0025 (ADR-003): etcd is the mesh registry, not only quorum #30

New issue

Open

opened 2026-09-30 10:10:42 -03:00 by marcos-mendez · 0 comments

marcos-mendez commented

2026-09-30 10:10:42 -03:00

(Migrated from github.com)

Decision (maintainer, 2026-09-30). etcd is the mesh registry: every appliance announces what it offers, where it is, and its site (availability zone). Control plane only, never application data; a few writes a minute tolerate 30 ms RTT. Quorum needs an odd number of voters; a third vote is mandatory for automatic failover. Site labels enforce that two replicas of the same data never share a site. No Corosync across a WAN.

Handbook: docs/decisions/0025-etcd-is-the-mesh-registry.md (handbook#26)

Amends: 0020 (etcd was one appliance's election lease; now the registry of the whole mesh; membership moves from Keel Cloud to etcd; the data-less voter becomes the external arbiter), brief section 2 (a registry is not orchestration).

First tasks

Specify the registry key layout: node, site, offered services, endpoints, health summary (0040).
etcd overlay in common, stopped by default (0036), with TLS peer and client certificates over the mesh.
Three-node gate on the build host bridge (0020's hard part 6).

**Decision (maintainer, 2026-09-30).** etcd is the mesh registry: every appliance announces what it offers, where it is, and its site (availability zone). Control plane only, never application data; a few writes a minute tolerate 30 ms RTT. Quorum needs an odd number of voters; a third vote is mandatory for automatic failover. Site labels enforce that two replicas of the same data never share a site. No Corosync across a WAN. **Handbook:** `docs/decisions/0025-etcd-is-the-mesh-registry.md` (handbook#26) **Amends:** 0020 (etcd was one appliance's election lease; now the registry of the whole mesh; membership moves from Keel Cloud to etcd; the data-less voter becomes the external arbiter), brief section 2 (a registry is not orchestration). **First tasks** - [ ] Specify the registry key layout: node, site, offered services, endpoints, health summary (0040). - [ ] etcd overlay in common, stopped by default (0036), with TLS peer and client certificates over the mesh. - [ ] Three-node gate on the build host bridge (0020's hard part 6).