Repository navigation
fix: etcd's health on its metrics listener, and no lone member kept - #39
Merged
Merged
Conversation
keel now runs etcd over the mesh with TLS and client certificates on 2379, so Monit's plain probe of /health there failed every cycle and restarted etcd forever. The check asks /health on the plain metrics listener keel renders on http://[::1]:2381. etcd-server's postinst starts etcd, which leaves a member of a cluster of one in /var/lib/etcd/default. The overlay's first installation now removes it when it was not there before (preinst notes it), so the node keeps no lone cluster before keel joins it to the mesh's. Data that was there before is never touched; keel refuses to start etcd over a member it did not start.
marcos-mendez
pushed a commit
to Keel-Linux/keel
that referenced
this pull request
Oct 4, 2026
…a CRL From the security review of keel#78: - only the node that holds the mesh's root CA forms the cluster, once (keel.mesh.etcdca): it signs a formation record (cluster token, members, root fingerprint) with the root and reserves the formation under its lock; a member takes a cluster, from a join's answer or a cluster message, only with that record, and a member in a cluster takes no other. Two concurrent invites can no longer start two clusters; another inviter of the third member admits it and names the holder; keel mesh etcd form runs only on the holder, and on a mesh with no CA makes the root where the operator runs it (a trust root); - leaves last 30 days and are renewed by their member; intermediates a year, renewed by the holder, which signs them with the root (an inviter relays the request; only when the holder is unreachable does it sign one level deeper, re-anchored at the first renewal); - every intermediate is name constrained to the mesh's /64 and ::1; - the holder keeps a CRL signed by the root, written to etcd's --peer-crl-file and --client-crl-file and carried by grants and rosters; keel mesh remove has the holder revoke the removed node's intermediates, and so its leaves and what it issued; - etcd serves /health on a plain metrics listener on [::1]:2381 for Monit (Keel-Linux/common#39); - apply refuses a member in /var/lib/etcd/default that keel never started, and the join or form that starts etcd removes it first.
trixie-backports replaced golang-1.26-go with golang-1.27-go, so the step 5 packages pinned in the web job, which build with 1.26, could no longer have their build dependencies installed (mk-build-deps exited 1 for libcoraza). The job takes 1.26.8-1~bpo13+1 from snapshot.debian.org as of 2026-10-01, pinned.
marcos-mendez
force-pushed
the
fix/etcd-health-and-stray-member
branch
from
October 4, 2026 12:11
2d8d0e8 to
6c1a787
Compare
keel removes a member directory only when it can prove it is the lone member etcd-server's own start made at the overlay's installation: the overlay's postinst now marks it in /var/lib/keel-overlay-etcd/ package-member rather than removing it, and never marks a member that was there before.
marcos-mendez
pushed a commit
to Keel-Linux/keel
that referenced
this pull request
Oct 4, 2026
From the third review of keel#78: - a revocation names no serial: the holder revokes the intermediates it recorded for that address and WireGuard key, and only for the node's admitter, a trust root or the node itself; it signs an intermediate for an address only with that address's own key; - records carry etcd's cluster ID, an epoch that only grows, a nonce and an expiry, name exactly the cluster's members for `new` and `existing` alike, and a member takes none no newer than the last; only the holder adds a learner, so every record is the root's; - only the root signs intermediates, relayed by the inviter (0048's third round, point 1, narrowed): one intermediate deep, path length 0, name constrained to its member's own /128 and ::1, each recorded. When the holder cannot be reached the join completes without etcd, the request is queued, keel-mesh-etcd.timer asks again, and keel mesh status says so; - a member directory keel never started is removed only when keel-overlay-etcd marked it as etcd-server's own (Keel-Linux/common#39); any other is refused, named; - a renewal that fails is alerted through the monitor's channels and shown by keel mesh status and keel diff (etcd.certificates), as is a certificate within seven days of its expiry.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Companion of Keel-Linux/keel#78; merge this first.
etcd-healthasks/healthon etcd's plain metrics listenerhttp://[::1]:2381(keel rendersETCD_LISTEN_METRICS_URLS). The client port is TLS with client certificates now, so the old plain probe of 2379 failed every cycle and restarted etcd forever./var/lib/etcd/defaultwhenetcd-server's own postinst made it in the same transaction (preinst notes whether a member existed before). That start leaves a lone cluster of one. Data that was there before is never touched, and keel refuses to start etcd over a member it never started.