Write each handler batch from its flush goroutine, outside the mutex (closes #65)

The AS, peer and prefix handlers wrote their batch to the database while
holding the mutex HandleMessage needs, and HandleMessage wrote a full batch
itself. For as long as a write took, the handler took nothing off its queue,
and the streamer dropped updates once the queue was half full.

Now only the flush goroutine writes: it takes the batch under the mutex,
releases it, then writes. A full batch wakes the flush goroutine instead.
Stop still writes what the handler holds.

While a write runs, the next batch has no size limit; the README memory
section says so.

Model: opus-5-5

README.md

@@ -141,7 +141,10 @@ routewatch/
 - **Batched Writes**: All database operations are batched for performance
 - **Backpressure**: Probabilistic message dropping when queues exceed 50% capacity
 - **Graceful Shutdown**: 60-second timeout, flushes all pending batches
-- **Reconnection**: Exponential backoff (5s-320s) with reset after 30s of stable connection
+- **Reconnection**: Any end of the stream other than shutdown, including RIS Live
+  ending the response normally, marks the feed disconnected and reconnects with
+  exponential backoff (5s-320s). The backoff resets after a stream that lasted
+  more than 30s. A stream with no line for 60s is dropped and reconnected.
 - **IP Lookup**: the most specific live route is found by looking up the
   address's prefix at each mask length, longest first, in the prefix index
   (at most 25 lookups for IPv4, 113 for IPv6); prefixes are stored in the
@@ -230,6 +233,10 @@ What happens at each limit:
   keeps running.
 - Handler queues: each of the four handler queues holds at most 20,000 messages.
   When a queue fills, the streamer drops messages instead of blocking.
+- Handler batches: while the AS, peer or prefix handler writes a batch to the
+  database, the updates that arrive collect in its next batch, which has no size
+  limit. A database that stays slower than the feed grows memory instead of
+  dropping updates.
 
 With `DEBUG=routewatch` the daemon logs a `System stats` line every 60 seconds
 with the goroutine count and Go memory figures.

internal/routewatch/ashandler.go

@@ -32,10 +32,13 @@ type ASHandler struct {
 	db     database.Store
 	logger *logger.Logger
 
-	// Batching
+	// Batching. mu guards batch only: the flush goroutine takes the batch
+	// under mu and writes it after releasing mu, so HandleMessage never
+	// waits for a database write.
 	mu          sync.Mutex
 	batch       []asnOp
-	lastFlush time.Time
+	batchFullCh chan struct{} // wakes the flush goroutine when batch is full
+	lastFlush   time.Time     // used only by the flush goroutine
 	stopCh      chan struct{}
 	wg          sync.WaitGroup
 }
@@ -55,6 +58,7 @@ func NewASHandler(db database.Store, logger *logger.Logger) *ASHandler {
 		db:          db,
 		logger:      logger,
 		batch:       make([]asnOp, 0, asnBatchSize),
+		batchFullCh: make(chan struct{}, 1),
 		lastFlush:   time.Now(),
 		stopCh:      make(chan struct{}),
 	}
@@ -85,8 +89,8 @@ func (h *ASHandler) QueueCapacity() int {
 // HandleMessage processes a RIS Live BGP message by extracting all ASNs from
 // the AS path and queuing them for batch insertion into the database. The
 // origin ASN (last element in the path) and all transit ASNs are recorded
-// with their associated timestamps. The batch is automatically flushed when
+// with their associated timestamps. When the batch reaches the configured size
-// it reaches the configured size threshold.
+// threshold, the flush goroutine is woken to write it.
 func (h *ASHandler) HandleMessage(msg *ristypes.RISMessage) {
 	// Use the pre-parsed timestamp
 	timestamp := msg.ParsedTimestamp
@@ -118,13 +122,18 @@ func (h *ASHandler) HandleMessage(msg *ristypes.RISMessage) {
 		}
 	}
 
-	// Check if we need to flush
+	// A full batch wakes the flush goroutine. If it is writing, the wake-up
+	// waits in batchFullCh and the batch keeps growing until the write ends.
 	if len(h.batch) >= asnBatchSize {
-		h.flushBatchLocked()
+		select {
+		case h.batchFullCh <- struct{}{}:
+		default: // already woken
+		}
 	}
 }
 
-// flushLoop runs in a goroutine and periodically flushes batches
+// flushLoop runs in a goroutine, the only one that writes the batch: it
+// flushes when the batch is full, periodically, and once more on Stop.
 func (h *ASHandler) flushLoop() {
 	defer h.wg.Done()
 	ticker := time.NewTicker(asnBatchTimeout)
@@ -133,31 +142,46 @@ func (h *ASHandler) flushLoop() {
 	for {
 		select {
 		case <-ticker.C:
-			h.mu.Lock()
 			if time.Since(h.lastFlush) >= asnBatchTimeout {
-				h.flushBatchLocked()
+				h.flushBatch()
 			}
-			h.mu.Unlock()
+		case <-h.batchFullCh:
+			h.flushBatch()
 		case <-h.stopCh:
 			// Final flush
-			h.mu.Lock()
+			h.flushBatch()
-			h.flushBatchLocked()
-			h.mu.Unlock()
 
 			return
 		}
 	}
 }
 
-// flushBatchLocked flushes the ASN batch to the database (must be called with mutex held)
+// takeBatch returns the pending operations and leaves an empty batch in their
-func (h *ASHandler) flushBatchLocked() {
+// place for HandleMessage to fill.
+func (h *ASHandler) takeBatch() []asnOp {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+
 	if len(h.batch) == 0 {
+		return nil
+	}
+	batch := h.batch
+	h.batch = make([]asnOp, 0, asnBatchSize)
+
+	return batch
+}
+
+// flushBatch takes the pending batch and writes it to the database, without
+// holding the mutex during the write.
+func (h *ASHandler) flushBatch() {
+	batch := h.takeBatch()
+	if len(batch) == 0 {
 		return
 	}
 
 	// Process ASNs first (deduped)
 	asnMap := make(map[int]time.Time)
-	for _, op := range h.batch {
+	for _, op := range batch {
 		if existing, ok := asnMap[op.number]; !ok || op.timestamp.After(existing) {
 			asnMap[op.number] = op.timestamp
 		}
@@ -168,8 +192,6 @@ func (h *ASHandler) flushBatchLocked() {
 		h.logger.Error("Failed to process ASN batch", "error", err, "count", len(asnMap))
 	}
 
-	// Clear batch
-	h.batch = h.batch[:0]
 	h.lastFlush = time.Now()
 }
 

internal/routewatch/batchwrite_test.go

@@ -0,0 +1,192 @@
+package routewatch
+
+import (
+	"slices"
+	"sync"
+	"testing"
+	"time"
+
+	"git.eeqj.de/sneak/routewatch/internal/database"
+	"git.eeqj.de/sneak/routewatch/internal/logger"
+	"git.eeqj.de/sneak/routewatch/internal/ristypes"
+)
+
+// handleMessageTimeout is how long a test waits for HandleMessage to return,
+// or for a write to start, before it fails.
+const handleMessageTimeout = 2 * time.Second
+
+// blockingStore is a mockStore whose batch writes for the AS, peer and prefix
+// handlers wait until the test closes release. writeStarted is closed when the
+// first of them starts. It records what each write was given.
+type blockingStore struct {
+	*mockStore
+
+	release      chan struct{}
+	writeStarted chan struct{}
+	startOnce    sync.Once
+
+	mu       sync.Mutex
+	prefixes map[string]bool
+	asns     map[int]bool
+	peers    map[string]bool
+}
+
+func newBlockingStore() *blockingStore {
+	return &blockingStore{
+		mockStore:    newMockStore(),
+		release:      make(chan struct{}),
+		writeStarted: make(chan struct{}),
+		prefixes:     make(map[string]bool),
+		asns:         make(map[int]bool),
+		peers:        make(map[string]bool),
+	}
+}
+
+func (s *blockingStore) waitForRelease() {
+	s.startOnce.Do(func() { close(s.writeStarted) })
+	<-s.release
+}
+
+func (s *blockingStore) UpsertLiveRouteBatch(routes []*database.LiveRoute) error {
+	s.waitForRelease()
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	for _, route := range routes {
+		s.prefixes[route.Prefix] = true
+	}
+
+	return nil
+}
+
+func (s *blockingStore) GetOrCreateASNBatch(asns map[int]time.Time) error {
+	s.waitForRelease()
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	for asn := range asns {
+		s.asns[asn] = true
+	}
+
+	return nil
+}
+
+func (s *blockingStore) UpdatePeerBatch(peers map[string]database.PeerUpdate) error {
+	s.waitForRelease()
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	for peerIP := range peers {
+		s.peers[peerIP] = true
+	}
+
+	return nil
+}
+
+// returnsInTime reports whether f returns within handleMessageTimeout.
+func returnsInTime(f func()) bool {
+	done := make(chan struct{})
+	go func() {
+		f()
+		close(done)
+	}()
+
+	select {
+	case <-done:
+		return true
+	case <-time.After(handleMessageTimeout):
+		return false
+	}
+}
+
+// handleWhileAWriteIsBlocked hands h the messages in fill, which fill its
+// batch so that the handler writes it to store, waits until the write starts,
+// and hands h the message extra while store holds the write. HandleMessage
+// must return each time. Then it lets the writes finish and stops h, which
+// writes what it holds.
+func handleWhileAWriteIsBlocked(
+	t *testing.T,
+	store *blockingStore,
+	h interface {
+		HandleMessage(*ristypes.RISMessage)
+		Stop()
+	},
+	fill []*ristypes.RISMessage,
+	extra *ristypes.RISMessage,
+) {
+	t.Helper()
+	defer h.Stop()
+	defer close(store.release)
+
+	if !returnsInTime(func() {
+		for _, msg := range fill {
+			h.HandleMessage(msg)
+		}
+	}) {
+		t.Fatal("HandleMessage did not return after filling the batch: it waits for the database write")
+	}
+
+	select {
+	case <-store.writeStarted:
+	case <-time.After(handleMessageTimeout):
+		t.Fatal("the handler did not write its full batch")
+	}
+
+	if !returnsInTime(func() { h.HandleMessage(extra) }) {
+		t.Fatal("HandleMessage did not return while a database write was blocked")
+	}
+}
+
+func TestPrefixHandlerHandlesMessagesWhileAWriteIsBlocked(t *testing.T) {
+	const (
+		fillPrefix  = "192.0.2.0/24"
+		extraPrefix = "198.51.100.0/24"
+	)
+
+	announce := func(prefixes []string) *ristypes.RISMessage {
+		return &ristypes.RISMessage{
+			Peer:          testPeerIP,
+			Path:          ristypes.ASPath{testASNA},
+			Announcements: []ristypes.RISAnnouncement{{NextHop: testPeerIP, Prefixes: prefixes}},
+		}
+	}
+
+	store := newBlockingStore()
+	h := NewPrefixHandler(store, logger.New())
+	handleWhileAWriteIsBlocked(t, store, h,
+		[]*ristypes.RISMessage{announce(slices.Repeat([]string{fillPrefix}, prefixBatchSize))},
+		announce([]string{extraPrefix}))
+
+	store.mu.Lock()
+	defer store.mu.Unlock()
+	if !store.prefixes[extraPrefix] {
+		t.Errorf("Stop did not write %s, handled while a write was blocked", extraPrefix)
+	}
+}
+
+func TestASHandlerHandlesMessagesWhileAWriteIsBlocked(t *testing.T) {
+	store := newBlockingStore()
+	h := NewASHandler(store, logger.New())
+	handleWhileAWriteIsBlocked(t, store, h,
+		[]*ristypes.RISMessage{{Path: slices.Repeat(ristypes.ASPath{testASNA}, asnBatchSize)}},
+		&ristypes.RISMessage{Path: ristypes.ASPath{testASNB}})
+
+	store.mu.Lock()
+	defer store.mu.Unlock()
+	if !store.asns[testASNB] {
+		t.Errorf("Stop did not write AS%d, handled while a write was blocked", testASNB)
+	}
+}
+
+func TestPeerHandlerHandlesMessagesWhileAWriteIsBlocked(t *testing.T) {
+	const otherPeerIP = "2001:db8:ffff::2"
+
+	store := newBlockingStore()
+	h := NewPeerHandler(store, logger.New())
+	handleWhileAWriteIsBlocked(t, store, h,
+		slices.Repeat([]*ristypes.RISMessage{{Peer: testPeerIP}}, peerBatchSize),
+		&ristypes.RISMessage{Peer: otherPeerIP})
+
+	store.mu.Lock()
+	defer store.mu.Unlock()
+	if !store.peers[otherPeerIP] {
+		t.Errorf("Stop did not write peer %s, handled while a write was blocked", otherPeerIP)
+	}
+}

internal/routewatch/peerhandler.go

@@ -35,10 +35,13 @@ type PeerHandler struct {
 	db     database.Store
 	logger *logger.Logger
 
-	// Batching
+	// Batching. mu guards peerBatch only: the flush goroutine takes the batch
+	// under mu and writes it after releasing mu, so HandleMessage never
+	// waits for a database write.
 	mu          sync.Mutex
 	peerBatch   []peerUpdate
-	lastFlush time.Time
+	batchFullCh chan struct{} // wakes the flush goroutine when peerBatch is full
+	lastFlush   time.Time     // used only by the flush goroutine
 	stopCh      chan struct{}
 	wg          sync.WaitGroup
 }
@@ -59,6 +62,7 @@ func NewPeerHandler(db database.Store, logger *logger.Logger) *PeerHandler {
 		db:          db,
 		logger:      logger,
 		peerBatch:   make([]peerUpdate, 0, peerBatchSize),
+		batchFullCh: make(chan struct{}, 1),
 		lastFlush:   time.Now(),
 		stopCh:      make(chan struct{}),
 	}
@@ -87,8 +91,8 @@ func (h *PeerHandler) QueueCapacity() int {
 
 // HandleMessage processes a RIS message to track peer information.
 // It extracts the peer IP address and ASN from the message and adds
-// the update to an internal batch. When the batch reaches peerBatchSize
+// the update to an internal batch. The flush goroutine writes the batch to the
-// or the batch timeout expires, the batch is flushed to the database.
+// database when it reaches peerBatchSize or the batch timeout expires.
 func (h *PeerHandler) HandleMessage(msg *ristypes.RISMessage) {
 	// Parse peer ASN from string
 	peerASN := 0
@@ -109,13 +113,18 @@ func (h *PeerHandler) HandleMessage(msg *ristypes.RISMessage) {
 		timestamp:   msg.ParsedTimestamp,
 	})
 
-	// Check if we need to flush
+	// A full batch wakes the flush goroutine. If it is writing, the wake-up
+	// waits in batchFullCh and the batch keeps growing until the write ends.
 	if len(h.peerBatch) >= peerBatchSize {
-		h.flushBatchLocked()
+		select {
+		case h.batchFullCh <- struct{}{}:
+		default: // already woken
+		}
 	}
 }
 
-// flushLoop runs in a goroutine and periodically flushes batches
+// flushLoop runs in a goroutine, the only one that writes the batch: it
+// flushes when the batch is full, periodically, and once more on Stop.
 func (h *PeerHandler) flushLoop() {
 	defer h.wg.Done()
 	ticker := time.NewTicker(peerBatchTimeout)
@@ -124,31 +133,46 @@ func (h *PeerHandler) flushLoop() {
 	for {
 		select {
 		case <-ticker.C:
-			h.mu.Lock()
 			if time.Since(h.lastFlush) >= peerBatchTimeout {
-				h.flushBatchLocked()
+				h.flushBatch()
 			}
-			h.mu.Unlock()
+		case <-h.batchFullCh:
+			h.flushBatch()
 		case <-h.stopCh:
 			// Final flush
-			h.mu.Lock()
+			h.flushBatch()
-			h.flushBatchLocked()
-			h.mu.Unlock()
 
 			return
 		}
 	}
 }
 
-// flushBatchLocked flushes the peer batch to the database (must be called with mutex held)
+// takeBatch returns the pending updates and leaves an empty batch in their
-func (h *PeerHandler) flushBatchLocked() {
+// place for HandleMessage to fill.
+func (h *PeerHandler) takeBatch() []peerUpdate {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+
 	if len(h.peerBatch) == 0 {
+		return nil
+	}
+	batch := h.peerBatch
+	h.peerBatch = make([]peerUpdate, 0, peerBatchSize)
+
+	return batch
+}
+
+// flushBatch takes the pending batch and writes it to the database, without
+// holding the mutex during the write.
+func (h *PeerHandler) flushBatch() {
+	batch := h.takeBatch()
+	if len(batch) == 0 {
 		return
 	}
 
 	// Deduplicate by peer IP, keeping the latest update for each peer
 	peerMap := make(map[string]peerUpdate)
-	for _, update := range h.peerBatch {
+	for _, update := range batch {
 		if existing, ok := peerMap[update.peerIP]; !ok || update.timestamp.After(existing.timestamp) {
 			peerMap[update.peerIP] = update
 		}
@@ -170,8 +194,6 @@ func (h *PeerHandler) flushBatchLocked() {
 		h.logger.Error("Failed to process peer batch", "error", err, "count", len(dbPeerMap))
 	}
 
-	// Clear batch
-	h.peerBatch = h.peerBatch[:0]
 	h.lastFlush = time.Now()
 }
 

internal/routewatch/prefixhandler.go

@@ -40,10 +40,13 @@ type PrefixHandler struct {
 	logger  *logger.Logger
 	metrics *metrics.Tracker
 
-	// Batching
+	// Batching. mu guards batch only: the flush goroutine takes the batch
+	// under mu and writes it after releasing mu, so HandleMessage never
+	// waits for a database write.
 	mu          sync.Mutex
 	batch       []prefixUpdate
-	lastFlush time.Time
+	batchFullCh chan struct{} // wakes the flush goroutine when batch is full
+	lastFlush   time.Time     // used only by the flush goroutine
 	stopCh      chan struct{}
 	wg          sync.WaitGroup
 }
@@ -63,6 +66,7 @@ func NewPrefixHandler(db database.Store, logger *logger.Logger) *PrefixHandler {
 		db:          db,
 		logger:      logger,
 		batch:       make([]prefixUpdate, 0, prefixBatchSize),
+		batchFullCh: make(chan struct{}, 1),
 		lastFlush:   time.Now(),
 		stopCh:      make(chan struct{}),
 	}
@@ -140,13 +144,18 @@ func (h *PrefixHandler) HandleMessage(msg *ristypes.RISMessage) {
 		}
 	}
 
-	// Check if we need to flush
+	// A full batch wakes the flush goroutine. If it is writing, the wake-up
+	// waits in batchFullCh and the batch keeps growing until the write ends.
 	if len(h.batch) >= prefixBatchSize {
-		h.flushBatchLocked()
+		select {
+		case h.batchFullCh <- struct{}{}:
+		default: // already woken
+		}
 	}
 }
 
-// flushLoop runs in a goroutine and periodically flushes batches
+// flushLoop runs in a goroutine, the only one that writes the batch: it
+// flushes when the batch is full, periodically, and once more on Stop.
 func (h *PrefixHandler) flushLoop() {
 	defer h.wg.Done()
 	ticker := time.NewTicker(prefixBatchTimeout)
@@ -155,30 +164,45 @@ func (h *PrefixHandler) flushLoop() {
 	for {
 		select {
 		case <-ticker.C:
-			h.mu.Lock()
 			if time.Since(h.lastFlush) >= prefixBatchTimeout {
-				h.flushBatchLocked()
+				h.flushBatch()
 			}
-			h.mu.Unlock()
+		case <-h.batchFullCh:
+			h.flushBatch()
 		case <-h.stopCh:
 			// Final flush
-			h.mu.Lock()
+			h.flushBatch()
-			h.flushBatchLocked()
-			h.mu.Unlock()
 
 			return
 		}
 	}
 }
 
-// flushBatchLocked flushes the prefix batch to the database (must be called with mutex held)
+// takeBatch returns the pending updates and leaves an empty batch in their
-func (h *PrefixHandler) flushBatchLocked() {
+// place for HandleMessage to fill.
+func (h *PrefixHandler) takeBatch() []prefixUpdate {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+
 	if len(h.batch) == 0 {
+		return nil
+	}
+	batch := h.batch
+	h.batch = make([]prefixUpdate, 0, prefixBatchSize)
+
+	return batch
+}
+
+// flushBatch takes the pending batch and writes it to the database, without
+// holding the mutex during the write.
+func (h *PrefixHandler) flushBatch() {
+	batch := h.takeBatch()
+	if len(batch) == 0 {
 		return
 	}
 
 	startTime := time.Now()
-	batchSize := len(h.batch)
+	batchSize := len(batch)
 
 	// For each prefix and peer, keep only the last update the feed delivered.
 	// Updates from different peers for the same prefix are all kept. Arrival
@@ -186,7 +210,7 @@ func (h *PrefixHandler) flushBatchLocked() {
 	// updates within one second carry the same one.
 	type prefixAndPeer struct{ prefix, peer string }
 	lastUpdates := make(map[prefixAndPeer]prefixUpdate)
-	for _, update := range h.batch {
+	for _, update := range batch {
 		lastUpdates[prefixAndPeer{update.prefix, update.peer}] = update
 	}
 
@@ -261,8 +285,6 @@ func (h *PrefixHandler) flushBatchLocked() {
 		"duration_ms", elapsed.Milliseconds(),
 	)
 
-	// Clear batch
-	h.batch = h.batch[:0]
 	h.lastFlush = time.Now()
 }
 

internal/routewatch/prefixhandler_test.go

@@ -30,11 +30,6 @@ func TestPrefixHandlerStoresPrefixesTheIPLookupFinds(t *testing.T) {
 
 	// Built without NewPrefixHandler's flush timer, so the test flushes itself.
 	h := &PrefixHandler{db: db, logger: logger.New()}
-	flush := func() {
-		h.mu.Lock()
-		defer h.mu.Unlock()
-		h.flushBatchLocked()
-	}
 
 	ts := time.Date(2026, 1, 2, 3, 4, 5, 0, time.UTC)
 	h.HandleMessage(&ristypes.RISMessage{
@@ -46,7 +41,7 @@ func TestPrefixHandlerStoresPrefixesTheIPLookupFinds(t *testing.T) {
 			Prefixes: []string{"2001:db8:1::/48", "192.0.2.0/24"},
 		}},
 	})
-	flush()
+	h.flushBatch()
 
 	for ip, want := range map[string]string{
 		"2001:db8:1::1": "2001:db8:1::/48",
@@ -66,7 +61,7 @@ func TestPrefixHandlerStoresPrefixesTheIPLookupFinds(t *testing.T) {
 		ParsedTimestamp: ts.Add(time.Minute),
 		Withdrawals:     []string{"2001:db8:1:0:0:0:0:0/48", "192.0.2.0/24"},
 	})
-	flush()
+	h.flushBatch()
 
 	for _, ip := range []string{"2001:db8:1::1", "192.0.2.1"} {
 		if info, err := db.GetIPInfo(ip); !errors.Is(err, database.ErrNoRoute) {
@@ -164,9 +159,7 @@ func TestPrefixHandlerAppliesEachPeersLastUpdateInABatch(t *testing.T) {
 				for _, msg := range batch {
 					h.HandleMessage(msg)
 				}
-				h.mu.Lock()
+				h.flushBatch()
-				h.flushBatchLocked()
-				h.mu.Unlock()
 			}
 
 			ctx := context.Background()

internal/streamer/streamer.go

@@ -7,6 +7,7 @@ import (
 	"compress/gzip"
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"math"
@@ -48,6 +49,7 @@ const (
 	metricsUpdateRate     = time.Second
 	minBackoffDelay       = 5 * time.Second
 	maxBackoffDelay       = 320 * time.Second
+	streamIdleTimeout     = 60 * time.Second // Reconnect when no line arrives for this long
 	metricsLogInterval    = 10 * time.Second
 	bytesPerKB            = 1024
 	bytesPerMB            = 1024 * 1024
@@ -58,6 +60,9 @@ const (
 	backpressureSlope     = 2.0 // Slope for linear drop probability increase
 )
 
+// errStreamEnded is returned by stream when RIS Live ends the response.
+var errStreamEnded = errors.New("RIS Live ended the stream")
+
 // MessageHandler defines the interface for processing RIS messages.
 // Implementations must specify which message types they want to receive,
 // how to process messages, and their desired queue capacity.
@@ -107,6 +112,8 @@ type Streamer struct {
 	logger       *logger.Logger
 	client       *http.Client
 	url          string
+	idleTimeout  time.Duration // Drop the connection when no line arrives for this long
+	minBackoff   time.Duration // First wait before reconnecting; doubles up to maxBackoffDelay
 	handlers     []*handlerInfo
 	rawHandler   RawMessageHandler
 	mu           sync.RWMutex
@@ -126,6 +133,8 @@ func New(logger *logger.Logger, metrics *metrics.Tracker) *Streamer {
 	return &Streamer{
 		logger:      logger,
 		url:         risLiveURL,
+		idleTimeout: streamIdleTimeout,
+		minBackoff:  minBackoffDelay,
 		client: &http.Client{
 			Timeout: 0, // No timeout for streaming
 			Transport: &http.Transport{
@@ -405,7 +414,7 @@ func (s *Streamer) updateMetrics(messageBytes int) {
 
 // streamWithReconnect handles streaming with automatic reconnection and exponential backoff
 func (s *Streamer) streamWithReconnect(ctx context.Context) {
-	backoffDelay := minBackoffDelay
+	backoffDelay := s.minBackoff
 	consecutiveFailures := 0
 
 	for {
@@ -422,8 +431,9 @@ func (s *Streamer) streamWithReconnect(ctx context.Context) {
 		err := s.stream(ctx)
 		streamDuration := time.Since(startTime)
 
-		if err == nil {
+		// Only Stop ends the streamer. A stream that ends any other way,
-			// Clean exit (context cancelled)
+		// with or without an error, is a disconnect and is reconnected.
+		if ctx.Err() != nil {
 			return
 		}
 
@@ -434,17 +444,13 @@ func (s *Streamer) streamWithReconnect(ctx context.Context) {
 			"stream_duration", streamDuration)
 		s.metrics.SetConnected(false)
 
-		// Check if context is cancelled
+		// A stream that lasted more than 30 seconds resets the backoff. That
-		if ctx.Err() != nil {
+		// includes a stream dropped after streamIdleTimeout with no line, so
-			return
+		// a silent feed is retried after the first backoff delay every time.
-		}
-
-		// If we streamed for more than 30 seconds, reset the backoff
-		// This indicates we had a successful connection that received data
 		if streamDuration > 30*time.Second {
 			s.logger.Info("Resetting backoff delay due to successful connection",
 				"stream_duration", streamDuration)
-			backoffDelay = minBackoffDelay
+			backoffDelay = s.minBackoff
 			consecutiveFailures = 0
 		} else {
 			// Increment consecutive failures
@@ -477,7 +483,13 @@ func (s *Streamer) stream(ctx context.Context) error {
 	connCtx, connCancel := context.WithCancel(ctx)
 	defer connCancel()
 
-	req, err := http.NewRequestWithContext(ctx, "GET", s.url, nil)
+	// The idle timer drops the connection, by cancelling connCtx, when no
+	// line arrives for s.idleTimeout. Each line read starts it over. It
+	// also covers the wait for the server to answer.
+	idleTimer := time.AfterFunc(s.idleTimeout, connCancel)
+	defer idleTimer.Stop()
+
+	req, err := http.NewRequestWithContext(connCtx, "GET", s.url, nil)
 	if err != nil {
 		return fmt.Errorf("failed to create request: %w", err)
 	}
@@ -559,6 +571,8 @@ func (s *Streamer) stream(ctx context.Context) error {
 	scanner := bufio.NewScanner(reader)
 
 	for scanner.Scan() {
+		idleTimer.Reset(s.idleTimeout)
+
 		select {
 		case <-ctx.Done():
 			s.logger.Info("Stream stopped by context")
@@ -707,11 +721,19 @@ func (s *Streamer) stream(ctx context.Context) error {
 		s.mu.RUnlock()
 	}
 
+	if ctx.Err() != nil {
+		return ctx.Err()
+	}
+
+	if connCtx.Err() != nil {
+		return fmt.Errorf("no line received for %s", s.idleTimeout)
+	}
+
 	if err := scanner.Err(); err != nil {
 		return fmt.Errorf("scanner error: %w", err)
 	}
 
-	return nil
+	return errStreamEnded
 }
 
 // shouldDropForBackpressure determines if a message should be dropped based on queue utilization

internal/streamer/streamer_test.go

@@ -7,6 +7,7 @@ import (
 	"net/http"
 	"net/http/httptest"
 	"runtime"
+	"sync/atomic"
 	"testing"
 	"time"
 
@@ -39,6 +40,14 @@ func TestNewStreamer(t *testing.T) {
 	if s.metrics != metricsTracker {
 		t.Error("metrics tracker not set correctly")
 	}
+
+	if s.idleTimeout != streamIdleTimeout {
+		t.Errorf("idleTimeout = %s, want %s", s.idleTimeout, streamIdleTimeout)
+	}
+
+	if s.minBackoff != minBackoffDelay {
+		t.Errorf("minBackoff = %s, want %s", s.minBackoff, minBackoffDelay)
+	}
 }
 
 // TestStreamDoesNotLeakTickersAcrossReconnects drives many short-lived
@@ -57,8 +66,8 @@ func TestStreamDoesNotLeakTickersAcrossReconnects(t *testing.T) {
 
 	// One warm-up connection so any persistent HTTP transport goroutine exists
 	// before we take the baseline.
-	if err := s.stream(context.Background()); err != nil {
+	if err := s.stream(context.Background()); !errors.Is(err, errStreamEnded) {
-		t.Fatalf("warm-up stream returned error: %v", err)
+		t.Fatalf("warm-up stream returned %v, want %v", err, errStreamEnded)
 	}
 	s.client.CloseIdleConnections()
 
@@ -66,8 +75,8 @@ func TestStreamDoesNotLeakTickersAcrossReconnects(t *testing.T) {
 
 	const reconnects = 20
 	for range reconnects {
-		if err := s.stream(context.Background()); err != nil {
+		if err := s.stream(context.Background()); !errors.Is(err, errStreamEnded) {
-			t.Fatalf("stream returned error: %v", err)
+			t.Fatalf("stream returned %v, want %v", err, errStreamEnded)
 		}
 	}
 	s.client.CloseIdleConnections()
@@ -78,23 +87,35 @@ func TestStreamDoesNotLeakTickersAcrossReconnects(t *testing.T) {
 	}
 }
 
-// updateHandler wants UPDATE messages and does nothing with them.
+// updateLine is one UPDATE message as RIS Live sends it.
-type updateHandler struct{}
+const updateLine = `{"type":"ris_message","data":{"type":"UPDATE","peer":"192.0.2.1",` +
+	`"peer_asn":"64496","timestamp":1700000000}}` + "\n"
+
+// updateHandler wants UPDATE messages. For each one it handles it sends on
+// handled if there is room, and otherwise does nothing; a nil handled is
+// never sent on.
+type updateHandler struct {
+	handled chan struct{}
+}
 
 func (updateHandler) WantsMessage(messageType string) bool { return messageType == "UPDATE" }
-func (updateHandler) HandleMessage(*ristypes.RISMessage)   {}
 func (updateHandler) QueueCapacity() int                   { return 10 }
 
+func (h updateHandler) HandleMessage(*ristypes.RISMessage) {
+	select {
+	case h.handled <- struct{}{}:
+	default:
+	}
+}
+
 // TestStopBeforeMessageReachesQueues stops the streamer after the read loop
 // has checked for cancellation but before it hands the message to the handler
 // queues. That is the gap a stop from another goroutine can land in, and it
 // used to end in "send on closed channel". The raw handler runs in that gap on
 // the read loop itself, so calling Stop from it hits the gap every time.
 func TestStopBeforeMessageReachesQueues(t *testing.T) {
-	const line = `{"type":"ris_message","data":{"type":"UPDATE","peer":"192.0.2.1",` +
-		`"peer_asn":"64496","timestamp":1700000000}}` + "\n"
 	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
-		_, _ = io.WriteString(w, line)
+		_, _ = io.WriteString(w, updateLine)
 	}))
 	defer srv.Close()
 
@@ -117,6 +138,117 @@ func TestStopBeforeMessageReachesQueues(t *testing.T) {
 	s.Stop()
 }
 
+// TestReconnectWhenStreamEnds has the server end the first response
+// normally, as when RIS Live restarts. The streamer must treat that as a
+// disconnect and connect again.
+func TestReconnectWhenStreamEnds(t *testing.T) {
+	endResponse := func(http.ResponseWriter, *http.Request) {}
+	checkReconnect(t, endResponse, streamIdleTimeout)
+}
+
+// TestReconnectWhenStreamGoesIdle has the server hold the first response
+// open without sending anything. The streamer must drop it once the idle
+// timeout passes and connect again.
+func TestReconnectWhenStreamGoesIdle(t *testing.T) {
+	const idleTimeout = 50 * time.Millisecond
+	holdResponse := func(w http.ResponseWriter, r *http.Request) {
+		w.(http.Flusher).Flush()
+		<-r.Context().Done()
+	}
+	checkReconnect(t, holdResponse, idleTimeout)
+}
+
+// TestReconnectWhenServerNeverAnswers has the server accept the first
+// request and never answer it, not even with headers. The streamer must give
+// up on it once the idle timeout passes and connect again.
+func TestReconnectWhenServerNeverAnswers(t *testing.T) {
+	const idleTimeout = 50 * time.Millisecond
+	neverAnswer := func(_ http.ResponseWriter, r *http.Request) {
+		<-r.Context().Done()
+	}
+	checkReconnect(t, neverAnswer, idleTimeout)
+}
+
+// TestStreamNotDroppedWhileLinesArrive has the server send lines well within
+// the idle timeout for several times the timeout, then end the response. Each
+// line must start the idle timeout over, so the stream lasts until the server
+// ends it instead of being dropped as idle.
+func TestStreamNotDroppedWhileLinesArrive(t *testing.T) {
+	const (
+		idleTimeout  = 100 * time.Millisecond
+		lineInterval = idleTimeout / 10
+		sendFor      = 5 * idleTimeout
+	)
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		deadline := time.Now().Add(sendFor)
+		for time.Now().Before(deadline) {
+			_, _ = io.WriteString(w, updateLine)
+			w.(http.Flusher).Flush()
+			time.Sleep(lineInterval)
+		}
+	}))
+	defer srv.Close()
+
+	s := New(logger.New(), metrics.New())
+	s.url = srv.URL
+	s.idleTimeout = idleTimeout
+
+	if err := s.stream(context.Background()); !errors.Is(err, errStreamEnded) {
+		t.Fatalf("stream returned %v, want %v", err, errStreamEnded)
+	}
+}
+
+// checkReconnect runs a streamer against a fake RIS Live server. The first
+// response is served by firstResponse; every later one carries one UPDATE
+// line and then stays open. The streamer must come back for a second
+// response, must not be marked connected when it does, and must handle the
+// UPDATE in it.
+func checkReconnect(t *testing.T, firstResponse http.HandlerFunc, idleTimeout time.Duration) {
+	t.Helper()
+
+	const maxWait = 5 * time.Second
+
+	tracker := metrics.New()
+	connectedAtReconnect := make(chan bool, 1)
+	var requests atomic.Int32
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if requests.Add(1) == 1 {
+			firstResponse(w, r)
+
+			return
+		}
+		select {
+		case connectedAtReconnect <- tracker.IsConnected():
+		default:
+		}
+		_, _ = io.WriteString(w, updateLine)
+		w.(http.Flusher).Flush()
+		<-r.Context().Done()
+	}))
+	defer srv.Close()
+
+	s := New(logger.New(), tracker)
+	s.url = srv.URL
+	s.idleTimeout = idleTimeout
+	s.minBackoff = time.Millisecond
+	handler := updateHandler{handled: make(chan struct{}, 1)}
+	s.RegisterHandler(handler)
+	if err := s.Start(); err != nil {
+		t.Fatalf("Start returned error: %v", err)
+	}
+	defer s.Stop()
+
+	select {
+	case <-handler.handled:
+	case <-time.After(maxWait):
+		t.Fatalf("no message from a second response handled within %s", maxWait)
+	}
+
+	if <-connectedAtReconnect {
+		t.Error("streamer was still marked connected when it came back for a second response")
+	}
+}
+
 // settledGoroutineCount lets transient goroutines finish, then reports the
 // current count.
 func settledGoroutineCount() int {