LoginSign Up
← Back to all sparks
V

VerneMQ

COMMS
Velocity0.0

Distributed MQTT message broker

vernemq.com ↗

A slow, careful MQTT broker that spent two years landing one breaking storage change.

mqttclusteringerlangrelease-candidatesbreaking-changesself-hosted
◆Current state
VerneMQ is in low-cadence maintenance on the 2.1.x line, with releases spaced months apart and a 2.1.3 release candidate open since April. The work is concentrated in the clustering metadata store (vmq_swc), the MQTT session state machine, and WebSocket listener behaviour — the parts operators actually file bugs against. Commercial binaries remain EULA-gated, so the open-source repo is the product surface and the packages are the business.
◆Where it's heading
The centre of gravity is correctness in clustered deployments, not new protocol capability. Recent releases keep tightening the same three areas — connection accounting under churn, proxy-header handling behind load balancers, and metadata sync for joining nodes — which reads as a broker being hardened for larger clusters rather than broadened. The RC-heavy release pattern (three RCs before 2.1.0, one open for 2.1.3) suggests a small team leaning on community testing before committing.
◆Prediction
2.1.3 ships as a fix-consolidation release absorbing the remaining open PRs, with the retain-expiry default change carried through. Nothing in these entries points to a new protocol or storage direction after that.

◆Recent moves

  1. 3mo ago

    2.1.3 RC1: socket-close accounting fix, retain-expiry check off by default

    An open release candidate that continues the cluster-hardening arc: a fix for connection counting when clients reconnect rapidly, ETS mirroring of event subscriptions in the SWC store for reload, and retain-expiry checking disabled by default. The notes explicitly say more open PRs will land before the final 2.1.3.

    View source ↗
  2. 4mo ago

    2.1.2 fixes WebSocket connection counts and X-Forwarded-For handling

    The clearest example of the operator-facing focus: active connection metrics now sum all WebSocket connections, X-Forwarded-For is read only from the actual proxy hop, and the session state machine sends SUBACKs on every error path. Routing-table memory management also moved off gen_server2.

    View source ↗
  3. 10mo ago

    2.1.1 fixes syslog boot, drops default local listeners

    A cleanup release that fixes a syslog boot failure, adds a syslog level setting, repairs an HTTPS config type regression introduced in 2.1.0, and stops auto-adding local MQTT and WebSocket listeners. The listener change is the one operators will notice on upgrade.

    View source ↗
  4. 1y ago

    2.1.0 lands a breaking on-disk metadata format

    ⚡ SPARK

    This is the release the whole 2.1.x line was built around: a breaking on-disk format change means a 2.1.0 node cannot read metadata state written by any earlier release. Everything since has been fixes settling on top of that foundation.

    View source ↗
  5. 1y ago

    2.1.0-rc3 makes inter-node connect options configurable

    The last RC before 2.1.0 fixes reconnect-after-failed-connect and keeps the patch directory first on the code load path, but the substantive change is making outgoing connect options configurable for inter-node connections — another knob for operators running across constrained networks.

    View source ↗
  6. 1y ago

    2.1.0-rc2: single init_sync flag fix

    A one-commit release candidate correcting the init_sync flag state when booting from disk. Visible only as a bug that no longer happens during the 2.1.0 metadata migration.

    View source ↗