v0.222.0: ask whether a supervised member is DEAD before whether one is UNHEALTHY (R-384), and stop promising an undo copy nobody looked for (R-383)

R-384. aggregateState returned StateUnhealthy the moment unhealthy > 0, and the R-51 mixed-case block that asks "is a supervised member dead?" sat below it. A two-container app whose database exits goes unhealthy BECAUSE it cannot reach that database - so the symptom the dead database causes was what suppressed the alarm for it. unhealthy is not a down state, so classifyRunStates never marked the app down and app_start_failed never fired. Measured live on demo-hp 2026-08-22: bookstack-db stopped at 21:27:01 and the F-OBS heartbeat printed "0 currently down" throughout. R-51's 18-hour immich failure, back through a different door. Two things moved, and either alone leaves the defect standing: the supervised test is hoisted above the unhealthy/starting/restarting returns, and "some members are up" now counts ANY member not in the down bucket. The old guard was running > 0, which made the R-51 block unreachable in exactly the case it was written for. IsDownState is byte-identical - unhealthy stays excluded, because an unhealthy container is running and folding it in reintroduces the flapping that exclusion exists to stop. No new state was minted. Only the ORDER changed. The priority comment was rewritten because it asserted an ordering the code no longer has. Three subtests in TestAggregateState_UnchangedBranches were AMENDED: they asserted an unhealthy/starting/restarting member beat an exited peer on unless-stopped, which pinned the defect as settled behaviour. They keep their intent with the down member given a benign policy. R-383. The double-failure message said the previous state's backup EXISTS, built from the returned path without asking the filesystem - and a missing file is one of the two ways that rollback fails. undoCopyPhrase now describes the copy from disk: present, partial, missing (still naming where it should be), or never written. Zero-length counts as missing. Test count 1494 -> 1504. Four red-proofs planted, four seen failing; the two halves of R-384 convict independently.
2026-08-23 07:25:00 +02:00
parent da75603553
commit 5da11c4480
7 changed files with 638 additions and 27 deletions
@@ -87,30 +87,171 @@ func TestAggregateState_OneShotExitedMemberIsBenign(t *testing.T) {
 }

 // The unchanged branches: R-51 must not move any state the pre-existing aggregation produced.
+//
+// R-384 (v0.222.0) AMENDED THREE OF THESE CASES, and the amendment is the fix, not an accommodation.
+// They read `{a: unhealthy|starting|restarting, b: EXITED}` with BOTH members on `unless-stopped`,
+// and asserted that the live member's state won. That is precisely the defect R-384 closes: member
+// `b` is a dead SUPERVISED container, and the case was pinning the wrong answer as if it were
+// settled. A dead database beside an unhealthy front end read `unhealthy`, which is not a down
+// state, so nothing ever alarmed — measured live on `demo-hp` 2026-08-22.
+//
+// The cases keep their ORIGINAL INTENT — "the live member's state still wins over a down member" —
+// by giving `b` a BENIGN policy, which is the only situation in which that sentence was ever true.
+// The supervised versions moved to TestR384_DeadSupervisedMemberIsAskedAboutFirst, where they now
+// assert `degraded`.
 func TestAggregateState_UnchangedBranches(t *testing.T) {
 	all := policyMap(t, map[string]string{"a": "unless-stopped", "b": "unless-stopped"})
+	// `b` is a one-shot that finished: the live member's state must still win over it.
+	benignB := policyMap(t, map[string]string{"a": "unless-stopped", "b": "no"})
 	cases := []struct {
 		name       string
 		containers []ContainerInfo
+		policy     restartPolicyLookup
 		want       ContainerState
 	}{
-		{"no containers", nil, StateNotDeployed},
-		{"all running", []ContainerInfo{{Name: "a", State: StateRunning}, {Name: "b", State: StateRunning}}, StateRunning},
-		{"all stopped", []ContainerInfo{{Name: "a", State: StateExited}, {Name: "b", State: StateStopped}}, StateStopped},
-		{"any unhealthy wins", []ContainerInfo{{Name: "a", State: StateUnhealthy}, {Name: "b", State: StateExited}}, StateUnhealthy},
-		{"any starting wins over exited", []ContainerInfo{{Name: "a", State: StateStarting}, {Name: "b", State: StateExited}}, StateStarting},
-		{"any restarting wins over exited", []ContainerInfo{{Name: "a", State: StateRestarting}, {Name: "b", State: StateExited}}, StateRestarting},
-		{"single container exited", []ContainerInfo{{Name: "a", State: StateExited}}, StateStopped},
+		{"no containers", nil, all, StateNotDeployed},
+		{"all running", []ContainerInfo{{Name: "a", State: StateRunning}, {Name: "b", State: StateRunning}}, all, StateRunning},
+		{"all stopped", []ContainerInfo{{Name: "a", State: StateExited}, {Name: "b", State: StateStopped}}, all, StateStopped},
+		{"any unhealthy wins over a BENIGN exited", []ContainerInfo{{Name: "a", State: StateUnhealthy}, {Name: "b", State: StateExited}}, benignB, StateUnhealthy},
+		{"any starting wins over a BENIGN exited", []ContainerInfo{{Name: "a", State: StateStarting}, {Name: "b", State: StateExited}}, benignB, StateStarting},
+		{"any restarting wins over a BENIGN exited", []ContainerInfo{{Name: "a", State: StateRestarting}, {Name: "b", State: StateExited}}, benignB, StateRestarting},
+		{"single container exited", []ContainerInfo{{Name: "a", State: StateExited}}, all, StateStopped},
 	}
 	for _, tc := range cases {
 		t.Run(tc.name, func(t *testing.T) {
-			if got := aggregateState(tc.containers, all); got != tc.want {
+			if got := aggregateState(tc.containers, tc.policy); got != tc.want {
 				t.Fatalf("aggregateState = %q, want %q", got, tc.want)
 			}
 		})
 	}
 }

+// R-384 (v0.222.0) — the ORDER was the defect, not the `unhealthy` exclusion.
+//
+// THE LIVE FAILURE THIS PINS. On `demo-hp` 2026-08-22 `bookstack-db` (MariaDB, `unless-stopped`) was
+// stopped at 21:27:01. Its front end went `unhealthy` seconds later because it could not reach its
+// database. `aggregateState` returned StateUnhealthy at the `unhealthy > 0` line and never reached
+// the R-51 mixed-case block, so `classifyRunStates` never marked the app down and the F-OBS heartbeat
+// printed "0 currently down" throughout. The dead database hid behind the unhealthy survivor.
+//
+// Two things had to move, and both are asserted here:
+//  1. the supervised-down test runs BEFORE the unhealthy/starting/restarting returns;
+//  2. "some members are up" counts ANY member not in the down bucket. The old guard was `running > 0`
+//     counting StateRunning alone, which made the block unreachable in exactly this case — an
+//     unhealthy survivor counted as nothing up.
+//
+// RED-PROOF (observed, see REPORT.md): move the hoisted block back below the `unhealthy > 0` return
+// and the `unhealthy survivor` subtest fails with `aggregateState = "unhealthy", want "degraded"` —
+// the exact state the live box reported. Narrowing `up` back to `running` alone fails the same
+// subtest identically, so BOTH halves of the fix are convicted.
+func TestR384_DeadSupervisedMemberIsAskedAboutFirst(t *testing.T) {
+	supervised := policyMap(t, map[string]string{"web": "unless-stopped", "db": "unless-stopped"})
+	cases := []struct {
+		name     string
+		survivor ContainerState
+	}{
+		{"unhealthy survivor — the bookstack case measured live", StateUnhealthy},
+		{"starting survivor", StateStarting},
+		{"restarting survivor", StateRestarting},
+		{"running survivor — R-51's original case, must not regress", StateRunning},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			got := aggregateState([]ContainerInfo{
+				{Name: "web", State: tc.survivor, Status: "Up 3 minutes (unhealthy)"},
+				{Name: "db", State: StateExited, Status: "Exited (0) 3 minutes ago"},
+			}, supervised)
+			if got != StateDegraded {
+				t.Fatalf("survivor %q beside a dead SUPERVISED member: aggregateState = %q, want %q "+
+					"— a dead database must not hide behind it", tc.survivor, got, StateDegraded)
+			}
+			// Non-hollow: the state is only worth anything if it actually alarms.
+			if !IsDownState(got) {
+				t.Fatalf("IsDownState(%q) = false — the state changed but nothing downstream alarms", got)
+			}
+		})
+	}
+}
+
+// The fenced act (§12): `IsDownState` must stay byte-identical. R-384 works by asking a PRIOR
+// question, never by folding a running-but-failing container into the down set — that is the
+// flapping the exclusion exists to stop. This is the consequence-level guard: if a future change
+// widens IsDownState instead of reordering, this fails even though R-384's own tests still pass.
+func TestR384_IsDownStateWasNotWidened(t *testing.T) {
+	for _, s := range []ContainerState{StateUnhealthy, StateRestarting, StateStarting, StatePaused, StateUnknown, StateDeploying} {
+		if IsDownState(s) {
+			t.Fatalf("IsDownState(%q) = true — R-384 must reorder the question, never widen the down set", s)
+		}
+	}
+	for _, s := range []ContainerState{StateStopped, StateExited, StateDegraded} {
+		if !IsDownState(s) {
+			t.Fatalf("IsDownState(%q) = false — R-384 must not narrow the down set either", s)
+		}
+	}
+}
+
+// Scenario C, at the layer the defect would live: a migration step that finished must never alarm,
+// and R-384 hoisted the code that decides it. Every catalog app with an init container passes through
+// this, so a regression here alarms fleet-wide on every start.
+func TestR384_FinishedOneShotStaysBenignBehindAnUnhealthyMember(t *testing.T) {
+	for _, policy := range []string{"no", "on-failure"} {
+		t.Run(policy, func(t *testing.T) {
+			got := aggregateState([]ContainerInfo{
+				{Name: "app-web", State: StateUnhealthy, Status: "Up 1 minute (unhealthy)"},
+				{Name: "app-migrate", State: StateExited, Status: "Exited (0) 1 minute ago"},
+			}, policyMap(t, map[string]string{"app-web": "unless-stopped", "app-migrate": policy}))
+			if got != StateUnhealthy {
+				t.Fatalf("policy %q: aggregateState = %q, want %q — a finished migration is benign "+
+					"and must not turn an unhealthy app into a down one", policy, got, StateUnhealthy)
+			}
+			if IsDownState(got) {
+				t.Fatalf("policy %q: a finished one-shot made the stack alarm", policy)
+			}
+		})
+	}
+}
+
+// Scenario B, unchanged and byte-identical: unhealthy with NOTHING dead must stay unhealthy and must
+// not alarm. This is the flapping case the IsDownState exclusion exists to stop, and re-creating it
+// would be the over-correction rather than the fix.
+func TestR384_UnhealthyWithNothingDeadDoesNotAlarm(t *testing.T) {
+	for _, containers := range [][]ContainerInfo{
+		{{Name: "solo", State: StateUnhealthy}},
+		{{Name: "a", State: StateUnhealthy}, {Name: "b", State: StateUnhealthy}},
+		{{Name: "a", State: StateUnhealthy}, {Name: "b", State: StateRunning}},
+	} {
+		got := aggregateState(containers, policyMap(t, map[string]string{"solo": "unless-stopped", "a": "unless-stopped", "b": "unless-stopped"}))
+		if got != StateUnhealthy {
+			t.Fatalf("%d container(s), none down: aggregateState = %q, want %q", len(containers), got, StateUnhealthy)
+		}
+		if IsDownState(got) {
+			t.Fatalf("%d container(s), none down: the stack alarmed", len(containers))
+		}
+	}
+}
+
+// Scenario F: every member down. R-384 is a MEASUREMENT here, not a change (§4 of the task) — the
+// all-down path must be byte-identical, so that whatever the live walk finds about single-container
+// crashes is a separate, later decision and not something this change quietly moved.
+func TestR384_AllMembersDownIsUnchanged(t *testing.T) {
+	supervised := policyMap(t, map[string]string{"a": "unless-stopped", "b": "unless-stopped"})
+	cases := []struct {
+		name       string
+		containers []ContainerInfo
+	}{
+		{"single exited", []ContainerInfo{{Name: "a", State: StateExited}}},
+		{"single stopped", []ContainerInfo{{Name: "a", State: StateStopped}}},
+		{"both down", []ContainerInfo{{Name: "a", State: StateExited}, {Name: "b", State: StateStopped}}},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := aggregateState(tc.containers, supervised); got != StateStopped {
+				t.Fatalf("aggregateState = %q, want %q — R-384 must not move the all-down path", got, StateStopped)
+			}
+		})
+	}
+}
+
 // The unhealthy/restarting/paused/unknown exclusions are the fix-3 contract (downstate_test.go owns
 // them). This asserts the ONE addition, so a future reader can see R-51 widened the set by exactly
 // one state and by nothing else.
@@ -277,3 +418,52 @@ func TestRefreshStatus_InspectFailureStillDegrades(t *testing.T) {
 		t.Fatalf("failed inspect was cached: %v", m.restartPolicyCache)
 	}
 }
+
+// R-384 production-path wiring: the bookstack shape, through RefreshStatus.
+//
+// This is the SAME chain the box runs — docker ps → aggregateState → docker inspect → the stack map —
+// with the fixture measured live on `demo-hp` 2026-08-22: `bookstack-db` stopped, `bookstack` itself
+// still up but failing its healthcheck because it cannot reach its database. An aggregateState-only
+// test proves the function and not the caller, and this defect lived in the caller's reading of it.
+//
+// The classifier half of the chain is pinned by
+// TestR384_ADeadDatabaseBehindAnUnhealthyAppAlarms in cmd/controller — this test ends at the state,
+// that one starts from it, and TestR384_IsDownStateWasNotWidened pins the join.
+//
+// RED-PROOF (observed): with the hoisted block moved back below the `unhealthy > 0` return, this
+// fails with `bookstack state = "unhealthy", want "degraded"`.
+func TestR384_WiresTheDeadDatabaseThroughTheRealPath(t *testing.T) {
+	dock := &scriptedDocker{
+		ps: strings.Join([]string{
+			psLine("bookstack", "running", "Up 3 minutes (unhealthy)", "bookstack"),
+			psLine("bookstack-db", "exited", "Exited (0) 3 minutes ago", "bookstack"),
+			psLine("docmost", "running", "Up 3 hours (healthy)", "docmost"),
+		}, "\n"),
+		policies: map[string]string{"bookstack-db": "unless-stopped"},
+	}
+	m := &Manager{
+		cfg:    &config.Config{},
+		logger: log.New(io.Discard, "", 0),
+		execFn: dock.exec,
+		stacks: map[string]*Stack{
+			"bookstack": {Name: "bookstack", Deployed: true},
+			"docmost":   {Name: "docmost", Deployed: true},
+		},
+	}
+	if err := m.RefreshStatus(); err != nil {
+		t.Fatalf("RefreshStatus: %v", err)
+	}
+	got := m.stacks["bookstack"].State
+	if got != StateDegraded {
+		t.Fatalf("bookstack state = %q, want %q — a dead database must not hide behind its own "+
+			"unhealthy front end (measured live 2026-08-22: this read %q and nothing alarmed)",
+			got, StateDegraded, StateUnhealthy)
+	}
+	// Non-hollow: the state only matters because it alarms.
+	if !IsDownState(got) {
+		t.Fatalf("IsDownState(%q) = false — the state moved but the app is still silently down", got)
+	}
+	if got := m.stacks["docmost"].State; got != StateRunning {
+		t.Fatalf("docmost state = %q, want %q — a healthy app must be untouched", got, StateRunning)
+	}
+}
@@ -775,10 +775,49 @@ func supervisedPolicy(policy string) bool {
 }

 // aggregateState determines the overall stack state from its containers.
-// Priority: unhealthy/starting > restarting > all-running > degraded > stopped
 //
-// policyOf is consulted ONLY for the mixed case (some members up, some down) and ONLY for the down
-// members — see the mix branch. nil is allowed (every exited member then reads as supervised).
+// Priority: **degraded > unhealthy/starting > restarting > all-running > stopped**
+//
+// R-384 (v0.222.0) MOVED `degraded` to the front, and the move is the whole fix. It used to sit last,
+// below the `unhealthy > 0` return, which meant a different question was answering it.
+//
+// ── WHY DEGRADED IS ASKED FIRST ──────────────────────────────────────────────────────────────
+//
+// "Is a SUPERVISED member of this app dead?" and "is a RUNNING member failing its healthcheck?" are
+// two different questions, and until v0.222.0 the second was allowed to answer the first. A two-
+// container app whose database exits will normally have its front end go `unhealthy` moments later —
+// it cannot reach its database. The old order returned `StateUnhealthy` immediately at that point and
+// never reached the R-51 mixed-case block below, so the dead database was never asked about.
+// `unhealthy` is deliberately NOT a down state (see IsDownState), so `classifyRunStates` never marked
+// the app down and `NotifyAppStartFailures` never fired.
+//
+// **Measured live on `demo-hp` 2026-08-22:** `bookstack-db` stopped at 21:27:01 and the F-OBS
+// heartbeat printed "0 currently down" throughout. That is R-51's own 18-hour immich failure back
+// through a different door — the dead member now hiding behind an unhealthy survivor instead of
+// behind three live helpers.
+//
+// ── WHAT "SOME MEMBERS ARE UP" MEANS, AND WHY IT WIDENED ─────────────────────────────────────
+//
+// The R-51 block was additionally guarded by `running > 0`, where `running` counts only StateRunning.
+// That guard made the block UNREACHABLE in exactly the case it was written for: an unhealthy survivor
+// beside a dead database counted as nothing up. "Up" now means **any member not in the down bucket** —
+// running, unhealthy, starting or restarting — because each of those is a container Docker still has,
+// and a dead supervised peer beside any of them is a fault either way.
+//
+// ── WHAT DID NOT CHANGE, DELIBERATELY ────────────────────────────────────────────────────────
+//
+// `IsDownState` is byte-identical. `unhealthy` is still excluded from it, for the reason recorded
+// there: an unhealthy container is RUNNING, and folding it in reintroduces the flapping that
+// exclusion exists to stop. This change does not fold it in — it asks a prior question first. No new
+// state was minted either: `StateDegraded` already means precisely this and every consumer already
+// handles it.
+//
+// The benign case is untouched: a one-shot init/migrate container that has legitimately finished has
+// policy `no`/`on-failure`, `supervisedPolicy` returns false, and the stack reads exactly as before.
+// Without that filter every app with a migration step would alarm on every start.
+//
+// policyOf is consulted ONLY when some members are up and some are down, and ONLY for the down
+// members. nil is allowed (every exited member then reads as supervised).
 func aggregateState(containers []ContainerInfo, policyOf restartPolicyLookup) ContainerState {
 	if len(containers) == 0 {
 		return StateNotDeployed
@@ -809,6 +848,22 @@ func aggregateState(containers []ContainerInfo, policyOf restartPolicyLookup) Co

 	total := len(containers)

+	// R-384 — ASKED FIRST. "Is a supervised member dead?" must be answered before "is a running
+	// member unhealthy?", because a dying database drags its front end unhealthy and the unhealthy
+	// return below then swallowed the whole question. `up` counts every member NOT in the down
+	// bucket; `running > 0` alone made this unreachable in the exact case R-51 was written for.
+	if up := running + unhealthy + starting + restarting; up > 0 && len(down) > 0 {
+		for _, c := range down {
+			policy := ""
+			if policyOf != nil {
+				policy = policyOf(c.Name)
+			}
+			if supervisedPolicy(policy) {
+				return StateDegraded
+			}
+		}
+	}
+
 	// Any unhealthy → whole stack is unhealthy
 	if unhealthy > 0 {
 		return StateUnhealthy
@@ -832,19 +887,14 @@ func aggregateState(containers []ContainerInfo, policyOf restartPolicyLookup) Co
 	// Mix (some members up, some down) — R-51. Until v0.156.0 this reported StateRunning
 	// unconditionally ("partial"), which is why a dead immich-server behind three live helpers was
 	// invisible to fix-3 for 18 hours. A down member whose restart policy says docker should be
-	// keeping it up is a FAULT → the whole stack is degraded (and degraded is a down state). A down
-	// member with policy `no`/`on-failure` is a one-shot init/migrate container that has legitimately
-	// finished → benign, the stack stays running.
+	// keeping it up is a FAULT → the whole stack is degraded (and degraded is a down state).
+	//
+	// R-384 (v0.222.0) HOISTED that supervised test to the top of this function, so by the time
+	// control reaches here every down member has already been proven BENIGN — policy `no`/`on-failure`,
+	// a one-shot init/migrate container that has legitimately finished. The stack stays running. This
+	// branch therefore no longer decides anything about supervision; it only records the benign
+	// verdict, and the test that pins it is the one that says a finished migration must not alarm.
 	if running > 0 {
-		for _, c := range down {
-			policy := ""
-			if policyOf != nil {
-				policy = policyOf(c.Name)
-			}
-			if supervisedPolicy(policy) {
-				return StateDegraded
-			}
-		}
 		return StateRunning
 	}