Merge pull request 'Issues 260, 262: a service waits for what it reads; a mesh name has no IPv6 address rather than no name' (#111) from issues/260-262 into main

2026-10-05 20:08:22 +00:00
parent 553d0c8893 eb4b4256a1
commit 072489e653
2 changed files with 58 additions and 2 deletions
@@ -1,8 +1,8 @@
 ---
-status: open
+status: located
 opened: 2026-10-05
 located-in: [mesh-host, mesh-catalog]
-fixed-by:
+fixed-by: novox/mesh-host#25 (a566add)
 amended-design:
 ---

@@ -38,3 +38,13 @@ machine's name resolution with it.

 - Should the host run every restart and reload after all files of the apply are written?
 - Or should it order each restart after every resource the service names in `restart-on`?
+
+## Fix
+
+The host now orders the apply so that a service comes after every resource it names under
+`restart-on` or `reload-on`. Only services move, and only as far as the last of what they name; every
+other resource keeps its declared place. The first open question above is answered that way, per
+service, rather than by moving every restart to the end. This is checked by mesh-host's
+`restart_order_test.go`: a service restarting on two files, one declared after it, is restarted once
+both exist, and a later change to the second file still restarts it. The test fails without the
+ordering.
@@ -0,0 +1,46 @@
+---
+status: located
+opened: 2026-10-05
+located-in: [mesh-catalog]
+fixed-by: novox/mesh-catalog#71 (20603b6)
+amended-design:
+---
+
+# 262. An Alpine container could not find a machine by its mesh name
+
+## Symptom
+
+After the mesh moved to one resolver (ADR 0194, 0196), a workflow container on the home server was
+restarted so that it would ask the mesh's resolver. It then crash-looped every fourteen seconds:
+
+> getaddrinfo ENOTFOUND <anchor>.internal
+
+From the same container, `getent hosts <anchor>.internal` answered correctly. So did the machine
+itself, every time, in about forty-five milliseconds.
+
+## Cause
+
+The mesh's resolver answered a machine's name only through a wildcard rule:
+
+- asked for the IPv4 address, it gave the address;
+- asked for the IPv6 address, it answered NXDOMAIN, "no such name", where the correct answer is
+  NODATA, "the name exists and has no such record".
+
+glibc ignores that. musl, the C library of every Alpine image, asks for both records and takes the
+NXDOMAIN as final, so the whole lookup failed. The per-machine resolvers this one replaced answered
+each machine's name from `/etc/hosts`, which gives NODATA, so nothing had depended on the difference
+before.
+
+Reproduced with a throwaway resolver of the same version, given only the wildcard rule.
+
+## Fix
+
+Each machine's name is now also a host record in the machine list the resolver reads. Asked for an
+IPv6 address, the resolver then answers that the name exists and has none. A name under a machine,
+answered only by the wildcard, still answers NXDOMAIN for IPv6, exactly as it did before the move.
+
+## How it is checked
+
+Ask the mesh's resolver for the IPv6 address of a machine's name. The answer must be NOERROR with no
+records. Today that check is done by hand. A controller test that renders the machine list and
+requires one host record per machine should be added.