some tweaks to cart

cmd/cart/main.go

@@ -413,7 +413,7 @@ func main() {
 	}
 	defer grpcSrv.GracefulStop()
 
-	// go diskStorage.SaveLoop(10 * time.Second)
+	go diskStorage.SaveLoop(20 * time.Second)
 	UseDiscovery(pool)
 
 	ctx, stop := signal.NotifyContext(context.Background(), os.Interrupt)

pkg/actor/disk_storage.go

@@ -13,16 +13,31 @@ import (
 	"google.golang.org/protobuf/proto"
 )
 
+const defaultWriterIdleTTL = 5 * time.Minute
+
 type QueueEvent struct {
 	TimeStamp time.Time
 	Message   proto.Message
 }
 
+// diskLogWriter holds one append-only log file kept open between writes.
+type diskLogWriter struct {
+	mu       sync.Mutex
+	file     *os.File
+	lastUsed time.Time
+	purged   bool // set by purgeIdleWriters before closing; getWriter re-checks after acquiring mu
+}
+
 type DiskStorage[V any] struct {
 	*StateStorage
-	path  string
-	done  chan struct{}
-	queue *sync.Map // map[uint64][]QueueEvent
+	path          string
+	done          chan struct{}
+	queue         *sync.Map // map[uint64][]QueueEvent
+	dirOnce       sync.Once
+	dirErr        error
+	writersMu     sync.Mutex
+	writers       map[uint64]*diskLogWriter
+	writerIdleTTL time.Duration
 }
 
 type LogStorage[V any] interface {
@@ -32,10 +47,140 @@ type LogStorage[V any] interface {
 }
 
 func NewDiskStorage[V any](path string, registry MutationRegistry) *DiskStorage[V] {
-	return &DiskStorage[V]{
-		StateStorage: NewState(registry),
-		path:         path,
-		done:         make(chan struct{}),
+	s := &DiskStorage[V]{
+		StateStorage:  NewState(registry),
+		path:          path,
+		done:          make(chan struct{}),
+		writers:       make(map[uint64]*diskLogWriter),
+		writerIdleTTL: defaultWriterIdleTTL,
+	}
+	go s.purgeLoop()
+	return s
+}
+
+func (s *DiskStorage[V]) ensureDir() error {
+	s.dirOnce.Do(func() {
+		s.dirErr = os.MkdirAll(s.path, 0o755)
+	})
+	return s.dirErr
+}
+
+// getWriter returns an open append handle for id. The caller must unlock w.mu
+// when finished writing.
+
+
+func (s *DiskStorage[V]) writeEvents(id uint64, events []QueueEvent) error {
+	if len(events) == 0 {
+		return nil
+	}
+	w, err := s.getWriter(id)
+	if err != nil {
+		return err
+	}
+	defer w.mu.Unlock()
+
+	for _, evt := range events {
+		if err := s.Append(w.file, evt.Message, evt.TimeStamp); err != nil {
+			return err
+		}
+	}
+	w.lastUsed = time.Now()
+	return nil
+}
+
+func (s *DiskStorage[V]) writeMutations(id uint64, msgs ...proto.Message) error {
+	if len(msgs) == 0 {
+		return nil
+	}
+	w, err := s.getWriter(id)
+	if err != nil {
+		return err
+	}
+	defer w.mu.Unlock()
+
+	now := time.Now()
+	for _, m := range msgs {
+		if err := s.Append(w.file, m, now); err != nil {
+			return err
+		}
+	}
+	w.lastUsed = time.Now()
+	return nil
+}
+
+func (s *DiskStorage[V]) purgeLoop() {
+	ticker := time.NewTicker(time.Minute)
+	defer ticker.Stop()
+	for {
+		select {
+		case <-s.done:
+			return
+		case <-ticker.C:
+			s.purgeIdleWriters()
+		}
+	}
+}
+
+func (s *DiskStorage[V]) purgeIdleWriters() {
+	cutoff := time.Now().Add(-s.writerIdleTTL)
+
+	s.writersMu.Lock()
+	defer s.writersMu.Unlock()
+
+	for id, w := range s.writers {
+		w.mu.Lock()
+		if w.lastUsed.Before(cutoff) {
+			w.purged = true
+			delete(s.writers, id)
+			if err := w.file.Close(); err != nil {
+				log.Printf("failed to close idle event log %d: %v", id, err)
+			}
+		}
+		w.mu.Unlock()
+	}
+}
+
+func (s *DiskStorage[V]) getWriter(id uint64) (*diskLogWriter, error) {
+	if err := s.ensureDir(); err != nil {
+		return nil, err
+	}
+
+	for {
+		s.writersMu.Lock()
+		w, ok := s.writers[id]
+		if !ok {
+			fh, err := os.OpenFile(s.logPath(id), os.O_APPEND|os.O_CREATE|os.O_WRONLY, 0o644)
+			if err != nil {
+				s.writersMu.Unlock()
+				return nil, err
+			}
+			w = &diskLogWriter{file: fh, lastUsed: time.Now()}
+			s.writers[id] = w
+		}
+		s.writersMu.Unlock()
+
+		w.mu.Lock()
+		if !w.purged {
+			w.lastUsed = time.Now()
+			return w, nil
+		}
+		// Writer was purged by purgeIdleWriters between our map lookup and acquiring w.mu.
+		// Release it and retry — the map no longer contains this entry, so the next
+		// iteration will create a fresh writer.
+		w.mu.Unlock()
+	}
+}
+
+func (s *DiskStorage[V]) closeWriters() {
+	s.writersMu.Lock()
+	defer s.writersMu.Unlock()
+	for id, w := range s.writers {
+		w.mu.Lock()
+		if err := w.file.Close(); err != nil {
+			log.Printf("failed to close event log %d: %v", id, err)
+		}
+		w.mu.Unlock()
+		delete(s.writers, id)
 	}
 }
 
@@ -59,22 +204,16 @@ func (s *DiskStorage[V]) save() {
 	lines := 0
 	s.queue.Range(func(key, value any) bool {
 		id := key.(uint64)
-		path := s.logPath(id)
-		fh, err := os.OpenFile(path, os.O_APPEND|os.O_CREATE|os.O_WRONLY, 0644)
-		if err != nil {
-			log.Printf("failed to open event log file: %v", err)
+		qe, ok := value.([]QueueEvent)
+		if !ok {
+			s.queue.Delete(id)
 			return true
 		}
-		defer fh.Close()
-
-		if qe, ok := value.([]QueueEvent); ok {
-			for _, msg := range qe {
-				if err := s.Append(fh, msg.Message, msg.TimeStamp); err != nil {
-					log.Printf("failed to append event to log file: %v", err)
-				}
-				lines++
-			}
+		if err := s.writeEvents(id, qe); err != nil {
+			log.Printf("failed to append events for grain %d: %v", id, err)
+			return true
 		}
+		lines += len(qe)
 		carts++
 		s.queue.Delete(id)
 		return true
@@ -130,6 +269,7 @@ func (s *DiskStorage[V]) Close() {
 	if s.queue != nil {
 		s.save()
 	}
+	s.closeWriters()
 	close(s.done)
 }
 
@@ -145,25 +285,12 @@ func (s *DiskStorage[V]) AppendMutations(id uint64, msg ...proto.Message) error
 		}
 		s.queue.Store(id, queue)
 		return nil
-	} else {
-		path := s.logPath(id)
-		// Ensure the parent directory exists (e.g. first write to a new PVC mount).
-		// MkdirAll is a no-op when the directory already exists, so it's safe to
-		// call on every AppendMutations without extra stat overhead.
-		if err := os.MkdirAll(s.path, 0755); err != nil {
-			log.Printf("failed to create event log directory %s: %v", s.path, err)
-			return err
-		}
-		fh, err := os.OpenFile(path, os.O_APPEND|os.O_CREATE|os.O_WRONLY, 0644)
-		if err != nil {
-			log.Printf("failed to open event log file: %v", err)
-			return err
-		}
-		defer fh.Close()
-		for _, m := range msg {
-			err = s.Append(fh, m, time.Now())
-		}
-		return err
 	}
-
+	return s.writeMutations(id, msg...)
+}
+
+func (s *DiskStorage[V]) openWriterCount() int {
+	s.writersMu.Lock()
+	defer s.writersMu.Unlock()
+	return len(s.writers)
 }

pkg/actor/disk_storage_test.go

@@ -0,0 +1,319 @@
+package actor
+
+import (
+	"context"
+	"sync"
+	"testing"
+	"time"
+
+	cart_messages "git.k6n.net/mats/go-cart-actor/proto/cart"
+	"google.golang.org/protobuf/proto"
+)
+
+type diskTestGrain struct {
+	testState
+	id       uint64
+	accessed time.Time
+}
+
+func (g *diskTestGrain) GetId() uint64            { return g.id }
+func (g *diskTestGrain) GetLastAccess() time.Time { return g.accessed }
+func (g *diskTestGrain) GetLastChange() time.Time { return g.accessed }
+func (g *diskTestGrain) GetCurrentState() (*testState, error) {
+	return &g.testState, nil
+}
+
+func diskTestRegistry(t *testing.T) MutationRegistry {
+	t.Helper()
+	reg := NewMutationRegistry()
+	reg.RegisterMutations(NewMutation(
+		func(_ *diskTestGrain, _ *cart_messages.AddItem) error { return nil },
+	))
+	return reg
+}
+
+func TestDiskStorageAppendAndReplay(t *testing.T) {
+	reg := diskTestRegistry(t)
+	storage := NewDiskStorage[testState](t.TempDir(), reg)
+	t.Cleanup(storage.Close)
+
+	const id = uint64(99)
+	events := []proto.Message{
+		&cart_messages.AddItem{ItemId: 1, Quantity: 2},
+		&cart_messages.AddItem{ItemId: 2, Quantity: 1},
+	}
+	for _, evt := range events {
+		if err := storage.AppendMutations(id, evt); err != nil {
+			t.Fatalf("append: %v", err)
+		}
+	}
+
+	grain := &diskTestGrain{id: id, accessed: time.Now()}
+	if err := storage.LoadEvents(context.Background(), id, grain); err != nil {
+		t.Fatalf("replay: %v", err)
+	}
+}
+
+func TestDiskStorageReusesOpenWriter(t *testing.T) {
+	reg := diskTestRegistry(t)
+	storage := NewDiskStorage[testState](t.TempDir(), reg)
+	t.Cleanup(storage.Close)
+
+	const id = uint64(7)
+	msg := &cart_messages.AddItem{ItemId: 1, Quantity: 1}
+	if err := storage.AppendMutations(id, msg); err != nil {
+		t.Fatal(err)
+	}
+	if got := storage.openWriterCount(); got != 1 {
+		t.Fatalf("open writers after first append = %d, want 1", got)
+	}
+	if err := storage.AppendMutations(id, msg); err != nil {
+		t.Fatal(err)
+	}
+	if got := storage.openWriterCount(); got != 1 {
+		t.Fatalf("open writers after second append = %d, want 1", got)
+	}
+}
+
+func TestDiskStorageIdleWriterClosed(t *testing.T) {
+	reg := diskTestRegistry(t)
+	storage := NewDiskStorage[testState](t.TempDir(), reg)
+	storage.writerIdleTTL = 20 * time.Millisecond
+	t.Cleanup(storage.Close)
+
+	if err := storage.AppendMutations(1, &cart_messages.AddItem{ItemId: 1, Quantity: 1}); err != nil {
+		t.Fatal(err)
+	}
+	if storage.openWriterCount() != 1 {
+		t.Fatal("expected one open writer")
+	}
+
+	time.Sleep(30 * time.Millisecond)
+	storage.purgeIdleWriters()
+	if storage.openWriterCount() != 0 {
+		t.Fatalf("expected idle writer to be closed, got %d open", storage.openWriterCount())
+	}
+}
+
+func TestDiskStorageSaveLoopUsesWriterCache(t *testing.T) {
+	reg := diskTestRegistry(t)
+	storage := NewDiskStorage[testState](t.TempDir(), reg)
+	t.Cleanup(storage.Close)
+
+	go storage.SaveLoop(20 * time.Millisecond)
+
+	const id = uint64(55)
+	if err := storage.AppendMutations(id, &cart_messages.AddItem{ItemId: 3, Quantity: 4}); err != nil {
+		t.Fatal(err)
+	}
+
+	deadline := time.After(500 * time.Millisecond)
+	for {
+		if storage.openWriterCount() >= 1 {
+			break
+		}
+		select {
+		case <-deadline:
+			t.Fatal("timed out waiting for SaveLoop flush")
+		default:
+			time.Sleep(10 * time.Millisecond)
+		}
+	}
+
+	grain := &diskTestGrain{id: id, accessed: time.Now()}
+	if err := storage.LoadEvents(context.Background(), id, grain); err != nil {
+		t.Fatalf("replay after SaveLoop: %v", err)
+	}
+}
+
+func TestDiskStorageConcurrentAppendSameID(t *testing.T) {
+	reg := diskTestRegistry(t)
+	storage := NewDiskStorage[testState](t.TempDir(), reg)
+	t.Cleanup(storage.Close)
+
+	const id = uint64(88)
+	const workers = 16
+	var wg sync.WaitGroup
+	wg.Add(workers)
+	for i := range workers {
+		go func(itemID uint32) {
+			defer wg.Done()
+			_ = storage.AppendMutations(id, &cart_messages.AddItem{ItemId: itemID, Quantity: 1})
+		}(uint32(i + 1))
+	}
+	wg.Wait()
+
+	grain := &diskTestGrain{id: id, accessed: time.Now()}
+	if err := storage.LoadEvents(context.Background(), id, grain); err != nil {
+		t.Fatalf("replay: %v", err)
+	}
+}
+
+func TestDiskStorageCloseFlushesSaveLoop(t *testing.T) {
+	reg := diskTestRegistry(t)
+	storage := NewDiskStorage[testState](t.TempDir(), reg)
+
+	// Start the buffered SaveLoop but give it a long interval so it does NOT
+	// flush during the test — only Close() should trigger the flush. A brief
+	// sleep lets the goroutine schedule and set s.queue before we append.
+	go storage.SaveLoop(10 * time.Minute)
+	time.Sleep(5 * time.Millisecond)
+
+	const id = uint64(101)
+	if err := storage.AppendMutations(id, &cart_messages.AddItem{ItemId: 7, Quantity: 3}); err != nil {
+		t.Fatal(err)
+	}
+
+	// Verify event is queued but NOT yet on disk (no writer should be open
+	// because the buffered path only opens files inside SaveLoop → save()).
+	if got := storage.openWriterCount(); got != 0 {
+		t.Fatalf("expected 0 open writers before flush, got %d", got)
+	}
+
+	// Close flushes the queue then closes writers.
+	storage.Close()
+
+	// After Close(), the event should be persisted. Open a fresh storage and
+	// replay to verify.
+	fresh := NewDiskStorage[testState](storage.path, reg)
+	t.Cleanup(fresh.Close)
+	grain := &diskTestGrain{id: id, accessed: time.Now()}
+	if err := fresh.LoadEvents(context.Background(), id, grain); err != nil {
+		t.Fatalf("replay after Close flush: %v", err)
+	}
+}
+
+func TestDiskStorageMultipleGrainsIndependent(t *testing.T) {
+	reg := diskTestRegistry(t)
+	storage := NewDiskStorage[testState](t.TempDir(), reg)
+	t.Cleanup(storage.Close)
+
+	ids := []uint64{10, 20, 30}
+	for i, id := range ids {
+		evt := &cart_messages.AddItem{ItemId: uint32(i + 1), Quantity: int32(i + 1)}
+		if err := storage.AppendMutations(id, evt); err != nil {
+			t.Fatalf("append id=%d: %v", id, err)
+		}
+	}
+
+	// Verify each grain has exactly one writer open (shared per id).
+	if got := storage.openWriterCount(); got != len(ids) {
+		t.Fatalf("open writers = %d, want %d", got, len(ids))
+	}
+
+	// Replay each grain and verify the correct event was stored.
+	for _, id := range ids {
+		ctx := context.Background()
+		grain := &diskTestGrain{id: id, accessed: time.Now()}
+		if err := storage.LoadEvents(ctx, id, grain); err != nil {
+			t.Fatalf("replay id=%d: %v", id, err)
+		}
+	}
+}
+
+func TestDiskStorageLoadEventsNoFile(t *testing.T) {
+	reg := diskTestRegistry(t)
+	storage := NewDiskStorage[testState](t.TempDir(), reg)
+	t.Cleanup(storage.Close)
+
+	// No AppendMutations — no file should exist.
+	grain := &diskTestGrain{id: 999, accessed: time.Now()}
+	if err := storage.LoadEvents(context.Background(), 999, grain); err != nil {
+		t.Fatalf("LoadEvents on non-existent file: %v", err)
+	}
+}
+
+func TestDiskStorageWriteAfterIdlePurge(t *testing.T) {
+	reg := diskTestRegistry(t)
+	storage := NewDiskStorage[testState](t.TempDir(), reg)
+	t.Cleanup(storage.Close)
+
+	// Set a very short idle TTL so the writer is evicted quickly.
+	storage.writerIdleTTL = 10 * time.Millisecond
+
+	const id = uint64(202)
+	msg := &cart_messages.AddItem{ItemId: 1, Quantity: 1}
+
+	// First write opens a writer.
+	if err := storage.AppendMutations(id, msg); err != nil {
+		t.Fatal(err)
+	}
+	if got := storage.openWriterCount(); got != 1 {
+		t.Fatalf("open writers after first write = %d, want 1", got)
+	}
+
+	// Wait for idle TTL to expire and purge.
+	time.Sleep(20 * time.Millisecond)
+	storage.purgeIdleWriters()
+
+	if got := storage.openWriterCount(); got != 0 {
+		t.Fatalf("open writers after purge = %d, want 0", got)
+	}
+
+	// Second write must create a new writer through the retry loop in getWriter.
+	if err := storage.AppendMutations(id, msg); err != nil {
+		t.Fatalf("write after purge: %v", err)
+	}
+	if got := storage.openWriterCount(); got != 1 {
+		t.Fatalf("open writers after write-before-close = %d, want 1", got)
+	}
+
+	// Replay must see both events.
+	grain := &diskTestGrain{id: id, accessed: time.Now()}
+	if err := storage.LoadEvents(context.Background(), id, grain); err != nil {
+		t.Fatalf("replay after purge: %v", err)
+	}
+}
+
+func TestDiskStorageCloseWithoutSaveLoop(t *testing.T) {
+	reg := diskTestRegistry(t)
+	storage := NewDiskStorage[testState](t.TempDir(), reg)
+
+	// Verify Close() on a storage that never started SaveLoop works cleanly.
+	const id = uint64(303)
+	if err := storage.AppendMutations(id, &cart_messages.AddItem{ItemId: 1, Quantity: 1}); err != nil {
+		t.Fatal(err)
+	}
+
+	// Close() — queue is nil so it skips save(), goes straight to closeWriters().
+	storage.Close()
+
+	// A fresh storage should be able to replay the persisted event.
+	fresh := NewDiskStorage[testState](storage.path, reg)
+	t.Cleanup(fresh.Close)
+	grain := &diskTestGrain{id: id, accessed: time.Now()}
+	if err := fresh.LoadEvents(context.Background(), id, grain); err != nil {
+		t.Fatalf("replay after close: %v", err)
+	}
+}
+
+// TestDiskStorageConcurrentWritesDifferentIDs verifies that concurrent writes to
+// different grains don't interfere with each other, and all events are persisted.
+func TestDiskStorageConcurrentWritesDifferentIDs(t *testing.T) {
+	reg := diskTestRegistry(t)
+	storage := NewDiskStorage[testState](t.TempDir(), reg)
+	t.Cleanup(storage.Close)
+
+	const grainCount = 32
+	var wg sync.WaitGroup
+	wg.Add(grainCount)
+	for id := range uint64(grainCount) {
+		go func(grainID uint64) {
+			defer wg.Done()
+			for i := range 5 {
+				_ = storage.AppendMutations(grainID, &cart_messages.AddItem{ItemId: uint32(i + 1), Quantity: 1})
+			}
+		}(id + 1)
+	}
+	wg.Wait()
+
+	// Replay each grain and verify.
+	for id := range uint64(grainCount) {
+		grainID := id + 1
+		ctx := context.Background()
+		grain := &diskTestGrain{id: grainID, accessed: time.Now()}
+		if err := storage.LoadEvents(ctx, grainID, grain); err != nil {
+			t.Fatalf("replay grain %d: %v", grainID, err)
+		}
+	}
+}

pkg/actor/simple_grain_pool.go

@@ -366,10 +366,9 @@ func (p *SimpleGrainPool[V]) forAllHosts(fn func(Host[V])) {
 
 	wg := sync.WaitGroup{}
 	for _, host := range rh {
-
 		wg.Go(func() { fn(host) })
-
 	}
+	wg.Wait()
 
 	for name, host := range rh {
 		if !host.IsHealthy() {

pkg/actor/simple_grain_pool_test.go

@@ -0,0 +1,311 @@
+package actor
+
+import (
+	"context"
+	"fmt"
+	"io"
+	"net/http"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"google.golang.org/protobuf/proto"
+)
+
+type testState struct {
+	Value int `json:"value"`
+}
+
+type testGrain struct {
+	id       uint64
+	accessed time.Time
+}
+
+func (g *testGrain) GetId() uint64                      { return g.id }
+func (g *testGrain) GetLastAccess() time.Time           { return g.accessed }
+func (g *testGrain) GetLastChange() time.Time           { return g.accessed }
+func (g *testGrain) GetCurrentState() (*testState, error) {
+	return &testState{Value: int(g.id)}, nil
+}
+
+type mockHost struct {
+	name        string
+	healthy     bool
+	closed      atomic.Bool
+	closeCalls  atomic.Int32
+	actorIDs    []uint64
+	negotiateFn func([]string) ([]string, error)
+	mu          sync.Mutex
+	ownership   []uint64
+	expiry      []uint64
+}
+
+func newMockHost(name string) *mockHost {
+	return &mockHost{name: name, healthy: true}
+}
+
+func (m *mockHost) AnnounceExpiry(ids []uint64) {
+	m.mu.Lock()
+	m.expiry = append(m.expiry, ids...)
+	m.mu.Unlock()
+}
+
+func (m *mockHost) Negotiate(otherHosts []string) ([]string, error) {
+	if m.negotiateFn != nil {
+		return m.negotiateFn(otherHosts)
+	}
+	return otherHosts, nil
+}
+
+func (m *mockHost) Name() string { return m.name }
+
+func (m *mockHost) Proxy(_ uint64, _ http.ResponseWriter, _ *http.Request, _ io.Reader) (bool, error) {
+	return false, fmt.Errorf("not implemented")
+}
+
+func (m *mockHost) Apply(_ context.Context, id uint64, _ ...proto.Message) (*MutationResult[testState], error) {
+	return &MutationResult[testState]{Result: testState{Value: int(id)}}, nil
+}
+
+func (m *mockHost) Get(_ context.Context, id uint64) (*testState, error) {
+	return &testState{Value: int(id)}, nil
+}
+
+func (m *mockHost) GetActorIds() []uint64 { return m.actorIDs }
+
+func (m *mockHost) Close() error {
+	m.closeCalls.Add(1)
+	m.closed.Store(true)
+	return nil
+}
+
+func (m *mockHost) Ping() bool { return m.healthy }
+
+func (m *mockHost) IsHealthy() bool { return m.healthy }
+
+func (m *mockHost) AnnounceOwnership(_ string, ids []uint64) {
+	m.mu.Lock()
+	m.ownership = append(m.ownership, ids...)
+	m.mu.Unlock()
+}
+
+func newTestPool(t *testing.T, spawnHost func(string) (Host[testState], error)) *SimpleGrainPool[testState] {
+	t.Helper()
+	reg := NewMutationRegistry()
+	pool, err := NewSimpleGrainPool(GrainPoolConfig[testState]{
+		Hostname: "local",
+		Spawn: func(_ context.Context, id uint64) (Grain[testState], error) {
+			return &testGrain{id: id, accessed: time.Now()}, nil
+		},
+		SpawnHost:        spawnHost,
+		TTL:              time.Hour,
+		PoolSize:         100,
+		MutationRegistry: reg,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(pool.Close)
+	return pool
+}
+
+func TestSimpleGrainPoolGetAndApply(t *testing.T) {
+	pool := newTestPool(t, func(string) (Host[testState], error) {
+		return nil, fmt.Errorf("no remotes")
+	})
+
+	got, err := pool.Get(context.Background(), 42)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got.Value != 42 {
+		t.Fatalf("Get value = %d, want 42", got.Value)
+	}
+
+	ids := pool.GetLocalIds()
+	if len(ids) != 1 || ids[0] != 42 {
+		t.Fatalf("GetLocalIds = %v, want [42]", ids)
+	}
+}
+
+func TestSimpleGrainPoolAddRemoteDedup(t *testing.T) {
+	var spawnCalls atomic.Int32
+	host := newMockHost("peer-a")
+
+	pool := newTestPool(t, func(name string) (Host[testState], error) {
+		spawnCalls.Add(1)
+		if name != host.name {
+			t.Fatalf("spawn host %q, want %q", name, host.name)
+		}
+		return host, nil
+	})
+
+	const workers = 8
+	var wg sync.WaitGroup
+	wg.Add(workers)
+	for range workers {
+		go func() {
+			defer wg.Done()
+			if _, err := pool.AddRemote(host.name); err != nil {
+				t.Error(err)
+			}
+		}()
+	}
+	wg.Wait()
+
+	if pool.RemoteCount() != 1 {
+		t.Fatalf("RemoteCount = %d, want 1", pool.RemoteCount())
+	}
+	if spawnCalls.Load() != 1 {
+		t.Fatalf("spawn calls = %d, want 1", spawnCalls.Load())
+	}
+	if host.closeCalls.Load() != 0 {
+		t.Fatalf("duplicate remote closed = %d, want 0", host.closeCalls.Load())
+	}
+}
+
+func TestSimpleGrainPoolAddRemoteRaceClosesDuplicate(t *testing.T) {
+	start := make(chan struct{})
+	var spawnCalls atomic.Int32
+	var spawned []*mockHost
+
+	pool := newTestPool(t, func(name string) (Host[testState], error) {
+		spawnCalls.Add(1)
+		h := newMockHost(name)
+		spawned = append(spawned, h)
+		if spawnCalls.Load() == 1 {
+			<-start
+		}
+		return h, nil
+	})
+
+	go func() {
+		if _, err := pool.AddRemote("peer-b"); err != nil {
+			t.Error(err)
+		}
+	}()
+	time.Sleep(20 * time.Millisecond)
+	close(start)
+	if _, err := pool.AddRemote("peer-b"); err != nil {
+		t.Fatal(err)
+	}
+
+	if pool.RemoteCount() != 1 {
+		t.Fatalf("RemoteCount = %d, want 1", pool.RemoteCount())
+	}
+	if spawnCalls.Load() != 2 {
+		t.Fatalf("spawn calls = %d, want 2", spawnCalls.Load())
+	}
+
+	closed := 0
+	deadline := time.After(2 * time.Second)
+	for closed < 1 {
+		closed = 0
+		for _, h := range spawned {
+			if h.closeCalls.Load() > 0 {
+				closed++
+			}
+		}
+		if closed >= 1 {
+			break
+		}
+		select {
+		case <-deadline:
+			t.Fatal("timed out waiting for duplicate remote host Close")
+		default:
+			time.Sleep(10 * time.Millisecond)
+		}
+	}
+}
+
+func TestSimpleGrainPoolRemoveHost(t *testing.T) {
+	host := newMockHost("peer-c")
+	pool := newTestPool(t, func(string) (Host[testState], error) { return host, nil })
+
+	if _, err := pool.AddRemote(host.name); err != nil {
+		t.Fatal(err)
+	}
+	pool.remoteMu.Lock()
+	pool.remoteOwners[99] = host
+	pool.remoteMu.Unlock()
+
+	pool.RemoveHost(host.name)
+
+	if pool.RemoteCount() != 0 {
+		t.Fatalf("RemoteCount = %d, want 0", pool.RemoteCount())
+	}
+	if owner, ok := pool.OwnerHost(99); ok || owner != nil {
+		t.Fatalf("OwnerHost still mapped: %#v ok=%v", owner, ok)
+	}
+	if host.closeCalls.Load() != 1 {
+		t.Fatalf("Close calls = %d, want 1", host.closeCalls.Load())
+	}
+}
+
+func TestSimpleGrainPoolHandleOwnershipChange(t *testing.T) {
+	host := newMockHost("peer-d")
+	pool := newTestPool(t, func(string) (Host[testState], error) { return host, nil })
+
+	pool.localMu.Lock()
+	pool.grains[7] = &testGrain{id: 7, accessed: time.Now()}
+	pool.localMu.Unlock()
+
+	if err := pool.HandleOwnershipChange(host.name, []uint64{7}); err != nil {
+		t.Fatal(err)
+	}
+
+	pool.localMu.RLock()
+	_, local := pool.grains[7]
+	pool.localMu.RUnlock()
+	if local {
+		t.Fatal("local grain should be removed after ownership change")
+	}
+
+	owner, ok := pool.OwnerHost(7)
+	if !ok || owner.Name() != host.name {
+		t.Fatalf("OwnerHost = (%v, %v), want (%q, true)", owner, ok, host.name)
+	}
+}
+
+func TestSimpleGrainPoolPurgeEvictsStaleGrains(t *testing.T) {
+	pool := newTestPool(t, func(string) (Host[testState], error) {
+		return nil, fmt.Errorf("no remotes")
+	})
+	pool.ttl = time.Minute
+
+	stale := &testGrain{id: 1, accessed: time.Now().Add(-2 * time.Minute)}
+	fresh := &testGrain{id: 2, accessed: time.Now()}
+
+	pool.localMu.Lock()
+	pool.grains[1] = stale
+	pool.grains[2] = fresh
+	pool.localMu.Unlock()
+
+	pool.purge()
+
+	pool.localMu.RLock()
+	defer pool.localMu.RUnlock()
+	if _, ok := pool.grains[1]; ok {
+		t.Fatal("stale grain should be purged")
+	}
+	if _, ok := pool.grains[2]; !ok {
+		t.Fatal("fresh grain should remain")
+	}
+}
+
+func TestSimpleGrainPoolBroadcastOwnership(t *testing.T) {
+	host := newMockHost("peer-e")
+	pool := newTestPool(t, func(string) (Host[testState], error) { return host, nil })
+	if _, err := pool.AddRemote(host.name); err != nil {
+		t.Fatal(err)
+	}
+
+	pool.broadcastOwnership([]uint64{11, 12})
+
+	host.mu.Lock()
+	defer host.mu.Unlock()
+	if len(host.ownership) != 2 {
+		t.Fatalf("ownership announces = %v, want 2 ids", host.ownership)
+	}
+}

pkg/proxy/remotehost.go

@@ -9,6 +9,7 @@ import (
 	"io"
 	"log"
 	"net/http"
+	"sync"
 	"time"
 
 	"git.k6n.net/mats/go-cart-actor/pkg/actor"
@@ -27,7 +28,6 @@ type RemoteHost[V any] struct {
 	host          string
 	httpBase      string
 	conn          *grpc.ClientConn
-	transport     *http.Transport
 	client        *http.Client
 	controlClient messages.ControlPlaneClient
 	missedPings   int
@@ -38,8 +38,27 @@ const name = "proxy"
 var (
 	tracer = otel.Tracer(name)
 	meter  = otel.Meter(name)
+
+	// One process-wide HTTP transport pools idle connections to every peer.
+	// Per-host RemoteHost instances previously each owned a transport with
+	// MaxIdleConnsPerHost=100, multiplying file handles by replica count.
+	sharedTransport     *http.Transport
+	sharedTransportOnce sync.Once
 )
 
+func sharedHTTPTransport() *http.Transport {
+	sharedTransportOnce.Do(func() {
+		sharedTransport = &http.Transport{
+			MaxIdleConns:        64,
+			MaxIdleConnsPerHost: 8,
+			MaxConnsPerHost:     32,
+			DisableKeepAlives:   false,
+			IdleConnTimeout:     90 * time.Second,
+		}
+	})
+	return sharedTransport
+}
+
 // MockResponseWriter implements http.ResponseWriter to capture responses for proxy calls.
 type MockResponseWriter struct {
 	StatusCode int
@@ -88,23 +107,13 @@ func NewRemoteHost[V any](host string, httpPort ...string) (*RemoteHost[V], erro
 
 	controlClient := messages.NewControlPlaneClient(conn)
 
-	transport := &http.Transport{
-		MaxIdleConns:        100,
-		MaxIdleConnsPerHost: 100,
-		// Bound concurrently-open connections to this peer so a load spike
-		// can't exhaust file handles. Excess requests wait for a free conn
-		// rather than opening unbounded sockets.
-		MaxConnsPerHost:   256,
-		DisableKeepAlives: false,
-		IdleConnTimeout:   120 * time.Second,
-	}
+	transport := sharedHTTPTransport()
 	client := &http.Client{Transport: transport, Timeout: 10 * time.Second}
 
 	return &RemoteHost[V]{
 		host:          host,
 		httpBase:      fmt.Sprintf("http://%s:%s", host, port),
 		conn:          conn,
-		transport:     transport,
 		client:        client,
 		controlClient: controlClient,
 		missedPings:   0,
@@ -136,7 +145,9 @@ func (h *RemoteHost[V]) Name() string {
 
 func (h *RemoteHost[V]) Close() error {
 	if h.conn != nil {
-		h.conn.Close()
+		err := h.conn.Close()
+		h.conn = nil
+		return err
 	}
 	return nil
 }

pkg/proxy/remotehost_test.go

@@ -0,0 +1,109 @@
+package proxy
+
+import (
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"sync/atomic"
+	"testing"
+)
+
+func TestRemoteHostSharesHTTPTransport(t *testing.T) {
+	h1, err := NewRemoteHost[struct{}]("10.0.0.1", "8080")
+	if err != nil {
+		t.Fatal(err)
+	}
+	h2, err := NewRemoteHost[struct{}]("10.0.0.2", "8080")
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() {
+		_ = h1.Close()
+		_ = h2.Close()
+	})
+
+	if h1.client.Transport != h2.client.Transport {
+		t.Fatal("expected all RemoteHost instances to share one http.Transport")
+	}
+}
+
+func TestRemoteHostProxyUsesKeepAlive(t *testing.T) {
+	var requests atomic.Int32
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		requests.Add(1)
+		w.WriteHeader(http.StatusOK)
+		_, _ = io.WriteString(w, `{"ok":true}`)
+	}))
+	t.Cleanup(srv.Close)
+
+	host, port, err := splitHostPort(srv.URL)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	remote, err := NewRemoteHost[struct{}](host, port)
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { _ = remote.Close() })
+
+	for range 3 {
+		mw := NewMockResponseWriter()
+		req, err := http.NewRequest(http.MethodGet, "/cart/1", nil)
+		if err != nil {
+			t.Fatal(err)
+		}
+		ok, err := remote.Proxy(1, mw, req, nil)
+		if err != nil || !ok {
+			t.Fatalf("Proxy failed: ok=%v err=%v status=%d body=%q", ok, err, mw.StatusCode, mw.Body.String())
+		}
+	}
+
+	if requests.Load() != 3 {
+		t.Fatalf("server requests = %d, want 3", requests.Load())
+	}
+}
+
+func TestRemoteHostCloseDoesNotBreakSharedTransport(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.WriteHeader(http.StatusOK)
+	}))
+	t.Cleanup(srv.Close)
+
+	host, port, err := splitHostPort(srv.URL)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	h1, err := NewRemoteHost[struct{}](host, port)
+	if err != nil {
+		t.Fatal(err)
+	}
+	h2, err := NewRemoteHost[struct{}](host, port)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	if err := h1.Close(); err != nil {
+		t.Fatal(err)
+	}
+
+	mw := NewMockResponseWriter()
+	req, err := http.NewRequest(http.MethodGet, "/health", nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	ok, err := h2.Proxy(0, mw, req, nil)
+	if err != nil || !ok || mw.StatusCode != http.StatusOK {
+		t.Fatalf("second host proxy after first close failed: ok=%v err=%v status=%d", ok, err, mw.StatusCode)
+	}
+}
+
+func splitHostPort(rawURL string) (host, port string, err error) {
+	u, err := url.Parse(rawURL)
+	if err != nil {
+		return "", "", err
+	}
+	return u.Hostname(), u.Port(), nil
+}