Repository navigation
v1.4.0 — multi-instance cluster layer
v1.4.0 — 2026-11-10
Multi-instance cluster layer. One binary, three modes; single-node
deployments keep byte-identical behavior with zero config changes.
Server (etunneld)
- Cluster modes —
ETUNNEL_CLUSTER=single(default) |mesh|sticky.mesh: memberlist gossip ownership + cross-peer forwarding; works
behind any L4 load balancer (round-robin fine).sticky: Redis routing table (et:route:<sub>, 15s TTL lease, 5s
heartbeat) for L7 load balancers.
- Postgres account store —
ETUNNEL_PG_DSN; in-memory hot path with
write-through + 5s poll; auto-migrated schema. Without it: JSON store
as before. - Redis global counters —
ETUNNEL_REDIS_URL; exact global per-IP /
per-account limits. Without it: per-instance limits as before. - Graceful drain —
/healthzflips 503 on SIGTERM, sessions drain,
SDKs re-land on any peer with the same URL (resume tokens verify via
sharedSESSION_SECRET). - Fix: brand-asset file server was claiming the global
/assets/
subtree — a tunneled WebUI's own/assets/*.js|cssreturned Go's
default 404 while the homepage proxied fine. Now host-scoped to the
base domain only.
Deployment artifacts
- Docker image (distroless, ~17MB) +
deploy/docker-compose.yml(3-node
mesh + PG + Redis, verified end-to-end). - Helm chart
deploy/helm/etunnel(StatefulSet + headless DNS, verified
on k3s: 3-pod mesh,K3S-CLUSTER-OKx3). - Bare-metal systemd template
deploy/systemd/etunneld-cluster@.service. - Full guide with mermaid diagrams:
deploy/CLUSTER.md.
CI
- Every build runs the live PG/Redis cluster-store tests against service
containers (postgres 16 + redis 7) and fails on SKIP, so store
coverage can never silently regress.
SDKs
- Go: reconnect-policy unit tests (grow/jitter/isPermanent).
- No API changes — all 20 SDKs at 1.4.0.
Upgrade notes
- Single-instance users: drop-in binary replacement; no new env vars
required. - To enable the cluster layer see
deploy/CLUSTER.md— the only required
secret in any multi-instance mode isSESSION_SECRET.