ingestion/internal/symlinkguard/symlinkguard.go

@@ -144,7 +144,11 @@ func Scan(brainDir string) (Result, error) {
 // scan waited on it a planted link stayed readable for several intervals.
 // The scan hands alerts over without blocking; any that do not fit are
 // counted in the summary. When ctx ends, Watch returns after a final flush
-// of what was handed over (bounded by shutdownFlushTimeout).
+// of what was handed over (bounded by shutdownFlushTimeout); alerts that
+// flush cannot deliver are counted in one error log line instead. That
+// shutdown path only runs for a caller that cancels ctx: cmd/server passes
+// context.Background(), so in production the flush on handoff is what keeps
+// alerts from waiting on a restart.
 func Watch(ctx context.Context, brainDir string, interval time.Duration, n Notifier) {
 	if interval <= 0 {
 		interval = DefaultInterval
@@ -211,7 +215,8 @@ func (s alertSender) push(msg string) {
 // over, and again every interval while any are pending (outage retries).
 // When alerts is closed (Watch is stopping) it drains it and flushes once
 // more, bounded by shutdownFlushTimeout, so an alert handed over just before
-// shutdown is not lost (PR #180 review).
+// shutdown is not lost (PR #180 review). Whatever that flush cannot deliver
+// is logged as one error with its counts, never dropped silently.
 func deliver(ctx context.Context, interval time.Duration, n Notifier, alerts <-chan string, overflow *atomic.Int64) {
 	t := time.NewTicker(interval)
 	defer t.Stop()
@@ -240,9 +245,14 @@ func deliver(ctx context.Context, interval time.Duration, n Notifier, alerts <-c
 				ok = drain()
 			}
 			if !ok {
+				q.suppressed += int(overflow.Swap(0))
 				fctx, cancel := context.WithTimeout(context.WithoutCancel(ctx), shutdownFlushTimeout)
 				q.flush(fctx, n)
 				cancel()
+				if len(q.items) > 0 || q.suppressed > 0 {
+					slog.Error("symlink tripwire alerts not delivered at shutdown",
+						"undelivered", len(q.items), "suppressed", q.suppressed)
+				}
 				return
 			}
 			q.flush(ctx, n)

ingestion/internal/symlinkguard/symlinkguard_test.go

@@ -1,10 +1,12 @@
 package symlinkguard
 
 import (
+	"bytes"
 	"context"
 	"errors"
 	"fmt"
 	"os"
+	"log/slog"
 	"path/filepath"
 	"strconv"
 	"strings"
@@ -345,13 +347,19 @@ func TestWatchDropsAMessageThatNeverSendsSoLaterAlertsFlow(t *testing.T) {
 	}, 3*time.Second, 10*time.Millisecond, "a later alert is delivered despite a poison message")
 }
 
-// toggleNotifier fails while down is set.
+// toggleNotifier fails while down is set. Like the real audit.NtfyNotifier
+// (http.NewRequestWithContext), it also fails, without recording, when its
+// ctx is already done (PR #180 round 2: notifiers that ignored ctx let a
+// final flush on the cancelled ctx pass).
 type toggleNotifier struct {
 	recNotifier
 	down atomic.Bool
 }
 
 func (r *toggleNotifier) Notify(ctx context.Context, m string) error {
+	if err := ctx.Err(); err != nil {
+		return err
+	}
 	if r.down.Load() {
 		return errors.New("ntfy down")
 	}
@@ -669,8 +677,9 @@ func TestWatchCountsHandoffOverflowWithASlowNotifier(t *testing.T) {
 
 // An alert still queued when Watch stops (its send failed, the notifier has
 // since recovered, the retry tick has not come) is flushed once more before
-// Watch returns.
-func TestWatchFlushesTheQueueOnceMoreAtShutdown(t *testing.T) {
+// Watch returns, on a ctx that is not the cancelled one: toggleNotifier fails
+// on a done ctx, as the real ntfy notifier does.
+func TestWatchFinalFlushSendsAQueuedAlertOnACtxThatOutlivesTheCancel(t *testing.T) {
 	brain := t.TempDir()
 	n := &toggleNotifier{}
 	n.down.Store(true)
@@ -686,3 +695,80 @@ func TestWatchFlushesTheQueueOnceMoreAtShutdown(t *testing.T) {
 	<-done
 	assert.Equal(t, 1, n.count(), "the queued alert was not flushed at shutdown")
 }
+
+// lockedBuffer is a goroutine-safe log sink: Watch goroutines of earlier
+// tests may still log while one test captures.
+type lockedBuffer struct {
+	mu  sync.Mutex
+	buf bytes.Buffer
+}
+
+func (b *lockedBuffer) Write(p []byte) (int, error) {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	return b.buf.Write(p)
+}
+
+func (b *lockedBuffer) String() string { b.mu.Lock(); defer b.mu.Unlock(); return b.buf.String() }
+
+func captureLog(t *testing.T) *lockedBuffer {
+	t.Helper()
+	b := &lockedBuffer{}
+	prev := slog.Default()
+	slog.SetDefault(slog.New(slog.NewTextHandler(b, nil)))
+	t.Cleanup(func() { slog.SetDefault(prev) })
+	return b
+}
+
+// PR #180 round 2: when the final flush fails (ntfy hung at shutdown) the
+// undelivered alerts vanished with only "will retry" warnings. Watch now logs
+// one error with both counts.
+func TestWatchLogsUndeliveredAndSuppressedCountsWhenTheFinalFlushFails(t *testing.T) {
+	prev := shutdownFlushTimeout
+	shutdownFlushTimeout = 50 * time.Millisecond
+	t.Cleanup(func() { shutdownFlushTimeout = prev })
+	logs := captureLog(t)
+	brain := t.TempDir()
+	ctx, cancel := context.WithCancel(context.Background())
+	done := make(chan struct{})
+	go func() { Watch(ctx, brain, 5*time.Millisecond, hungNotifier{latency: time.Hour}); close(done) }()
+	link := filepath.Join(brain, "lost")
+	require.NoError(t, os.Symlink(t.TempDir(), link))
+	waitGone(t, link)
+	cancel()
+	<-done
+	out := logs.String()
+	assert.Contains(t, out, "level=ERROR")
+	assert.Contains(t, out, `msg="symlink tripwire alerts not delivered at shutdown"`)
+	assert.Contains(t, out, "undelivered=1")
+	assert.Contains(t, out, "suppressed=0")
+}
+
+// deliver, given a handoff that already holds alerts and is closed, moves them
+// into the queue and runs the final flush, even though the ctx is cancelled:
+// the drain-on-close path, without depending on when the close lands.
+func TestDeliverFlushesAlertsDrainedFromAClosedHandoffOnACancelledCtx(t *testing.T) {
+	alerts := make(chan string, 3)
+	alerts <- "a1"
+	alerts <- "a2"
+	close(alerts)
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	n := &toggleNotifier{}
+	var overflow atomic.Int64
+	deliver(ctx, time.Hour, n, alerts, &overflow)
+	assert.Equal(t, []string{"a1", "a2"}, n.msgs)
+}
+
+// PR #180 round 2: when the first receive after close finds the handoff empty,
+// overflow counted by the scan was never moved into the summary.
+func TestDeliverCountsOverflowWhenTheHandoffIsClosedAndEmpty(t *testing.T) {
+	alerts := make(chan string)
+	close(alerts)
+	n := &recNotifier{}
+	var overflow atomic.Int64
+	overflow.Store(3)
+	deliver(context.Background(), time.Hour, n, alerts, &overflow)
+	require.Len(t, n.msgs, 1)
+	assert.Contains(t, n.msgs[0], "3 further alert(s)")
+}