fix(decaymap): fix lock convoy (#1106)

* fix(decaymap): fix lock convoy

Ref #1103

This uses the actor pattern to delay deletion instead of making things
fight over a lock. It also properly fixes locking logic to prevent the
convoy problem.

Signed-off-by: Xe Iaso <me@xeiaso.net>

* docs: update CHANGELOG

Signed-off-by: Xe Iaso <me@xeiaso.net>

---------

Signed-off-by: Xe Iaso <me@xeiaso.net>

decaymap/decaymap.go

@@ -14,6 +14,12 @@ func Zilch[T any]() T {
 type Impl[K comparable, V any] struct {
 	data map[K]decayMapEntry[V]
 	lock sync.RWMutex
+
+	// deleteCh receives decay-deletion requests from readers.
+	deleteCh chan deleteReq[K]
+	// stopCh stops the background cleanup worker.
+	stopCh chan struct{}
+	wg     sync.WaitGroup
 }
 
 type decayMapEntry[V any] struct {
@@ -21,30 +27,38 @@ type decayMapEntry[V any] struct {
 	expiry time.Time
 }
 
+// deleteReq is a request to remove a key if its expiry timestamp still matches
+// the observed one. This prevents racing with concurrent Set updates.
+type deleteReq[K comparable] struct {
+	key    K
+	expiry time.Time
+}
+
 // New creates a new DecayMap of key type K and value type V.
 //
 // Key types must be comparable to work with maps.
 func New[K comparable, V any]() *Impl[K, V] {
-	return &Impl[K, V]{
-		data: make(map[K]decayMapEntry[V]),
+	m := &Impl[K, V]{
+		data:     make(map[K]decayMapEntry[V]),
+		deleteCh: make(chan deleteReq[K], 1024),
+		stopCh:   make(chan struct{}),
 	}
+	m.wg.Add(1)
+	go m.cleanupWorker()
+	return m
 }
 
 // expire forcibly expires a key by setting its time-to-live one second in the past.
 func (m *Impl[K, V]) expire(key K) bool {
-	m.lock.RLock()
+	// Use a single write lock to avoid RUnlock->Lock convoy.
+	m.lock.Lock()
+	defer m.lock.Unlock()
 	val, ok := m.data[key]
-	m.lock.RUnlock()
-
 	if !ok {
 		return false
 	}
-
-	m.lock.Lock()
 	val.expiry = time.Now().Add(-1 * time.Second)
 	m.data[key] = val
-	m.lock.Unlock()
-
 	return true
 }
 
@@ -53,19 +67,14 @@ func (m *Impl[K, V]) expire(key K) bool {
 // If the value does not exist, return false. Return true after
 // deletion.
 func (m *Impl[K, V]) Delete(key K) bool {
-	m.lock.RLock()
-	_, ok := m.data[key]
-	m.lock.RUnlock()
-
-	if !ok {
-		return false
-	}
-
+	// Use a single write lock to avoid RUnlock->Lock convoy.
 	m.lock.Lock()
-	delete(m.data, key)
-	m.lock.Unlock()
-
-	return true
+	defer m.lock.Unlock()
+	_, ok := m.data[key]
+	if ok {
+		delete(m.data, key)
+	}
+	return ok
 }
 
 // Get gets a value from the DecayMap by key.
@@ -81,13 +90,12 @@ func (m *Impl[K, V]) Get(key K) (V, bool) {
 	}
 
 	if time.Now().After(value.expiry) {
-		m.lock.Lock()
-		// Since previously reading m.data[key], the value may have been updated.
-		// Delete the entry only if the expiry time is still the same.
-		if m.data[key].expiry.Equal(value.expiry) {
-			delete(m.data, key)
+		// Defer decay deletion to the background worker to avoid convoy.
+		select {
+		case m.deleteCh <- deleteReq[K]{key: key, expiry: value.expiry}:
+		default:
+			// Channel full: drop request; a future Cleanup() or Get will retry.
 		}
-		m.lock.Unlock()
 
 		return Zilch[V](), false
 	}
@@ -125,3 +133,64 @@ func (m *Impl[K, V]) Len() int {
 	defer m.lock.RUnlock()
 	return len(m.data)
 }
+
+// Close stops the background cleanup worker. It's optional to call; maps live
+// for the process lifetime in many cases. Call in tests or when you know you no
+// longer need the map to avoid goroutine leaks.
+func (m *Impl[K, V]) Close() {
+	close(m.stopCh)
+	m.wg.Wait()
+}
+
+// cleanupWorker batches decay deletions to minimize lock contention.
+func (m *Impl[K, V]) cleanupWorker() {
+	defer m.wg.Done()
+	batch := make([]deleteReq[K], 0, 64)
+	ticker := time.NewTicker(10 * time.Millisecond)
+	defer ticker.Stop()
+
+	flush := func() {
+		if len(batch) == 0 {
+			return
+		}
+		m.applyDeletes(batch)
+		// reset batch without reallocating
+		batch = batch[:0]
+	}
+
+	for {
+		select {
+		case req := <-m.deleteCh:
+			batch = append(batch, req)
+		case <-ticker.C:
+			flush()
+		case <-m.stopCh:
+			// Drain any remaining requests then exit
+			for {
+				select {
+				case req := <-m.deleteCh:
+					batch = append(batch, req)
+				default:
+					flush()
+					return
+				}
+			}
+		}
+	}
+}
+
+func (m *Impl[K, V]) applyDeletes(batch []deleteReq[K]) {
+	now := time.Now()
+	m.lock.Lock()
+	for _, req := range batch {
+		entry, ok := m.data[req.key]
+		if !ok {
+			continue
+		}
+		// Only delete if the expiry is unchanged and already past.
+		if entry.expiry.Equal(req.expiry) && now.After(entry.expiry) {
+			delete(m.data, req.key)
+		}
+	}
+	m.lock.Unlock()
+}