Manager paradigm change to an event-loop concurrency style, begun implementing batching in binance futures_websocket.go

services/data_service/cmd/data_service/main.go

@@ -17,10 +17,10 @@ import (
 func main() {
 	fmt.Println("Starting Data Service...")
 	// Setup
-	r := router.NewRouter()
+	r := router.NewRouter(2048)
 	m := manager.NewManager(r)
 	binanceFutures := binance.NewFuturesWebsocket()
-	m.AddProvider("binance_futures_websocket", binanceFutures)
+	_ = m.AddProvider("binance_futures_websocket", binanceFutures)
 
 	// gRPC Control Server
 	grpcControlServer := grpc.NewServer()

services/data_service/internal/manager/helper.go

@@ -0,0 +1,100 @@
+package manager
+
+import (
+	"fmt"
+
+	"gitlab.michelsen.id/phillmichelsen/tessera/services/data_service/internal/domain"
+	"gitlab.michelsen.id/phillmichelsen/tessera/services/data_service/internal/provider"
+)
+
+// Lightweight error helper to define package-level errors inline.
+type constErr string
+
+func (e constErr) Error() string { return string(e) }
+func errorf(s string) error      { return constErr(s) }
+
+// copySet copies a set of identifiers to a new map.
+func copySet(in map[domain.Identifier]struct{}) map[domain.Identifier]struct{} {
+	out := make(map[domain.Identifier]struct{}, len(in))
+	for k := range in {
+		out[k] = struct{}{}
+	}
+	return out
+}
+
+// identifierSetDifferences computes additions and deletions from old -> next.
+func identifierSetDifferences(old map[domain.Identifier]struct{}, next []domain.Identifier) (toAdd, toDel []domain.Identifier) {
+	newSet := make(map[domain.Identifier]struct{}, len(next))
+	for _, id := range next {
+		newSet[id] = struct{}{}
+		if _, ok := old[id]; !ok {
+			toAdd = append(toAdd, id)
+		}
+	}
+	for id := range old {
+		if _, ok := newSet[id]; !ok {
+			toDel = append(toDel, id)
+		}
+	}
+	return
+}
+
+// joinErrors aggregates multiple errors.
+type joined struct{ es []error }
+
+func (j joined) Error() string {
+	switch n := len(j.es); {
+	case n == 0:
+		return ""
+	case n == 1:
+		return j.es[0].Error()
+	default:
+		s := j.es[0].Error()
+		for i := 1; i < n; i++ {
+			s += "; " + j.es[i].Error()
+		}
+		return s
+	}
+}
+
+func join(es []error) error {
+	if len(es) == 0 {
+		return nil
+	}
+	return joined{es}
+}
+
+// resolveProvider parses a raw identifier and looks up the provider.
+func (m *Manager) resolveProvider(id domain.Identifier) (provider.Provider, string, error) {
+	provName, subj, ok := id.ProviderSubject()
+	if !ok || provName == "" || subj == "" {
+		return nil, "", ErrInvalidIdentifier
+	}
+	p := m.providers[provName]
+	if p == nil {
+		return nil, "", fmt.Errorf("%w: %s", ErrUnknownProvider, provName)
+	}
+	return p, subj, nil
+}
+
+// incrementStreamRefCount increments refcount and returns true if transitioning 0->1.
+func (m *Manager) incrementStreamRefCount(id domain.Identifier) bool {
+	rc := m.streamRef[id] + 1
+	m.streamRef[id] = rc
+	return rc == 1
+}
+
+// decrementStreamRefCount decrements refcount and returns true if transitioning 1->0.
+func (m *Manager) decrementStreamRefCount(id domain.Identifier) bool {
+	rc, ok := m.streamRef[id]
+	if !ok {
+		return false
+	}
+	rc--
+	if rc <= 0 {
+		delete(m.streamRef, id)
+		return true
+	}
+	m.streamRef[id] = rc
+	return false
+}

services/data_service/internal/manager/manager.go

@@ -1,10 +1,7 @@
 package manager
 
 import (
-	"context"
-	"errors"
 	"fmt"
-	"sync"
 	"time"
 
 	"github.com/google/uuid"
@@ -13,512 +10,376 @@ import (
 	"gitlab.michelsen.id/phillmichelsen/tessera/services/data_service/internal/router"
 )
 
-var (
+// Manager is a single-goroutine actor that owns all state.
-	ErrSessionNotFound    = errors.New("session not found")
-	ErrSessionClosed      = errors.New("session closed")
-	ErrInvalidIdentifier  = errors.New("invalid identifier")
-	ErrUnknownProvider    = errors.New("unknown provider")
-	ErrClientAlreadyBound = errors.New("client channels already bound")
-)
-
-const (
-	defaultInternalBuf = 1024
-	defaultClientBuf   = 256
-)
-
-type ChannelOpts struct {
-	InBufSize    int
-	OutBufSize   int
-	DropOutbound bool // If true, drop outbound to client when its buffer is full. If false, block.
-	DropInbound  bool // If true, drop inbound from client when internalIn is full. If false, block.
-}
-
-// Manager owns providers, sessions, and the router fanout.
 type Manager struct {
+	// Command channel
+	cmdCh chan any
+
+	// State (loop-owned)
 	providers map[string]provider.Provider
-	providerStreams   map[domain.Identifier]chan domain.Message
-	rawReferenceCount map[domain.Identifier]int
-
 	sessions  map[uuid.UUID]*session
+	streamRef map[domain.Identifier]int
 
+	// Router
 	router *router.Router
-	mu     sync.Mutex
 }
 
-type session struct {
+// New creates a manager and starts its run loop.
-	id uuid.UUID
+func New(r *router.Router) *Manager {
-
+	m := &Manager{
-	// Stable internal channels.
+		cmdCh:     make(chan any, 256),
-	internalIn  chan domain.Message // forwarded into router.IncomingChannel()
-	internalOut chan domain.Message // registered as router route target, forwarded to clientOut (or dropped if unattached)
-
-	// Client Channels (optional). Created on GetChannels and cleared on DetachClient.
-	clientIn  chan domain.Message // caller writes
-	clientOut chan domain.Message // caller reads
-
-	// Controls the permanent internalIn forwarder.
-	cancelInternal context.CancelFunc
-
-	// Permanent outbound drain control.
-	egressWG sync.WaitGroup
-
-	// Policy
-	dropWhenUnattached bool // always drop when no client attached
-	dropWhenSlow       bool // mirror ChannelOpts.DropOutbound
-	dropInbound        bool // mirror ChannelOpts.DropInbound
-
-	bound     map[domain.Identifier]struct{} // map for quick existence checks
-	closed    bool
-	idleAfter time.Duration
-	idleTimer *time.Timer
-}
-
-func NewManager(r *router.Router) *Manager {
-	go r.Run()
-	return &Manager{
 		providers: make(map[string]provider.Provider),
-		providerStreams:   make(map[domain.Identifier]chan domain.Message),
-		rawReferenceCount: make(map[domain.Identifier]int),
 		sessions:  make(map[uuid.UUID]*session),
+		streamRef: make(map[domain.Identifier]int),
 		router:    r,
 	}
+	go r.Run()
+	go m.run()
+	return m
 }
 
-// NewSession creates a session with stable internal channels and two permanent workers:
+// Public API (posts commands to loop)
-// 1) internalIn -> router.Incoming  2) internalOut -> clientOut (or discard if unattached)
+
+// AddProvider adds and starts a new provider.
+func (m *Manager) AddProvider(name string, p provider.Provider) error {
+	resp := make(chan error, 1)
+	m.cmdCh <- addProviderCmd{name: name, p: p, resp: resp}
+	return <-resp
+}
+
+// RemoveProvider stops and removes a provider, cleaning up all sessions.
+func (m *Manager) RemoveProvider(name string) error {
+	resp := make(chan error, 1)
+	m.cmdCh <- removeProviderCmd{name: name, resp: resp}
+	return <-resp
+}
+
+// NewSession creates a new session with the given idle timeout.
 func (m *Manager) NewSession(idleAfter time.Duration) (uuid.UUID, error) {
+	resp := make(chan struct {
+		id  uuid.UUID
+		err error
+	}, 1)
+	m.cmdCh <- newSessionCmd{idleAfter: idleAfter, resp: resp}
+	r := <-resp
+	return r.id, r.err
+}
+
+// AttachClient attaches a client to a session, creates and returns client channels for the session.
+func (m *Manager) AttachClient(id uuid.UUID, inBuf, outBuf int) (chan<- domain.Message, <-chan domain.Message, error) {
+	resp := make(chan struct {
+		cin  chan<- domain.Message
+		cout <-chan domain.Message
+		err  error
+	}, 1)
+	m.cmdCh <- attachCmd{sid: id, inBuf: inBuf, outBuf: outBuf, resp: resp}
+	r := <-resp
+	return r.cin, r.cout, r.err
+}
+
+// DetachClient detaches the client from the session, closes client channels and arms timeout.
+func (m *Manager) DetachClient(id uuid.UUID) error {
+	resp := make(chan error, 1)
+	m.cmdCh <- detachCmd{sid: id, resp: resp}
+	return <-resp
+}
+
+// ConfigureSession sets the next set of identifiers for the session, starting and stopping streams as needed.
+func (m *Manager) ConfigureSession(id uuid.UUID, next []domain.Identifier) error {
+	resp := make(chan error, 1)
+	m.cmdCh <- configureCmd{sid: id, next: next, resp: resp}
+	return <-resp
+}
+
+// CloseSession closes and removes the session, cleaning up all bindings.
+func (m *Manager) CloseSession(id uuid.UUID) error {
+	resp := make(chan error, 1)
+	m.cmdCh <- closeSessionCmd{sid: id, resp: resp}
+	return <-resp
+}
+
+// The main loop of the manager, processing commands serially.
+func (m *Manager) run() {
+	for {
+		msg := <-m.cmdCh
+		switch c := msg.(type) {
+		case addProviderCmd:
+			m.handleAddProvider(c)
+		case removeProviderCmd:
+			m.handleRemoveProvider(c)
+		case newSessionCmd:
+			m.handleNewSession(c)
+		case attachCmd:
+			m.handleAttach(c)
+		case detachCmd:
+			m.handleDetach(c)
+		case configureCmd:
+			m.handleConfigure(c)
+		case closeSessionCmd:
+			m.handleCloseSession(c)
+		}
+	}
+}
+
+// Command handlers, run in loop goroutine. With a single goroutine, no locking is needed.
+
+func (m *Manager) handleAddProvider(cmd addProviderCmd) {
+	if _, ok := m.providers[cmd.name]; ok {
+		cmd.resp <- fmt.Errorf("provider exists: %s", cmd.name)
+		return
+	}
+	if err := cmd.p.Start(); err != nil {
+		cmd.resp <- fmt.Errorf("start provider %s: %w", cmd.name, err)
+		return
+	}
+	m.providers[cmd.name] = cmd.p
+	cmd.resp <- nil
+}
+
+func (m *Manager) handleRemoveProvider(cmd removeProviderCmd) {
+	p, ok := m.providers[cmd.name]
+	if !ok {
+		cmd.resp <- fmt.Errorf("provider not found: %s", cmd.name)
+		return
+	}
+
+	// Clean all identifiers belonging to this provider. Iterates through sessions to reduce provider burden.
+	for _, s := range m.sessions {
+		for ident := range s.bound {
+			provName, subj, ok := ident.ProviderSubject()
+			if !ok || provName != cmd.name {
+				// TODO: add log warning, but basically should never ever happen
+				continue
+			}
+			if s.attached && s.clientOut != nil {
+				m.router.DeregisterRoute(ident, s.clientOut)
+			}
+			delete(s.bound, ident)
+
+			// decrementStreamRefCount returns true if this was the last ref. In which case we want to stop the stream.
+			if ident.IsRaw() && m.decrementStreamRefCount(ident) && subj != "" {
+				_ = p.StopStream(subj) // best-effort as we will remove the provider anyway
+			}
+		}
+	}
+
+	// first iteration above is sound, but as a precaution we also clean up any dangling streamRef entries here
+	for id := range m.streamRef {
+		provName, _, ok := id.ProviderSubject()
+		if !ok || provName != cmd.name {
+			continue
+		}
+		fmt.Printf("manager: warning — dangling streamRef for %s after removing provider %s\n", id.Key(), cmd.name)
+		delete(m.streamRef, id)
+	}
+
+	p.Stop()
+	delete(m.providers, cmd.name)
+	cmd.resp <- nil
+}
+
+func (m *Manager) handleNewSession(cmd newSessionCmd) {
 	s := &session{
 		id:        uuid.New(),
-		internalIn:         make(chan domain.Message, defaultInternalBuf),
-		internalOut:        make(chan domain.Message, defaultInternalBuf),
 		bound:     make(map[domain.Identifier]struct{}),
-		idleAfter:          idleAfter,
+		idleAfter: cmd.idleAfter,
-		dropWhenUnattached: true,
 	}
-	ctx, cancel := context.WithCancel(context.Background())
-	s.cancelInternal = cancel
 
-	m.mu.Lock()
+	// Arm idle timer to auto-close the session.
-	m.sessions[s.id] = s
+	s.idleTimer = time.AfterFunc(cmd.idleAfter, func() {
-	incoming := m.router.IncomingChannel()
+		m.cmdCh <- closeSessionCmd{sid: s.id, resp: make(chan error, 1)}
-	m.mu.Unlock()
+	})
 
-	// Permanent forwarder: internalIn -> router.Incoming
+	m.sessions[s.id] = s // added after arming in the case of immediate timeout or error in arming timer
-	go func(ctx context.Context, in <-chan domain.Message) {
+
-		for {
+	cmd.resp <- struct {
-			select {
+		id  uuid.UUID
-			case <-ctx.Done():
+		err error
-				return
+	}{id: s.id, err: nil}
-			case msg, ok := <-in:
+}
+
+func (m *Manager) handleAttach(cmd attachCmd) {
+	s, ok := m.sessions[cmd.sid]
 	if !ok {
+		cmd.resp <- struct {
+			cin  chan<- domain.Message
+			cout <-chan domain.Message
+			err  error
+		}{nil, nil, ErrSessionNotFound}
 		return
 	}
-				// Hook: filter/validate/meter/throttle inbound to router here.
+	if s.closed {
-				incoming <- msg
+		cmd.resp <- struct {
+			cin  chan<- domain.Message
+			cout <-chan domain.Message
+			err  error
+		}{nil, nil, ErrSessionClosed}
+		return
 	}
-		}
+	if s.attached {
-	}(ctx, s.internalIn)
+		cmd.resp <- struct {
-
+			cin  chan<- domain.Message
-	// Permanent drain: internalOut -> clientOut (drop if unattached)
+			cout <-chan domain.Message
-	s.egressWG.Add(1)
+			err  error
-	go func(sid uuid.UUID) {
+		}{nil, nil, ErrClientAlreadyAttached}
-		defer s.egressWG.Done()
+		return
-		for msg := range s.internalOut {
-			m.mu.Lock()
-			// Session might be gone; re-fetch safely.
-			s, ok := m.sessions[sid]
-			var cout chan domain.Message
-			var dropSlow, attached bool
-			if ok {
-				cout = s.clientOut
-				dropSlow = s.dropWhenSlow
-				attached = cout != nil
-			}
-			m.mu.Unlock()
-
-			switch {
-			case !attached:
-				// unattached => drop
-
-			case dropSlow: // typical case when attached
-				select {
-				case cout <- msg:
-				default:
-					// drop on slow consumer
 	}
 
-			default:
+	cin, cout, err := m.attachSession(s, cmd.inBuf, cmd.outBuf)
-				cout <- msg // push to client, block if slow
-			}
-		}
-	}(s.id)
 
-	return s.id, nil
+	cmd.resp <- struct {
+		cin  chan<- domain.Message
+		cout <-chan domain.Message
+		err  error
+	}{cin, cout, err}
 }
 
-// GetChannels creates a fresh client attachment and hooks inbound (clientIn -> internalIn).
+func (m *Manager) handleDetach(cmd detachCmd) {
-// Outbound delivery is handled by the permanent drain.
+	s, ok := m.sessions[cmd.sid]
-// Only one attachment at a time.
-func (m *Manager) GetChannels(id uuid.UUID, opts ChannelOpts) (chan<- domain.Message, <-chan domain.Message, error) {
-	if opts.InBufSize <= 0 {
-		opts.InBufSize = defaultClientBuf
-	}
-	if opts.OutBufSize <= 0 {
-		opts.OutBufSize = defaultClientBuf
-	}
-
-	m.mu.Lock()
-	s, ok := m.sessions[id]
 	if !ok {
-		m.mu.Unlock()
+		cmd.resp <- ErrSessionNotFound
-		return nil, nil, ErrSessionNotFound
+		return
 	}
 	if s.closed {
-		m.mu.Unlock()
+		cmd.resp <- ErrSessionClosed
-		return nil, nil, ErrSessionClosed
+		return
 	}
-	if s.clientIn != nil || s.clientOut != nil {
+	if !s.attached {
-		m.mu.Unlock()
+		cmd.resp <- ErrClientNotAttached
-		return nil, nil, ErrClientAlreadyBound
+		return
 	}
 
-	// Create attachment channels.
+	_ = m.detachSession(cmd.sid, s)
-	cin := make(chan domain.Message, opts.InBufSize)
-	cout := make(chan domain.Message, opts.OutBufSize)
-	s.clientIn, s.clientOut = cin, cout
-	s.dropWhenSlow = opts.DropOutbound
-	s.dropInbound = opts.DropInbound
 
-	// Stop idle timer while attached.
+	cmd.resp <- nil
-	if s.idleTimer != nil {
-		s.idleTimer.Stop()
-		s.idleTimer = nil
 }
 
-	internalIn := s.internalIn
+func (m *Manager) handleConfigure(c configureCmd) {
-	m.mu.Unlock()
+	s, ok := m.sessions[c.sid]
-
-	// Forward clientIn -> internalIn
-	go func(src <-chan domain.Message, dst chan<- domain.Message, drop bool) {
-		for msg := range src {
-			if drop {
-				select {
-				case dst <- msg:
-				default:
-					// drop inbound on internal backpressure
-				}
-			} else {
-				dst <- msg
-			}
-		}
-		// client closed input; forwarder exits
-	}(cin, internalIn, opts.DropInbound)
-
-	// Return directional views.
-	return (chan<- domain.Message)(cin), (<-chan domain.Message)(cout), nil
-}
-
-// DetachClient clears the client attachment and starts the idle close timer if configured.
-// Does not close clientOut to avoid send-on-closed races with the permanent drain.
-func (m *Manager) DetachClient(id uuid.UUID) error {
-	m.mu.Lock()
-	s, ok := m.sessions[id]
 	if !ok {
-		m.mu.Unlock()
+		c.resp <- ErrSessionNotFound
-		return ErrSessionNotFound
+		return
 	}
 	if s.closed {
-		m.mu.Unlock()
+		c.resp <- ErrSessionClosed
-		return ErrSessionClosed
+		return
-	}
-	cin := s.clientIn
-	// Make unattached; permanent drain will drop while nil.
-	s.clientIn, s.clientOut = nil, nil
-	after := s.idleAfter
-	m.mu.Unlock()
-
-	if cin != nil {
-		// We own the channel. Closing signals writers to stop.
-		close(cin)
 	}
 
-	if after > 0 {
+	old := copySet(s.bound)
-		m.mu.Lock()
+	toAdd, toDel := identifierSetDifferences(old, c.next)
-		ss, ok := m.sessions[id]
-		if ok && !ss.closed && ss.clientOut == nil && ss.idleTimer == nil {
-			ss.idleTimer = time.AfterFunc(after, func() { _ = m.CloseSession(id) })
-		}
-		m.mu.Unlock()
-	}
-	return nil
-}
 
-func (m *Manager) Subscribe(id uuid.UUID, ids ...domain.Identifier) error {
+	// 1) Handle removals first.
-	m.mu.Lock()
+	for _, ident := range toDel {
-	s, ok := m.sessions[id]
+		if s.attached && s.clientOut != nil {
-	if !ok {
+			m.router.DeregisterRoute(ident, s.clientOut)
-		m.mu.Unlock()
-		return ErrSessionNotFound
-	}
-	out := s.internalOut
-	m.mu.Unlock()
-
-	for _, ident := range ids {
-		m.mu.Lock()
-		if _, exists := s.bound[ident]; exists {
-			m.mu.Unlock()
-			continue
-		}
-		s.bound[ident] = struct{}{}
-		m.mu.Unlock()
-
-		if ident.IsRaw() {
-			if err := m.provisionRawStream(ident); err != nil {
-				return err
-			}
-		}
-		m.router.RegisterRoute(ident, out)
-	}
-	return nil
-}
-
-func (m *Manager) Unsubscribe(id uuid.UUID, ids ...domain.Identifier) error {
-	m.mu.Lock()
-	s, ok := m.sessions[id]
-	if !ok {
-		m.mu.Unlock()
-		return ErrSessionNotFound
-	}
-	out := s.internalOut
-	m.mu.Unlock()
-
-	for _, ident := range ids {
-		m.mu.Lock()
-		if _, exists := s.bound[ident]; !exists {
-			m.mu.Unlock()
-			continue
 		}
 		delete(s.bound, ident)
-		m.mu.Unlock()
 
-		m.router.DeregisterRoute(ident, out)
 		if ident.IsRaw() {
-			m.releaseRawStreamIfUnused(ident)
+			if m.decrementStreamRefCount(ident) {
+				if p, subj, err := m.resolveProvider(ident); err == nil {
+					_ = p.StopStream(subj) // fire-and-forget
+				}
 			}
 		}
-	return nil
 	}
 
-func (m *Manager) SetSubscriptions(id uuid.UUID, next []domain.Identifier) error {
+	// 2) Handle additions. Collect starts to await.
-	m.mu.Lock()
+	type startItem struct {
-	s, ok := m.sessions[id]
+		id domain.Identifier
-	if !ok {
+		ch <-chan error
-		m.mu.Unlock()
-		return ErrSessionNotFound
 	}
-	old := make(map[domain.Identifier]struct{}, len(s.bound))
+	var starts []startItem
-	for k := range s.bound {
+	var initErrs []error
-		old[k] = struct{}{}
-	}
-	out := s.internalOut
-	m.mu.Unlock()
-
-	toAdd, toDel := m.identifierSetDifferences(old, next)
 
 	for _, ident := range toAdd {
-		m.mu.Lock()
+		// Bind intent now.
 		s.bound[ident] = struct{}{}
-		m.mu.Unlock()
 
-		if ident.IsRaw() {
+		if !ident.IsRaw() {
-			if err := m.provisionRawStream(ident); err != nil {
+			if s.attached && s.clientOut != nil {
-				return err
+				m.router.RegisterRoute(ident, s.clientOut)
 			}
-		}
+			continue
-		m.router.RegisterRoute(ident, out)
 		}
 
-	for _, ident := range toDel {
+		p, subj, err := m.resolveProvider(ident)
-		m.mu.Lock()
+		if err != nil {
-		_, exists := s.bound[ident]
 			delete(s.bound, ident)
-		m.mu.Unlock()
+			initErrs = append(initErrs, err)
-
+			continue
-		if exists {
-			m.router.DeregisterRoute(ident, out)
-			if ident.IsRaw() {
-				m.releaseRawStreamIfUnused(ident)
 		}
-		}
+		if !p.IsValidSubject(subj, false) {
-	}
+			delete(s.bound, ident)
-	return nil
+			initErrs = append(initErrs, fmt.Errorf("invalid subject %q for provider", subj))
+			continue
 		}
 
-func (m *Manager) CloseSession(id uuid.UUID) error {
+		first := m.incrementStreamRefCount(ident)
-	m.mu.Lock()
-	s, ok := m.sessions[id]
-	if !ok {
-		m.mu.Unlock()
-		return ErrSessionNotFound
-	}
-	if s.closed {
-		m.mu.Unlock()
-		return nil
-	}
-	s.closed = true
-	if s.idleTimer != nil {
-		s.idleTimer.Stop()
-		s.idleTimer = nil
-	}
-	out := s.internalOut
-	ids := make([]domain.Identifier, 0, len(s.bound))
-	for k := range s.bound {
-		ids = append(ids, k)
-	}
-	cancelInternal := s.cancelInternal
-	// Snapshot clientIn/Out for shutdown signals after unlock.
-	cin := s.clientIn
-	cout := s.clientOut
-	// Remove from map before unlock to prevent new work.
-	delete(m.sessions, id)
-	m.mu.Unlock()
 
-	// Deregister all routes and release raw streams.
+		if first || !p.IsStreamActive(subj) {
-	for _, ident := range ids {
+			ch := p.StartStream(subj, m.router.IncomingChannel())
-		m.router.DeregisterRoute(ident, out)
+			starts = append(starts, startItem{id: ident, ch: ch})
-		if ident.IsRaw() {
+		} else if s.attached && s.clientOut != nil {
-			m.releaseRawStreamIfUnused(ident)
+			// Already active, just register for this session.
+			m.router.RegisterRoute(ident, s.clientOut)
 		}
 	}
 
-	// Stop inbound forwarder and close internals.
+	// 3) Wait for starts initiated by this call, each with its own timeout.
-	if cancelInternal != nil {
+	if len(starts) == 0 {
-		cancelInternal()
+		c.resp <- join(initErrs)
-	}
-	close(s.internalIn)  // end internalIn forwarder
-	close(s.internalOut) // signal drain to finish
-
-	// Wait drain exit, then close clientOut if attached at close time.
-	s.egressWG.Wait()
-	if cout != nil {
-		close(cout)
-	}
-	// Close clientIn to stop client writers if still attached.
-	if cin != nil {
-		close(cin)
-	}
-	return nil
-}
-
-func (m *Manager) AddProvider(name string, p provider.Provider) error {
-	m.mu.Lock()
-	if _, exists := m.providers[name]; exists {
-		m.mu.Unlock()
-		return fmt.Errorf("provider exists: %s", name)
-	}
-	m.mu.Unlock()
-
-	if err := p.Start(); err != nil {
-		return fmt.Errorf("start provider %s: %w", name, err)
-	}
-
-	m.mu.Lock()
-	m.providers[name] = p
-	m.mu.Unlock()
-	return nil
-}
-
-func (m *Manager) RemoveProvider(name string) error {
-	m.mu.Lock()
-	_, ok := m.providers[name]
-	m.mu.Unlock()
-	if !ok {
-		return fmt.Errorf("provider not found: %s", name)
-	}
-	// TODO: implement full drain and cancel of all streams for this provider if needed.
-	return fmt.Errorf("RemoveProvider not implemented")
-}
-
-func (m *Manager) provisionRawStream(id domain.Identifier) error {
-	providerName, subject, ok := id.ProviderSubject()
-	if !ok || providerName == "" || subject == "" {
-		return ErrInvalidIdentifier
-	}
-
-	m.mu.Lock()
-	prov, exists := m.providers[providerName]
-	if !exists {
-		m.mu.Unlock()
-		return ErrUnknownProvider
-	}
-	if !prov.IsValidSubject(subject, false) {
-		m.mu.Unlock()
-		return fmt.Errorf("invalid subject %q for provider %s", subject, providerName)
-	}
-
-	if ch, ok := m.providerStreams[id]; ok {
-		m.rawReferenceCount[id] = m.rawReferenceCount[id] + 1
-		m.mu.Unlock()
-		_ = ch
-		return nil
-	}
-
-	ch := make(chan domain.Message, 64)
-	if err := prov.RequestStream(subject, ch); err != nil {
-		m.mu.Unlock()
-		return fmt.Errorf("provision %v: %w", id, err)
-	}
-	m.providerStreams[id] = ch
-	m.rawReferenceCount[id] = 1
-	incoming := m.router.IncomingChannel()
-	m.mu.Unlock()
-
-	// Provider stream -> router.Incoming
-	go func(c chan domain.Message) {
-		for msg := range c {
-			incoming <- msg
-		}
-	}(ch)
-
-	return nil
-}
-
-func (m *Manager) releaseRawStreamIfUnused(id domain.Identifier) {
-	providerName, subject, ok := id.ProviderSubject()
-	if !ok {
 		return
 	}
 
-	m.mu.Lock()
+	type result struct {
-	rc := m.rawReferenceCount[id] - 1
+		id  domain.Identifier
-	if rc <= 0 {
+		err error
-		if ch, ok := m.providerStreams[id]; ok {
-			if prov, exists := m.providers[providerName]; exists {
-				prov.CancelStream(subject)
 	}
-			close(ch)
+	done := make(chan result, len(starts))
-			delete(m.providerStreams, id)
+
+	for _, si := range starts {
+		// Per-start waiter.
+		go func(id domain.Identifier, ch <-chan error) {
+			select {
+			case err := <-ch:
+				done <- result{id: id, err: err}
+			case <-time.After(statusWaitTotal):
+				done <- result{id: id, err: fmt.Errorf("timeout")}
 			}
-		delete(m.rawReferenceCount, id)
+		}(si.id, si.ch)
-		m.mu.Unlock()
-		return
-	}
-	m.rawReferenceCount[id] = rc
-	m.mu.Unlock()
 	}
 
-func (m *Manager) identifierSetDifferences(old map[domain.Identifier]struct{}, next []domain.Identifier) (toAdd, toDel []domain.Identifier) {
+	// Collect results and apply.
-	newSet := make(map[domain.Identifier]struct{}, len(next))
+	for i := 0; i < len(starts); i++ {
-	for _, id := range next {
+		r := <-done
-		newSet[id] = struct{}{}
+		if r.err != nil {
-		if _, ok := old[id]; !ok {
+			// Roll back this session's bind and drop ref.
-			toAdd = append(toAdd, id)
+			delete(s.bound, r.id)
+			_ = m.decrementStreamRefCount(r.id)
+			initErrs = append(initErrs, fmt.Errorf("start %v: %w", r.id, r.err))
+			continue
+		}
+		// Success: register for any attached sessions that are bound.
+		for _, sess := range m.sessions {
+			if !sess.attached || sess.clientOut == nil {
+				continue
+			}
+			if _, bound := sess.bound[r.id]; bound {
+				m.router.RegisterRoute(r.id, sess.clientOut)
 			}
 		}
-	for id := range old {
-		if _, ok := newSet[id]; !ok {
-			toDel = append(toDel, id)
 	}
+
+	c.resp <- join(initErrs)
 }
+
+func (m *Manager) handleCloseSession(c closeSessionCmd) {
+	s, ok := m.sessions[c.sid]
+	if !ok {
+		c.resp <- ErrSessionNotFound
 		return
 	}
+	m.closeSession(c.sid, s)
+	c.resp <- nil
+}

services/data_service/internal/manager/session.go

@@ -0,0 +1,114 @@
+package manager
+
+import (
+	"time"
+
+	"github.com/google/uuid"
+	"gitlab.michelsen.id/phillmichelsen/tessera/services/data_service/internal/domain"
+)
+
+// attachSession wires channels, stops idle timer, and registers ready routes.
+// Precondition: session exists and is not attached/closed. Runs in loop.
+func (m *Manager) attachSession(s *session, inBuf, outBuf int) (chan<- domain.Message, <-chan domain.Message, error) {
+	if inBuf <= 0 {
+		inBuf = defaultClientBuf
+	}
+	if outBuf <= 0 {
+		outBuf = defaultClientBuf
+	}
+
+	cin := make(chan domain.Message, inBuf)
+	cout := make(chan domain.Message, outBuf)
+	s.clientIn, s.clientOut = cin, cout
+
+	if s.idleTimer != nil {
+		s.idleTimer.Stop()
+		s.idleTimer = nil
+	}
+
+	// Forward clientIn to router.Incoming with drop on backpressure.
+	go func(src <-chan domain.Message, dst chan<- domain.Message) {
+		for msg := range src {
+			select {
+			case dst <- msg:
+			default:
+				// drop
+			}
+		}
+	}(cin, m.router.IncomingChannel())
+
+	// Register all currently bound that are ready.
+	for ident := range s.bound {
+		if !ident.IsRaw() {
+			m.router.RegisterRoute(ident, cout)
+			continue
+		}
+		// Raw: register only if provider stream is active.
+		if p, subj, err := m.resolveProvider(ident); err == nil && p.IsStreamActive(subj) {
+			m.router.RegisterRoute(ident, cout)
+		}
+	}
+
+	s.attached = true
+	return cin, cout, nil
+}
+
+// detachSession deregisters all routes, closes channels, and arms idle timer.
+// Precondition: session exists and is attached. Runs in loop.
+func (m *Manager) detachSession(sid uuid.UUID, s *session) error {
+	if s.clientOut != nil {
+		for ident := range s.bound {
+			m.router.DeregisterRoute(ident, s.clientOut)
+		}
+		close(s.clientOut)
+	}
+	if s.clientIn != nil {
+		close(s.clientIn)
+	}
+	s.clientIn, s.clientOut = nil, nil
+	s.attached = false
+
+	// Arm idle timer to auto-close the session.
+	s.idleTimer = time.AfterFunc(s.idleAfter, func() {
+		m.cmdCh <- closeSessionCmd{sid: sid, resp: make(chan error, 1)}
+	})
+	return nil
+}
+
+// closeSession performs full teardown and refcount drops. Runs in loop.
+func (m *Manager) closeSession(sid uuid.UUID, s *session) {
+	if s.closed {
+		return
+	}
+	s.closed = true
+
+	// Detach if attached.
+	if s.attached {
+		if s.clientOut != nil {
+			for ident := range s.bound {
+				m.router.DeregisterRoute(ident, s.clientOut)
+			}
+			close(s.clientOut)
+		}
+		if s.clientIn != nil {
+			close(s.clientIn)
+		}
+	} else if s.idleTimer != nil {
+		s.idleTimer.Stop()
+		s.idleTimer = nil
+	}
+
+	// Drop refs for raw identifiers and stop streams if last ref. Fire-and-forget.
+	for ident := range s.bound {
+		if !ident.IsRaw() {
+			continue
+		}
+		if last := m.decrementStreamRefCount(ident); last {
+			if p, subj, err := m.resolveProvider(ident); err == nil {
+				_ = p.StopStream(subj) // do not wait
+			}
+		}
+	}
+
+	delete(m.sessions, sid)
+}

services/data_service/internal/manager/types.go

@@ -0,0 +1,86 @@
+package manager
+
+import (
+	"time"
+
+	"github.com/google/uuid"
+	"gitlab.michelsen.id/phillmichelsen/tessera/services/data_service/internal/domain"
+	"gitlab.michelsen.id/phillmichelsen/tessera/services/data_service/internal/provider"
+)
+
+// Shared constants.
+const (
+	defaultClientBuf = 256
+	statusWaitTotal  = 8 * time.Second
+)
+
+// Manager-level errors.
+var (
+	ErrSessionNotFound       = errorf("session not found")
+	ErrSessionClosed         = errorf("session closed")
+	ErrClientAlreadyAttached = errorf("client already attached")
+	ErrClientNotAttached     = errorf("client not attached")
+	ErrInvalidIdentifier     = errorf("invalid identifier")
+	ErrUnknownProvider       = errorf("unknown provider")
+)
+
+// Session holds per-session state. Owned by the manager loop.
+type session struct {
+	id uuid.UUID
+
+	clientIn  chan domain.Message // caller writes
+	clientOut chan domain.Message // caller reads
+
+	bound map[domain.Identifier]struct{}
+
+	closed    bool
+	attached  bool
+	idleAfter time.Duration
+	idleTimer *time.Timer
+}
+
+// Commands posted into the manager loop. One struct per action.
+type addProviderCmd struct {
+	name string
+	p    provider.Provider
+	resp chan error
+}
+
+type removeProviderCmd struct {
+	name string
+	resp chan error
+}
+
+type newSessionCmd struct {
+	idleAfter time.Duration
+	resp      chan struct {
+		id  uuid.UUID
+		err error
+	}
+}
+
+type attachCmd struct {
+	sid           uuid.UUID
+	inBuf, outBuf int
+	resp          chan struct {
+		cin  chan<- domain.Message
+		cout <-chan domain.Message
+		err  error
+	}
+}
+
+type detachCmd struct {
+	sid  uuid.UUID
+	resp chan error
+}
+
+type configureCmd struct {
+	sid  uuid.UUID
+	next []domain.Identifier
+	resp chan error // returns after starts from this call succeed or timeout
+}
+
+type closeSessionCmd struct {
+	sid  uuid.UUID
+	resp chan error
+}

services/data_service/internal/provider/binance/futures_websocket.go

@@ -2,139 +2,411 @@ package binance
 
 import (
 	"encoding/json"
+	"errors"
 	"fmt"
+	"net/http"
 	"sync"
+	"sync/atomic"
+	"time"
 
 	"github.com/gorilla/websocket"
 	"gitlab.michelsen.id/phillmichelsen/tessera/services/data_service/internal/domain"
 )
 
+const (
+	wsURL              = "wss://fstream.binance.com/stream"
+	writeRatePerSecond = 8               // hard cap per second
+	writeBurst         = 8               // token bucket burst
+	writeWait          = 5 * time.Second // per write deadline
+
+	batchPeriod = 1 * time.Second // batch SUB/UNSUB every second
+
+	reconnectMin = 500 * time.Millisecond
+	reconnectMax = 10 * time.Second
+)
+
+// internal stream states (provider stays simple; manager relies on IsStreamActive)
+type streamState uint8
+
+const (
+	stateUnknown streamState = iota
+	statePendingSub
+	stateActive
+	statePendingUnsub
+	stateInactive
+	stateError
+)
+
 type FuturesWebsocket struct {
-	conn          *websocket.Conn
+	dial websocket.Dialer
-	activeStreams map[string]chan domain.Message
+	hdr  http.Header
+
+	// desired subscriptions and sinks
 	mu      sync.Mutex
+	desired map[string]bool                  // subject -> want subscribed
+	sinks   map[string]chan<- domain.Message // subject -> destination
+	states  map[string]streamState           // subject -> state
+
+	// waiters per subject
+	startWaiters map[string][]chan error
+	stopWaiters  map[string][]chan error
+
+	// batching queues
+	subQ   chan string
+	unsubQ chan string
+
+	// websocket
+	writeMu sync.Mutex
+	conn    *websocket.Conn
+
+	// rate limit tokens
+	tokensCh chan struct{}
+	stopRate chan struct{}
+
+	// lifecycle
+	stopCh chan struct{}
+	wg     sync.WaitGroup
+
+	// ack tracking
+	ackMu    sync.Mutex
+	idSeq    uint64
+	pendingA map[int64]ackBatch
+}
+
+type ackBatch struct {
+	method   string // "SUBSCRIBE" or "UNSUBSCRIBE"
+	subjects []string
 }
 
 func NewFuturesWebsocket() *FuturesWebsocket {
 	return &FuturesWebsocket{
-		activeStreams: make(map[string]chan domain.Message),
+		desired:      make(map[string]bool),
+		sinks:        make(map[string]chan<- domain.Message),
+		states:       make(map[string]streamState),
+		startWaiters: make(map[string][]chan error),
+		stopWaiters:  make(map[string][]chan error),
+		subQ:         make(chan string, 4096),
+		unsubQ:       make(chan string, 4096),
+		tokensCh:     make(chan struct{}, writeBurst),
+		stopRate:     make(chan struct{}),
+		stopCh:       make(chan struct{}),
+		pendingA:     make(map[int64]ackBatch),
 	}
 }
 
+/* provider.Provider */
+
 func (b *FuturesWebsocket) Start() error {
-	c, _, err := websocket.DefaultDialer.Dial("wss://fstream.binance.com/stream", nil)
+	// token bucket
-	if err != nil {
+	b.wg.Add(1)
-		return fmt.Errorf("connect failed: %w", err)
+	go func() {
+		defer b.wg.Done()
+		t := time.NewTicker(time.Second / writeRatePerSecond)
+		defer t.Stop()
+		// prime burst
+		for i := 0; i < writeBurst; i++ {
+			select {
+			case b.tokensCh <- struct{}{}:
+			default:
 			}
-	b.conn = c
+		}
-	go b.readLoop()
+		for {
+			select {
+			case <-b.stopRate:
+				return
+			case <-t.C:
+				select {
+				case b.tokensCh <- struct{}{}:
+				default:
+				}
+			}
+		}
+	}()
+
+	// connection manager
+	b.wg.Add(1)
+	go b.run()
+
+	// batcher
+	b.wg.Add(1)
+	go b.batcher()
+
 	return nil
 }
 
 func (b *FuturesWebsocket) Stop() {
+	close(b.stopCh)
+	close(b.stopRate)
+
+	b.writeMu.Lock()
 	if b.conn != nil {
-		err := b.conn.Close()
+		_ = b.conn.WriteMessage(websocket.CloseMessage, websocket.FormatCloseMessage(websocket.CloseNormalClosure, "bye"))
-		if err != nil {
+		_ = b.conn.Close()
-			panic(fmt.Errorf("failed to close websocket connection: %w", err))
+		b.conn = nil
-		}
-	}
 	}
+	b.writeMu.Unlock()
 
-func (b *FuturesWebsocket) RequestStream(subject string, ch chan domain.Message) error {
+	b.wg.Wait()
+
+	// resolve any remaining waiters with an error
 	b.mu.Lock()
 	defer b.mu.Unlock()
-
+	for subj, ws := range b.startWaiters {
-	if _, ok := b.activeStreams[subject]; ok {
+		for _, ch := range ws {
-		return nil
+			select {
+			case ch <- errors.New("provider stopped"):
+			default:
+			}
+			close(ch)
+		}
+		delete(b.startWaiters, subj)
+	}
+	for subj, ws := range b.stopWaiters {
+		for _, ch := range ws {
+			select {
+			case ch <- errors.New("provider stopped"):
+			default:
+			}
+			close(ch)
+		}
+		delete(b.stopWaiters, subj)
+	}
 }
 
-	msg := map[string]interface{}{
+func (b *FuturesWebsocket) StartStream(subject string, dst chan<- domain.Message) <-chan error {
-		"method": "SUBSCRIBE",
+	fmt.Println("Starting stream for subject:", subject)
-		"params": []string{subject},
+	ch := make(chan error, 1)
-		"id":     len(b.activeStreams) + 1,
+
-	}
+	if subject == "" {
-	if err := b.conn.WriteJSON(msg); err != nil {
+		ch <- fmt.Errorf("empty subject")
-		return fmt.Errorf("subscribe failed: %w", err)
+		close(ch)
+		return ch
 	}
 
-	b.activeStreams[subject] = ch
-	fmt.Println("Subscribed to stream:", subject)
-	return nil
-}
-
-func (b *FuturesWebsocket) CancelStream(subject string) {
 	b.mu.Lock()
-	defer b.mu.Unlock()
+	// mark desired, update sink
+	b.desired[subject] = true
+	b.sinks[subject] = dst
 
-	if _, ok := b.activeStreams[subject]; !ok {
+	// fast path: already active
-		return
+	if b.states[subject] == stateActive {
+		b.mu.Unlock()
+		ch <- nil
+		close(ch)
+		return ch
 	}
 
-	msg := map[string]interface{}{
+	// enqueue waiter and transition if needed
-		"method": "UNSUBSCRIBE",
+	b.startWaiters[subject] = append(b.startWaiters[subject], ch)
-		"params": []string{subject},
+	if b.states[subject] != statePendingSub {
-		"id":     len(b.activeStreams) + 1000,
+		b.states[subject] = statePendingSub
+		select {
+		case b.subQ <- subject:
+		default:
+			// queue full → fail fast
+			ws := b.startWaiters[subject]
+			delete(b.startWaiters, subject)
+			b.states[subject] = stateError
+			b.mu.Unlock()
+			for _, w := range ws {
+				w <- fmt.Errorf("subscribe queue full")
+				close(w)
 			}
-	_ = b.conn.WriteJSON(msg)
+			return ch
-
+		}
-	fmt.Println("Unsubscribed from stream:", subject)
+	}
-
+	b.mu.Unlock()
-	delete(b.activeStreams, subject)
+	return ch
+}
+
+func (b *FuturesWebsocket) StopStream(subject string) <-chan error {
+	fmt.Println("Stopping stream for subject:", subject)
+	ch := make(chan error, 1)
+
+	if subject == "" {
+		ch <- fmt.Errorf("empty subject")
+		close(ch)
+		return ch
 	}
 
-func (b *FuturesWebsocket) GetActiveStreams() []string {
 	b.mu.Lock()
-	defer b.mu.Unlock()
+	// mark no longer desired; keep sink until UNSUB ack to avoid drops
+	b.desired[subject] = false
 
-	var streams []string
+	// already inactive
-	for k := range b.activeStreams {
+	if b.states[subject] == stateInactive {
-		streams = append(streams, k)
+		b.mu.Unlock()
+		ch <- nil
+		close(ch)
+		return ch
 	}
-	return streams
+
+	// enqueue waiter and transition if needed
+	b.stopWaiters[subject] = append(b.stopWaiters[subject], ch)
+	if b.states[subject] != statePendingUnsub {
+		b.states[subject] = statePendingUnsub
+		select {
+		case b.unsubQ <- subject:
+		default:
+			// queue full → fail fast
+			ws := b.stopWaiters[subject]
+			delete(b.stopWaiters, subject)
+			b.states[subject] = stateError
+			b.mu.Unlock()
+			for _, w := range ws {
+				w <- fmt.Errorf("unsubscribe queue full")
+				close(w)
+			}
+			return ch
+		}
+	}
+	b.mu.Unlock()
+	return ch
+}
+
+func (b *FuturesWebsocket) Fetch(_ string) (domain.Message, error) {
+	return domain.Message{}, fmt.Errorf("fetch not supported")
 }
 
 func (b *FuturesWebsocket) IsStreamActive(subject string) bool {
 	b.mu.Lock()
 	defer b.mu.Unlock()
-
+	return b.states[subject] == stateActive
-	_, ok := b.activeStreams[subject]
-	return ok
-}
-
-func (b *FuturesWebsocket) Fetch(_ string) (domain.Message, error) {
-	return domain.Message{}, fmt.Errorf("not supported: websocket provider does not implement fetch")
 }
 
 func (b *FuturesWebsocket) IsValidSubject(subject string, isFetch bool) bool {
-	if isFetch {
+	return !isFetch && subject != ""
-		return false
-	}
-	return len(subject) > 0
 }
 
-func (b *FuturesWebsocket) readLoop() {
+/* internals */
-	for {
+
-		_, msgBytes, err := b.conn.ReadMessage()
+func (b *FuturesWebsocket) run() {
+	defer b.wg.Done()
+
+	backoff := reconnectMin
+
+	dial := func() (*websocket.Conn, error) {
+		c, _, err := b.dial.Dial(wsURL, b.hdr)
 		if err != nil {
-			if websocket.IsCloseError(err, websocket.CloseNormalClosure) {
+			return nil, err
-				return
 		}
-			fmt.Printf("read error: %v\n", err)
+		return c, nil
+	}
+
+	for {
+		select {
+		case <-b.stopCh:
+			return
+		default:
+		}
+
+		c, err := dial()
+		if err != nil {
+			time.Sleep(backoff)
+			backoff = minDur(backoff*2, reconnectMax)
 			continue
 		}
+		backoff = reconnectMin
 
+		b.writeMu.Lock()
+		b.conn = c
+		b.writeMu.Unlock()
+
+		// Resubscribe desired subjects in one batched SUB.
+		want := b.snapshotDesired(true) // only desired==true
+		if len(want) > 0 {
+			_ = b.sendSubscribe(want)
+			b.mu.Lock()
+			for _, s := range want {
+				if b.states[s] != stateActive {
+					b.states[s] = statePendingSub
+				}
+			}
+			b.mu.Unlock()
+		}
+
+		err = b.readLoop(c)
+
+		// tear down connection
+		b.writeMu.Lock()
+		if b.conn != nil {
+			_ = b.conn.Close()
+			b.conn = nil
+		}
+		b.writeMu.Unlock()
+
+		select {
+		case <-b.stopCh:
+			return
+		default:
+			time.Sleep(backoff)
+			backoff = minDur(backoff*2, reconnectMax)
+		}
+	}
+}
+
+func (b *FuturesWebsocket) batcher() {
+	defer b.wg.Done()
+
+	t := time.NewTicker(batchPeriod)
+	defer t.Stop()
+
+	var subs, unsubs []string
+
+	flush := func() {
+		if len(subs) > 0 {
+			_ = b.sendSubscribe(subs)
+			subs = subs[:0]
+		}
+		if len(unsubs) > 0 {
+			_ = b.sendUnsubscribe(unsubs)
+			unsubs = unsubs[:0]
+		}
+	}
+
+	for {
+		select {
+		case <-b.stopCh:
+			return
+		case s := <-b.subQ:
+			if s != "" {
+				subs = append(subs, s)
+			}
+		case s := <-b.unsubQ:
+			if s != "" {
+				unsubs = append(unsubs, s)
+			}
+		case <-t.C:
+			flush()
+		}
+	}
+}
+
+func (b *FuturesWebsocket) readLoop(c *websocket.Conn) error {
+	for {
+		_, raw, err := c.ReadMessage()
+		if err != nil {
+			return err
+		}
+
+		fmt.Println("Received message:", string(raw))
+
+		// Stream data or command ack
+		if hasField(raw, `"stream"`) {
 			var container struct {
 				Stream string          `json:"stream"`
 				Data   json.RawMessage `json:"data"`
 			}
-		if err := json.Unmarshal(msgBytes, &container); err != nil {
+			if err := json.Unmarshal(raw, &container); err != nil || container.Stream == "" {
 				continue
 			}
 
 			b.mu.Lock()
-		ch, ok := b.activeStreams[container.Stream]
+			dst, ok := b.sinks[container.Stream]
+			st := b.states[container.Stream]
 			b.mu.Unlock()
-		if !ok {
+
+			if !ok || st == stateInactive || st == statePendingUnsub {
 				continue
 			}
 
@@ -142,7 +414,6 @@ func (b *FuturesWebsocket) readLoop() {
 			if err != nil {
 				continue
 			}
-
 			msg := domain.Message{
 				Identifier: id,
 				Payload:    container.Data,
@@ -150,9 +421,223 @@ func (b *FuturesWebsocket) readLoop() {
 			}
 
 			select {
-		case ch <- msg:
+			case dst <- msg:
 			default:
-			fmt.Printf("channel for %s is full, dropping message\n", container.Stream)
+				// drop on backpressure
+			}
+			continue
+		}
+
+		// Ack path
+		var ack struct {
+			Result json.RawMessage `json:"result"`
+			ID     int64           `json:"id"`
+		}
+		if err := json.Unmarshal(raw, &ack); err != nil || ack.ID == 0 {
+			continue
+		}
+
+		b.ackMu.Lock()
+		batch, ok := b.pendingA[ack.ID]
+		if ok {
+			delete(b.pendingA, ack.ID)
+		}
+		b.ackMu.Unlock()
+		if !ok {
+			continue
+		}
+
+		ackErr := (len(ack.Result) > 0 && string(ack.Result) != "null")
+
+		switch batch.method {
+		case "SUBSCRIBE":
+			b.mu.Lock()
+			for _, s := range batch.subjects {
+				if ackErr {
+					b.states[s] = stateError
+					// fail all start waiters
+					ws := b.startWaiters[s]
+					delete(b.startWaiters, s)
+					b.mu.Unlock()
+					for _, ch := range ws {
+						ch <- fmt.Errorf("subscribe failed")
+						close(ch)
+					}
+					b.mu.Lock()
+					continue
+				}
+				// success
+				b.states[s] = stateActive
+				ws := b.startWaiters[s]
+				delete(b.startWaiters, s)
+				dst := b.sinks[s]
+				b.mu.Unlock()
+
+				for _, ch := range ws {
+					ch <- nil
+					close(ch)
+				}
+				_ = dst // messages will flow via readLoop
+				b.mu.Lock()
+			}
+			b.mu.Unlock()
+
+		case "UNSUBSCRIBE":
+			b.mu.Lock()
+			for _, s := range batch.subjects {
+				if ackErr {
+					b.states[s] = stateError
+					ws := b.stopWaiters[s]
+					delete(b.stopWaiters, s)
+					b.mu.Unlock()
+					for _, ch := range ws {
+						ch <- fmt.Errorf("unsubscribe failed")
+						close(ch)
+					}
+					b.mu.Lock()
+					continue
+				}
+				// success
+				b.states[s] = stateInactive
+				delete(b.sinks, s) // stop delivering
+				ws := b.stopWaiters[s]
+				delete(b.stopWaiters, s)
+				b.mu.Unlock()
+				for _, ch := range ws {
+					ch <- nil
+					close(ch)
+				}
+				b.mu.Lock()
+			}
+			b.mu.Unlock()
 		}
 	}
 }
+
+func (b *FuturesWebsocket) nextID() int64 {
+	return int64(atomic.AddUint64(&b.idSeq, 1))
+}
+
+func (b *FuturesWebsocket) sendSubscribe(subjects []string) error {
+	if len(subjects) == 0 {
+		return nil
+	}
+	id := b.nextID()
+	req := map[string]any{
+		"method": "SUBSCRIBE",
+		"params": subjects,
+		"id":     id,
+	}
+	if err := b.writeJSON(req); err != nil {
+		// mark error and fail waiters
+		b.mu.Lock()
+		for _, s := range subjects {
+			b.states[s] = stateError
+			ws := b.startWaiters[s]
+			delete(b.startWaiters, s)
+			b.mu.Unlock()
+			for _, ch := range ws {
+				ch <- fmt.Errorf("subscribe send failed")
+				close(ch)
+			}
+			b.mu.Lock()
+		}
+		b.mu.Unlock()
+		return err
+	}
+	b.ackMu.Lock()
+	b.pendingA[id] = ackBatch{method: "SUBSCRIBE", subjects: append([]string(nil), subjects...)}
+	b.ackMu.Unlock()
+	return nil
+}
+
+func (b *FuturesWebsocket) sendUnsubscribe(subjects []string) error {
+	if len(subjects) == 0 {
+		return nil
+	}
+	id := b.nextID()
+	req := map[string]any{
+		"method": "UNSUBSCRIBE",
+		"params": subjects,
+		"id":     id,
+	}
+	if err := b.writeJSON(req); err != nil {
+		b.mu.Lock()
+		for _, s := range subjects {
+			b.states[s] = stateError
+			ws := b.stopWaiters[s]
+			delete(b.stopWaiters, s)
+			b.mu.Unlock()
+			for _, ch := range ws {
+				ch <- fmt.Errorf("unsubscribe send failed")
+				close(ch)
+			}
+			b.mu.Lock()
+		}
+		b.mu.Unlock()
+		return err
+	}
+	b.ackMu.Lock()
+	b.pendingA[id] = ackBatch{method: "UNSUBSCRIBE", subjects: append([]string(nil), subjects...)}
+	b.ackMu.Unlock()
+	return nil
+}
+
+func (b *FuturesWebsocket) writeJSON(v any) error {
+	// token bucket
+	select {
+	case <-b.stopCh:
+		return fmt.Errorf("stopped")
+	case <-b.tokensCh:
+	}
+
+	b.writeMu.Lock()
+	c := b.conn
+	b.writeMu.Unlock()
+	if c == nil {
+		return fmt.Errorf("not connected")
+	}
+
+	_ = c.SetWriteDeadline(time.Now().Add(writeWait))
+	return c.WriteJSON(v)
+}
+
+/* utilities */
+
+func (b *FuturesWebsocket) snapshotDesired(onlyTrue bool) []string {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	var out []string
+	for s, want := range b.desired {
+		if !onlyTrue || want {
+			out = append(out, s)
+		}
+	}
+	return out
+}
+
+func minDur(a, b time.Duration) time.Duration {
+	if a < b {
+		return a
+	}
+	return b
+}
+
+func hasField(raw []byte, needle string) bool {
+	// cheap check; avoids another allocation if it's obviously an ACK
+	return json.Valid(raw) && byteContains(raw, needle)
+}
+
+func byteContains(b []byte, sub string) bool {
+	n := len(sub)
+	if n == 0 || len(b) < n {
+		return false
+	}
+	// naive search; sufficient for small frames
+	for i := 0; i <= len(b)-n; i++ {
+		if string(b[i:i+n]) == sub {
+			return true
+		}
+	}
+	return false
+}

services/data_service/internal/provider/provider.go

@@ -1,19 +1,16 @@
 package provider
 
-import (
+import "gitlab.michelsen.id/phillmichelsen/tessera/services/data_service/internal/domain"
-	"gitlab.michelsen.id/phillmichelsen/tessera/services/data_service/internal/domain"
-)
 
 type Provider interface {
 	Start() error
 	Stop()
 
-	RequestStream(subject string, channel chan domain.Message) error
+	StartStream(key string, destination chan<- domain.Message) <-chan error
-	CancelStream(subject string)
+	StopStream(key string) <-chan error
-	GetActiveStreams() []string
-	IsStreamActive(subject string) bool
 
-	Fetch(subject string) (domain.Message, error)
+	Fetch(key string) (domain.Message, error)
 
-	IsValidSubject(subject string, isFetch bool) bool
+	IsStreamActive(key string) bool
+	IsValidSubject(key string, isFetch bool) bool
 }

services/data_service/internal/router/router.go

@@ -13,9 +13,9 @@ type Router struct {
 	mu       sync.RWMutex
 }
 
-func NewRouter() *Router {
+func NewRouter(buffer int) *Router {
 	return &Router{
-		incoming: make(chan domain.Message, 512), // Buffered channel for incoming messages
+		incoming: make(chan domain.Message, buffer), // Buffered channel for incoming messages
 		routes:   make(map[domain.Identifier][]chan<- domain.Message),
 	}
 }
@@ -33,7 +33,7 @@ func (r *Router) Run() {
 			select {
 			case ch <- msg:
 			default:
-				fmt.Println("Dropped message, buffer full!!!") // TODO: Handle full buffer case more gracefully
+				fmt.Println("Router could not push message to a full buffer...") // TODO: Handle full buffer case more gracefully
 			}
 		}
 		r.mu.RUnlock()

services/data_service/internal/server/gprc_control_server.go

@@ -22,7 +22,7 @@ func NewGRPCControlServer(m *manager.Manager) *GRPCControlServer {
 }
 
 // StartStream creates a new session. It does NOT attach client channels.
-// Your streaming RPC should later call GetChannels(sessionID, opts).
+// Your streaming RPC should later call AttachClient(sessionID, opts).
 func (s *GRPCControlServer) StartStream(_ context.Context, req *pb.StartStreamRequest) (*pb.StartStreamResponse, error) {
 	sessionID, err := s.manager.NewSession(time.Duration(1) * time.Minute) // timeout set to 1 minute
 	if err != nil {
@@ -51,7 +51,7 @@ func (s *GRPCControlServer) ConfigureStream(_ context.Context, req *pb.Configure
 		ids = append(ids, id)
 	}
 
-	if err := s.manager.SetSubscriptions(streamID, ids); err != nil {
+	if err := s.manager.ConfigureSession(streamID, ids); err != nil {
 		// Map common manager errors to gRPC codes.
 		switch err {
 		case manager.ErrSessionNotFound:

services/data_service/internal/server/grpc_streaming_server.go

@@ -28,15 +28,7 @@ func (s *GRPCStreamingServer) ConnectStream(req *pb.ConnectStreamRequest, stream
 		return fmt.Errorf("invalid UUID: %w", err)
 	}
 
-	// Defaults; tune or map from req if your proto carries options.
+	_, out, err := s.manager.AttachClient(sessionID, 256, 1024)
-	opts := manager.ChannelOpts{
-		InBufSize:    256,
-		OutBufSize:   1024,
-		DropOutbound: true, // do not let slow clients stall producers
-		DropInbound:  true, // irrelevant here (we don't send inbound), safe default
-	}
-
-	_, out, err := s.manager.GetChannels(sessionID, opts)
 	if err != nil {
 		return fmt.Errorf("attach channels: %w", err)
 	}