fleetd #669: report collaborator reload restart
CI / shell-tests (pull_request) Failing after 11s
CI / contract (pull_request) Successful in 49s
CI / build (pull_request) Failing after 2m1s

This commit is contained in:
Dai Ha
2026-10-04 06:12:37 +02:00
parent 16b52fac6c
commit f82073717a
2 changed files with 100 additions and 2 deletions
@@ -615,7 +615,7 @@ public final class ConfigRef implements Supplier<FleetConfig> {
// may bind to, AND what a slot already bound still grants) through its own instance of that
// same supplier shape — see MemberRegistry.live and its class doc for the binding rule:
// removing a slot revokes ARCHITECT on the bound pane's very next request, and only the slot
// OCCUPANCY survives, so the demoted session keeps its slot key until it unbinds. Only
// OCCUPANCY survives, so the demoted session keeps its slot key until it unbinds.
// fleet.leaders is frozen (Fleetd.java:281 reads cfg.fleet().leaders() off the startup
// snapshot to build both the LeadTabScanner's tab-label-to-name map, wired into
// CallerResolver.withLeadsAndMembers at Fleetd.java:620/624, and — when herdr answered —
@@ -635,6 +635,11 @@ public final class ConfigRef implements Supplier<FleetConfig> {
+ "live through that same supplier for placement AND through a separate supplier "
+ "on MemberRegistry for spawn-time identity — both already applied");
}
if (!Objects.equals(collaboratorsOf(old), collaboratorsOf(fresh))) {
changed.add("fleet: fleet.collaborators (each collaborator's tab) is read once at "
+ "startup to build the LeadTabScanner's identity map, which is not rebuilt on "
+ "reload, so a collaborator added, removed, or given a new tab: needs a restart");
}
// Kept in step with SPLIT_KEYS the same way changedColdKeys is kept in step with COLD_KEYS —
// every message here must be traceable to one of the split keys the class doc documents.
// NOTE what this does NOT prove, per the javadoc above: it does not catch a SPLIT_KEYS
@@ -650,6 +655,11 @@ public final class ConfigRef implements Supplier<FleetConfig> {
return cfg.fleet() == null ? Map.of() : cfg.fleet().leaders();
}
/** {@code cfg.fleet().collaborators()}, defensively, in case a caller hands in a non-defaulted config. */
private static Map<String, FleetConfig.Collaborator> collaboratorsOf(FleetConfig cfg) {
return cfg.fleet() == null ? Map.of() : cfg.fleet().collaborators();
}
/**
* {@link FleetConfig.Profile} record components deliberately left out of
* {@link #sameLaunchSettings} because they are read <em>live</em>, not baked in at spawn — see
@@ -107,6 +107,8 @@ class ConfigRefTest {
assertTrue(out.applied());
assertTrue(out.deferred().isEmpty());
assertTrue(out.split().stream().noneMatch(s -> s.contains("fleet.collaborators")),
out.split().toString());
assertEquals("new charter", ref.get().fleet().charterFor(
dev.ltms.fleet.peer.MemberRole.ARCHITECT));
}
@@ -786,14 +788,100 @@ class ConfigRefTest {
assertTrue(out.split().getFirst().startsWith("fleet:"), out.split().toString());
assertTrue(out.split().getFirst().contains("restart"), out.split().toString());
assertTrue(out.split().getFirst().contains("live"), out.split().toString());
assertTrue(out.split().stream().noneMatch(s -> s.contains("fleet.collaborators")),
out.split().toString());
assertTrue(out.summary().contains("partially live"), out.summary());
// The snapshot still carries the new value — the LeadTabScanner's identity map and
// LeadLauncher's auto-launch are what wait for a restart; a reload rebuilds neither.
assertEquals("lead: opus-b", ref.get().fleet().leaders().get("opus").tab());
}
@Test
void addingAFleetCollaboratorIsReportedAsSplit(@TempDir Path dir) throws Exception {
assertCollaboratorChangeIsReported(dir, """
fleet:
collaborators:
alex:
tab: "collaborator: alex"
""", "add a collaborator");
}
@Test
void removingAFleetCollaboratorIsReportedAsSplit(@TempDir Path dir) throws Exception {
Path f = dir.resolve("fleetd.yaml");
Files.writeString(f, yaml("""
fleet:
collaborators:
alex:
tab: "collaborator: alex"
"""));
ConfigRef ref = refFor(f);
Files.writeString(f, yaml("fleet: {}\n"));
assertCollaboratorSplit(ref.reload(), "remove a collaborator");
}
@Test
void changingAFleetCollaboratorTabIsReportedAsSplit(@TempDir Path dir) throws Exception {
Path f = dir.resolve("fleetd.yaml");
Files.writeString(f, yaml("""
fleet:
collaborators:
alex:
tab: "collaborator: alex-a"
"""));
ConfigRef ref = refFor(f);
Files.writeString(f, yaml("""
fleet:
collaborators:
alex:
tab: "collaborator: alex-b"
"""));
assertCollaboratorSplit(ref.reload(), "change a collaborator tab");
}
@Test
void unchangedFleetCollaboratorsProduceNoSplitReport(@TempDir Path dir) throws Exception {
Path f = dir.resolve("fleetd.yaml");
String config = yaml("""
fleet:
collaborators:
alex:
tab: "collaborator: alex"
""");
Files.writeString(f, config);
ConfigRef ref = refFor(f);
Files.writeString(f, config);
ConfigRef.Outcome out = ref.reload();
assertTrue(out.applied());
assertTrue(out.split().isEmpty(), out.split().toString());
}
private static void assertCollaboratorChangeIsReported(Path dir, String changed, String action)
throws Exception {
Path f = dir.resolve("fleetd.yaml");
Files.writeString(f, yaml("fleet: {}\n"));
ConfigRef ref = refFor(f);
Files.writeString(f, yaml(changed));
assertCollaboratorSplit(ref.reload(), action);
}
private static void assertCollaboratorSplit(ConfigRef.Outcome out, String action) {
assertTrue(out.applied(), action);
assertTrue(out.deferred().isEmpty(), out.deferred().toString());
assertEquals(1, out.split().size(), out.split().toString());
String report = out.split().getFirst();
assertTrue(report.startsWith("fleet: fleet.collaborators"), report);
assertTrue(report.contains("read once at startup"), report);
assertTrue(report.contains("restart"), report);
}
/**
* fleetd #333: {@code fleet.leaders} is the ONLY frozen part of {@code fleet:}. A reload that
* fleetd #333: {@code fleet.leaders} is a frozen part of {@code fleet:}. A reload that
* changes {@code tabLabel} (or charters, or a role pool) without touching {@code fleet.leaders}
* must stay fully hot with nothing reported — proving {@link ConfigRef#changedSplitKeys}
* compares {@code fleet.leaders} specifically rather than the whole {@code Fleet} record, which