- Adapt gossip pacing using queue load, write service time, and correlated round-trip samples.
For the implementation details, see the Peer Discovery, Membership, and Gossip design and the Gossip Protocol guide. The Discovery Provider Contract describes how discovery providers contribute candidates without granting membership or authorization.
Durable replication and recovery
- Extend local operation handling across all six CRDT families, making a clear distinction between accepting an operation locally, confirming its durability in storage, and replicating it to peers.
- Add per-origin sequence tracking, replay deduplication, and sparse receipt tracking so replicas can identify missing operation ranges and request recovery.
- Bound recovery tracking and state transfer: the implementation places explicit limits on tracked origins, gaps, sparse positions, baseline size, and record count.
- Add versioned full-state baselines for cases where the required operation history is no longer available. A baseline must represent a complete, dominating state cut; partial additive-counter state is not merged as if it were complete.
- Bound delivery by bytes as well as messages. Operations are persisted before they enter the volatile delivery queue, so rejecting a send attempt does not discard the only recoverable copy.
- Harden shutdown and storage behavior: a membership leave is published only after its durable state is confirmed, and a timed-out shutdown can be retried while the actor completes that work.
The Peer Discovery design covers the durability, sequencing, and recovery contracts. The Management API guide and Observability guide explain how to manage and inspect peers and runtime behavior.
Compatibility, tests, and documentation
- Evolve the wire protocol and serialization-format negotiation, with incompatible protocol versions rejected during the handshake.
- Add deterministic peer-I/O simulation and replay tests, CRDT convergence and persistence coverage, and restart/recovery scenarios across all six families.
- Add packet-loss, partition-recovery, and membership load/soak test harnesses. These improve our ability to investigate behavior under stress; they are not a claim that the separate 100-node scalability goals have been met.
- Rewrite and expand the discovery, membership, recovery, SDK, CLI, and API documentation, with new diagrams and updated examples.
- Update workspace dependencies, including Wasmtime/WASI to 50.0.1.
The sync manager test guide documents how to run the replication, restart, recovery, and load scenarios.
Background reading
The project’s Foundations documentation links these references:
These are background references already linked by the project documentation;
Thank you
A huge thank you to @AndyFerns , who contributed two more improvements: Node.js ABI smoke tests for the C guest and the C++ guest. Andy is, as always, Numax’s number-one contributor, and I’m grateful for the care and consistency he brings to the project.
And a warm welcome to @hippi345 . Their first contribution adds focused nx-sdk tests for dynamic reads, capability decoding, module IDs, and event return-code handling (PR #216). Thank you for jumping in and helping make the SDK more reliable!
I may have built this release mostly on my own, but every contribution and every person taking the time to try Numax makes the project better.
Thank you, and yes, I’m still here. We’re moving forward.