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

@@ -144,8 +144,9 @@ routewatch/
 - **Reconnection**: Exponential backoff (5s-320s) with reset after 30s of stable connection
 - **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 33 lookups for IPv4, 129 for IPv6); prefixes are stored in the
-  text form Go's `net/netip` prints
+  (at most 25 lookups for IPv4, 113 for IPv6); prefixes are stored in the
+  text form Go's `net/netip` prints. Routes shorter than /8 for IPv4 or /16
+  for IPv6, such as a default route, are not used.
 
 ### Database Schema
 
@@ -229,6 +230,10 @@ 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/database/database.go

@@ -37,6 +37,14 @@ const (
 	ipv6Bits       = 128
 )
 
+// The IP lookup does not use routes shorter than these mask lengths. A default
+// route or a route this short only says where a peer sends traffic it has no
+// better route for; it does not say who holds the address.
+const (
+	minLookupMaskLengthV4 = 8
+	minLookupMaskLengthV6 = 16
+)
+
 // SQLite memory tuning. cache_size and busy_timeout go in the DSN so every
 // pooled connection gets them; the heap limits are process-wide and set once.
 const (
@@ -1876,11 +1884,12 @@ func (d *Database) GetIPInfo(ip string) (*IPInfo, error) {
 // GetIPInfoContext returns comprehensive IP information with context support.
 //
 // It finds the most specific live route for the address with point lookups on
-// the prefix index: for each mask length from the longest (32 or 128) down to 0
-// it looks up the address's network prefix at that length, and the first length
-// with a live route wins. That is at most 33 or 129 indexed lookups, however
-// many routes are live. The prefix is looked up in the text form net/netip
-// prints, which is the form the prefix handler stores.
+// the prefix index: for each mask length from the longest (32 or 128) down to
+// minLookupMaskLengthV4 or minLookupMaskLengthV6 it looks up the address's
+// network prefix at that length, and the first length with a live route wins.
+// That is at most 25 or 113 indexed lookups, however many routes are live. The
+// prefix is looked up in the text form net/netip prints, which is the form the
+// prefix handler stores.
 func (d *Database) GetIPInfoContext(ctx context.Context, ip string) (*IPInfo, error) {
 	addr, err := netip.ParseAddr(ip)
 	if err != nil {
@@ -1891,9 +1900,11 @@ func (d *Database) GetIPInfoContext(ctx context.Context, ip string) (*IPInfo, er
 
 	info := &IPInfo{IP: ip, IPVersion: ipVersionV6}
 	routesTable, prefixesTable := "live_routes_v6", "prefixes_v6"
+	minMaskLength := minLookupMaskLengthV6
 	if addr.Is4() {
 		info.IPVersion = ipVersionV4
 		routesTable, prefixesTable = "live_routes_v4", "prefixes_v4"
+		minMaskLength = minLookupMaskLengthV4
 	}
 
 	// Route info with peer count and prefix first_seen
@@ -1934,7 +1945,7 @@ func (d *Database) GetIPInfoContext(ctx context.Context, ip string) (*IPInfo, er
 	var whoisUpdatedAt sql.NullTime
 
 	found := false
-	for bits := addr.BitLen(); bits >= 0; bits-- {
+	for bits := addr.BitLen(); bits >= minMaskLength; bits-- {
 		prefix := netip.PrefixFrom(addr, bits).Masked().String()
 		err := stmt.QueryRowContext(ctx, prefix).Scan(
 			&info.Netblock,

internal/database/database_test.go

@@ -133,6 +133,78 @@ func TestGetIPInfoFindsMostSpecificLiveRoute(t *testing.T) {
 	lookup("2001:db8:1:2::3", "2001:db8:1::/48")
 }
 
+// TestGetIPInfoIgnoresShortRoutes checks that the IP lookup does not report a
+// default route, or a route shorter than /8 for IPv4 or /16 for IPv6, as the
+// route for an address, and that a route of exactly /8 or /16, or longer, still
+// is. The prefixes are written out so that a wrong limit fails the test.
+func TestGetIPInfoIgnoresShortRoutes(t *testing.T) {
+	cfg := &config.Config{StateDir: t.TempDir()}
+
+	db, err := New(cfg, logger.New())
+	if err != nil {
+		t.Fatalf("failed to create database: %v", err)
+	}
+	defer func() { _ = db.Close() }()
+
+	ts := time.Date(2026, 1, 2, 3, 4, 5, 0, time.UTC)
+	announce := func(prefix string, asn int) {
+		t.Helper()
+
+		err := db.UpsertLiveRouteBatch([]*LiveRoute{{
+			ID:          uuid.New(),
+			Prefix:      prefix,
+			MaskLength:  netip.MustParsePrefix(prefix).Bits(),
+			IPVersion:   detectIPVersion(prefix),
+			OriginASN:   asn,
+			PeerIP:      testPeerIP,
+			ASPath:      []int{asn},
+			NextHop:     testPeerIP,
+			LastUpdated: ts,
+		}})
+		if err != nil {
+			t.Fatalf("UpsertLiveRouteBatch(%s): %v", prefix, err)
+		}
+	}
+
+	// lookup checks that ip resolves to the live prefix netblock from asn.
+	lookup := func(ip, netblock string, asn int) {
+		t.Helper()
+
+		info, err := db.GetIPInfo(ip)
+		if err != nil {
+			t.Errorf("GetIPInfo(%s): %v", ip, err)
+
+			return
+		}
+		if info.Netblock != netblock || info.ASN != asn {
+			t.Errorf("GetIPInfo(%s) = %s AS%d, want %s AS%d", ip, info.Netblock, info.ASN, netblock, asn)
+		}
+	}
+
+	// Default and very short routes, and routes one bit shorter than each limit.
+	announce("2000::/3", 64500)
+	announce("0.0.0.0/0", 64500)
+	announce("2000::/15", 64500)
+	announce("192.0.0.0/7", 64500)
+
+	for _, ip := range []string{"2001:db8::1", "192.0.2.1"} {
+		if info, err := db.GetIPInfo(ip); !errors.Is(err, ErrNoRoute) {
+			t.Errorf("GetIPInfo(%s) with only short routes = %+v, %v; want ErrNoRoute", ip, info, err)
+		}
+	}
+
+	// Routes of exactly the minimum length are used.
+	announce("2001::/16", 64501)
+	announce("192.0.0.0/8", 64501)
+
+	lookup("2001:db8::1", "2001::/16", 64501)
+	lookup("192.0.2.1", "192.0.0.0/8", 64501)
+
+	announce("2001:db8::/32", 64502)
+
+	lookup("2001:db8::1", "2001:db8::/32", 64502)
+}
+
 // TestConnectionPoolPragmas holds several pooled connections open at once and
 // checks each one carries the per-connection settings from the DSN, plus the
 // process-wide hard heap limit, and sees the new file with auto_vacuum

internal/metrics/metrics.go

@@ -60,17 +60,22 @@ func New() *Tracker {
 	}
 }
 
-// SetConnected updates the connection status
+// SetConnected updates the connection status. Every connection after the
+// first counts one reconnect.
 func (t *Tracker) SetConnected(connected bool) {
-	wasConnected := t.isConnected.Swap(connected)
-	if connected {
-		t.mu.Lock()
-		t.connectedSince = time.Now()
-		t.mu.Unlock()
-		// Increment reconnect count (but not for the initial connection)
-		if wasConnected || t.reconnectCount.Load() > 0 {
-			t.reconnectCount.Add(1)
-		}
+	t.isConnected.Store(connected)
+	if !connected {
+		return
+	}
+
+	t.mu.Lock()
+	// connectedSince is zero until the first connection.
+	connectedBefore := !t.connectedSince.IsZero()
+	t.connectedSince = time.Now()
+	t.mu.Unlock()
+
+	if connectedBefore {
+		t.reconnectCount.Add(1)
 	}
 }
 

internal/metrics/metrics_test.go

@@ -0,0 +1,21 @@
+package metrics
+
+import "testing"
+
+// TestReconnectCount connects three times with a disconnect between each, the
+// way the streamer does, and expects every connection after the first to count
+// one reconnect.
+func TestReconnectCount(t *testing.T) {
+	tracker := New()
+
+	tracker.SetConnected(true)
+	tracker.SetConnected(false)
+	tracker.SetConnected(true)
+	tracker.SetConnected(false)
+	tracker.SetConnected(true)
+
+	const wantReconnects = 2
+	if got := tracker.GetReconnectCount(); got != wantReconnects {
+		t.Fatalf("reconnect count = %d, want %d", got, wantReconnects)
+	}
+}

internal/routewatch/ashandler.go

@@ -32,12 +32,15 @@ type ASHandler struct {
 	db     database.Store
 	logger *logger.Logger
 
-	// Batching
-	mu        sync.Mutex
-	batch     []asnOp
-	lastFlush time.Time
-	stopCh    chan struct{}
-	wg        sync.WaitGroup
+	// 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
+	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
 }
 
 type asnOp struct {
@@ -52,11 +55,12 @@ type asnOp struct {
 // background goroutine is terminated.
 func NewASHandler(db database.Store, logger *logger.Logger) *ASHandler {
 	h := &ASHandler{
-		db:        db,
-		logger:    logger,
-		batch:     make([]asnOp, 0, asnBatchSize),
-		lastFlush: time.Now(),
-		stopCh:    make(chan struct{}),
+		db:          db,
+		logger:      logger,
+		batch:       make([]asnOp, 0, asnBatchSize),
+		batchFullCh: make(chan struct{}, 1),
+		lastFlush:   time.Now(),
+		stopCh:      make(chan struct{}),
 	}
 
 	// Start the flush timer goroutine
@@ -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
-// it reaches the configured size threshold.
+// with their associated timestamps. When the batch reaches the configured size
+// 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.flushBatchLocked()
-			h.mu.Unlock()
+			h.flushBatch()
 
 			return
 		}
 	}
 }
 
-// flushBatchLocked flushes the ASN batch to the database (must be called with mutex held)
-func (h *ASHandler) flushBatchLocked() {
+// takeBatch returns the pending operations and leaves an empty batch in their
+// 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

@@ -35,12 +35,15 @@ type PeerHandler struct {
 	db     database.Store
 	logger *logger.Logger
 
-	// Batching
-	mu        sync.Mutex
-	peerBatch []peerUpdate
-	lastFlush time.Time
-	stopCh    chan struct{}
-	wg        sync.WaitGroup
+	// 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
+	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
 }
 
 type peerUpdate struct {
@@ -56,11 +59,12 @@ type peerUpdate struct {
 // should be stopped by calling Stop when it is no longer needed.
 func NewPeerHandler(db database.Store, logger *logger.Logger) *PeerHandler {
 	h := &PeerHandler{
-		db:        db,
-		logger:    logger,
-		peerBatch: make([]peerUpdate, 0, peerBatchSize),
-		lastFlush: time.Now(),
-		stopCh:    make(chan struct{}),
+		db:          db,
+		logger:      logger,
+		peerBatch:   make([]peerUpdate, 0, peerBatchSize),
+		batchFullCh: make(chan struct{}, 1),
+		lastFlush:   time.Now(),
+		stopCh:      make(chan struct{}),
 	}
 
 	// Start the flush timer goroutine
@@ -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
-// or the batch timeout expires, the batch is flushed to the database.
+// the update to an internal batch. The flush goroutine writes the batch to the
+// 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.flushBatchLocked()
-			h.mu.Unlock()
+			h.flushBatch()
 
 			return
 		}
 	}
 }
 
-// flushBatchLocked flushes the peer batch to the database (must be called with mutex held)
-func (h *PeerHandler) flushBatchLocked() {
+// takeBatch returns the pending updates and leaves an empty batch in their
+// 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

@@ -40,12 +40,15 @@ type PrefixHandler struct {
 	logger  *logger.Logger
 	metrics *metrics.Tracker
 
-	// Batching
-	mu        sync.Mutex
-	batch     []prefixUpdate
-	lastFlush time.Time
-	stopCh    chan struct{}
-	wg        sync.WaitGroup
+	// 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
+	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
 }
 
 type prefixUpdate struct {
@@ -60,11 +63,12 @@ type prefixUpdate struct {
 // NewPrefixHandler creates a new batched prefix tracking handler
 func NewPrefixHandler(db database.Store, logger *logger.Logger) *PrefixHandler {
 	h := &PrefixHandler{
-		db:        db,
-		logger:    logger,
-		batch:     make([]prefixUpdate, 0, prefixBatchSize),
-		lastFlush: time.Now(),
-		stopCh:    make(chan struct{}),
+		db:          db,
+		logger:      logger,
+		batch:       make([]prefixUpdate, 0, prefixBatchSize),
+		batchFullCh: make(chan struct{}, 1),
+		lastFlush:   time.Now(),
+		stopCh:      make(chan struct{}),
 	}
 
 	// Start the flush timer goroutine
@@ -105,7 +109,24 @@ func (h *PrefixHandler) HandleMessage(msg *ristypes.RISMessage) {
 	h.mu.Lock()
 	defer h.mu.Unlock()
 
-	// Process announcements
+	// Withdrawals are queued before announcements, because the batch keeps
+	// each peer's last update for a prefix: a message that lists a prefix as
+	// both withdrawn and announced keeps the route (RFC 4271 section 4.3).
+	for _, prefix := range msg.Withdrawals {
+		h.batch = append(h.batch, prefixUpdate{
+			prefix:      canonicalPrefix(prefix),
+			originASN:   originASN, // Use the originASN from path if available
+			peer:        msg.Peer,
+			messageType: "withdrawal",
+			timestamp:   timestamp,
+			path:        msg.Path,
+		})
+		// Record withdrawal in metrics
+		if h.metrics != nil {
+			h.metrics.RecordWithdrawal()
+		}
+	}
+
 	for _, announcement := range msg.Announcements {
 		for _, prefix := range announcement.Prefixes {
 			h.batch = append(h.batch, prefixUpdate{
@@ -123,29 +144,18 @@ func (h *PrefixHandler) HandleMessage(msg *ristypes.RISMessage) {
 		}
 	}
 
-	// Process withdrawals
-	for _, prefix := range msg.Withdrawals {
-		h.batch = append(h.batch, prefixUpdate{
-			prefix:      canonicalPrefix(prefix),
-			originASN:   originASN, // Use the originASN from path if available
-			peer:        msg.Peer,
-			messageType: "withdrawal",
-			timestamp:   timestamp,
-			path:        msg.Path,
-		})
-		// Record withdrawal in metrics
-		if h.metrics != nil {
-			h.metrics.RecordWithdrawal()
-		}
-	}
-
-	// 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)
@@ -154,39 +164,54 @@ 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.flushBatchLocked()
-			h.mu.Unlock()
+			h.flushBatch()
 
 			return
 		}
 	}
 }
 
-// flushBatchLocked flushes the prefix batch to the database (must be called with mutex held)
-func (h *PrefixHandler) flushBatchLocked() {
+// takeBatch returns the pending updates and leaves an empty batch in their
+// 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)
 
-	// Group updates by prefix to deduplicate
-	// For each prefix, keep the latest update
-	prefixMap := make(map[string]prefixUpdate)
-	for _, update := range h.batch {
-		key := update.prefix
-		if existing, ok := prefixMap[key]; !ok || update.timestamp.After(existing.timestamp) {
-			prefixMap[key] = update
-		}
+	// 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
+	// order decides, not the timestamp: timestamps are whole seconds, so
+	// updates within one second carry the same one.
+	type prefixAndPeer struct{ prefix, peer string }
+	lastUpdates := make(map[prefixAndPeer]prefixUpdate)
+	for _, update := range batch {
+		lastUpdates[prefixAndPeer{update.prefix, update.peer}] = update
 	}
 
 	// Collect routes to upsert and delete
@@ -196,7 +221,7 @@ func (h *PrefixHandler) flushBatchLocked() {
 	// Collect unique prefixes to update
 	prefixesToUpdate := make(map[string]time.Time)
 
-	for _, update := range prefixMap {
+	for _, update := range lastUpdates {
 		// Track prefix for both announcements and withdrawals
 		if _, exists := prefixesToUpdate[update.prefix]; !exists || update.timestamp.After(prefixesToUpdate[update.prefix]) {
 			prefixesToUpdate[update.prefix] = update.timestamp
@@ -255,13 +280,11 @@ func (h *PrefixHandler) flushBatchLocked() {
 	elapsed := time.Since(startTime)
 	h.logger.Debug("Flushed prefix batch",
 		"batch_size", batchSize,
-		"unique_prefixes", len(prefixMap),
+		"unique_prefixes", len(prefixesToUpdate),
 		"success", successCount,
 		"duration_ms", elapsed.Milliseconds(),
 	)
 
-	// Clear batch
-	h.batch = h.batch[:0]
 	h.lastFlush = time.Now()
 }
 

internal/routewatch/prefixhandler_test.go

@@ -1,7 +1,9 @@
 package routewatch
 
 import (
+	"context"
 	"errors"
+	"slices"
 	"testing"
 	"time"
 
@@ -28,11 +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{
@@ -44,7 +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",
@@ -64,7 +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) {
@@ -72,3 +69,138 @@ func TestPrefixHandlerStoresPrefixesTheIPLookupFinds(t *testing.T) {
 		}
 	}
 }
+
+// TestPrefixHandlerAppliesEachPeersLastUpdateInABatch runs updates for one
+// prefix through the prefix handler into a real database. All updates carry the
+// same timestamp, as updates within one second of the feed do. Within a batch,
+// the last update each peer sent for the prefix decides whether that peer has a
+// live route for it, and updates from different peers are all applied. The
+// in-memory route count and prefix distribution must match the route table.
+func TestPrefixHandlerAppliesEachPeersLastUpdateInABatch(t *testing.T) {
+	const (
+		prefix           = "198.51.100.0/24"
+		prefixMaskLength = 24
+		otherPeerIP      = "2001:db8:ffff::2"
+	)
+
+	ts := time.Date(2026, 1, 2, 3, 4, 5, 0, time.UTC)
+	announce := func(peer string) *ristypes.RISMessage {
+		return &ristypes.RISMessage{
+			Peer:            peer,
+			Path:            ristypes.ASPath{testASNA, testASNB},
+			ParsedTimestamp: ts,
+			Announcements:   []ristypes.RISAnnouncement{{NextHop: peer, Prefixes: []string{prefix}}},
+		}
+	}
+	withdraw := func(peer string) *ristypes.RISMessage {
+		return &ristypes.RISMessage{Peer: peer, ParsedTimestamp: ts, Withdrawals: []string{prefix}}
+	}
+	announceAndWithdraw := func(peer string) *ristypes.RISMessage {
+		msg := announce(peer)
+		msg.Withdrawals = []string{prefix}
+
+		return msg
+	}
+
+	tests := []struct {
+		name string
+		// Each inner list is handled and then flushed as one batch.
+		batches [][]*ristypes.RISMessage
+		// The peers left with a live route for the prefix, by peer IP.
+		wantPeers []string
+	}{
+		{
+			name:      "two peers announce the prefix",
+			batches:   [][]*ristypes.RISMessage{{announce(testPeerIP), announce(otherPeerIP)}},
+			wantPeers: []string{testPeerIP, otherPeerIP},
+		},
+		{
+			name: "two peers withdraw the prefix",
+			batches: [][]*ristypes.RISMessage{
+				{announce(testPeerIP)},
+				{announce(otherPeerIP)},
+				{withdraw(testPeerIP), withdraw(otherPeerIP)},
+			},
+		},
+		{
+			name:    "a peer announces and then withdraws the prefix",
+			batches: [][]*ristypes.RISMessage{{announce(testPeerIP), withdraw(testPeerIP)}},
+		},
+		{
+			name: "a peer withdraws and then re-announces the prefix",
+			batches: [][]*ristypes.RISMessage{
+				{announce(testPeerIP)},
+				{withdraw(testPeerIP), announce(testPeerIP)},
+			},
+			wantPeers: []string{testPeerIP},
+		},
+		{
+			// RFC 4271 section 4.3: the announcement counts, not the withdrawal.
+			name: "one message announces and withdraws the prefix",
+			batches: [][]*ristypes.RISMessage{
+				{announce(testPeerIP)},
+				{announceAndWithdraw(testPeerIP)},
+			},
+			wantPeers: []string{testPeerIP},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			db, err := database.New(&config.Config{StateDir: t.TempDir()}, logger.New())
+			if err != nil {
+				t.Fatalf("failed to create database: %v", err)
+			}
+			defer func() { _ = db.Close() }()
+
+			// Built without NewPrefixHandler's flush timer, so the test flushes itself.
+			h := &PrefixHandler{db: db, logger: logger.New()}
+			for _, batch := range tt.batches {
+				for _, msg := range batch {
+					h.HandleMessage(msg)
+				}
+				h.flushBatch()
+			}
+
+			ctx := context.Background()
+			routes, err := db.GetPrefixDetailsContext(ctx, prefix)
+			if err != nil && !errors.Is(err, database.ErrNoRoute) {
+				t.Fatalf("GetPrefixDetailsContext: %v", err)
+			}
+			var gotPeers []string
+			for _, route := range routes {
+				gotPeers = append(gotPeers, route.PeerIP)
+			}
+			if !slices.Equal(gotPeers, tt.wantPeers) {
+				t.Errorf("peers with a live route for %s = %v, want %v", prefix, gotPeers, tt.wantPeers)
+			}
+
+			stats, err := db.GetStatsContext(ctx)
+			if err != nil {
+				t.Fatalf("GetStatsContext: %v", err)
+			}
+			tableRoutesV4, _, err := db.GetLiveRouteCountsContext(ctx)
+			if err != nil {
+				t.Fatalf("GetLiveRouteCountsContext: %v", err)
+			}
+			if stats.IPv4Routes != len(tt.wantPeers) || tableRoutesV4 != len(tt.wantPeers) {
+				t.Errorf("IPv4 live routes: %d in memory, %d in the table, want %d",
+					stats.IPv4Routes, tableRoutesV4, len(tt.wantPeers))
+			}
+
+			var wantDistribution []database.PrefixDistribution
+			if len(tt.wantPeers) > 0 {
+				wantDistribution = []database.PrefixDistribution{{MaskLength: prefixMaskLength, Count: 1}}
+			}
+			queryDistributionV4, _, err := db.GetPrefixDistributionContext(ctx)
+			if err != nil {
+				t.Fatalf("GetPrefixDistributionContext: %v", err)
+			}
+			if !slices.Equal(stats.IPv4PrefixDistribution, wantDistribution) ||
+				!slices.Equal(queryDistributionV4, wantDistribution) {
+				t.Errorf("IPv4 prefix distribution: %v in memory, %v from the table, want %v",
+					stats.IPv4PrefixDistribution, queryDistributionV4, wantDistribution)
+			}
+		})
+	}
+}

internal/server/handlers.go

@@ -7,6 +7,7 @@ import (
 	"errors"
 	"net"
 	"net/http"
+	"net/netip"
 	"runtime"
 	"sort"
 	"strconv"
@@ -748,8 +749,15 @@ func (s *Server) handlePrefixDetailJSON() http.HandlerFunc {
 			return
 		}
 
-		// Combine prefix and length into CIDR notation
-		prefix := prefixParam + "/" + lenParam
+		// Look the prefix up in the text form net/netip prints, the form the
+		// tables store, so that any valid way of writing it finds the route.
+		parsed, err := netip.ParsePrefix(prefixParam + "/" + lenParam)
+		if err != nil {
+			writeJSONError(w, http.StatusBadRequest, err.Error())
+
+			return
+		}
+		prefix := parsed.Masked().String()
 
 		routes, err := s.db.GetPrefixDetailsContext(r.Context(), prefix)
 		if err != nil {
@@ -912,8 +920,15 @@ func (s *Server) handlePrefixDetail() http.HandlerFunc {
 			return
 		}
 
-		// Combine prefix and length into CIDR notation
-		prefix := prefixParam + "/" + lenParam
+		// Look the prefix up in the text form net/netip prints, the form the
+		// tables store, so that any valid way of writing it finds the route.
+		parsed, err := netip.ParsePrefix(prefixParam + "/" + lenParam)
+		if err != nil {
+			http.Error(w, "Invalid prefix", http.StatusBadRequest)
+
+			return
+		}
+		prefix := parsed.Masked().String()
 
 		routes, err := s.db.GetPrefixDetailsContext(r.Context(), prefix)
 		if err != nil {

internal/server/handlers_test.go

@@ -148,6 +148,58 @@ func TestStatsHandlersAnswerFromMemory(t *testing.T) {
 	}
 }
 
+// TestPrefixDetailHandlersParsePrefix checks that both prefix pages find a
+// stored route whatever valid text form the URL writes its prefix in, and
+// answer 400 for a prefix that does not parse. The tables store the form
+// net/netip prints. Before the fix the URL text was looked up as typed, so
+// "2001:DB8:0:0::/32" was not found although "2001:db8::/32" was stored, and
+// "foo/24" answered 500 (https://git.eeqj.de/sneak/routewatch/issues/71).
+func TestPrefixDetailHandlersParsePrefix(t *testing.T) {
+	db, err := database.New(&config.Config{StateDir: t.TempDir()}, logger.New())
+	if err != nil {
+		t.Fatalf("database.New: %v", err)
+	}
+	t.Cleanup(func() { _ = db.Close() })
+
+	ts := time.Date(2026, 1, 2, 3, 4, 5, 0, time.UTC)
+	if err := db.UpsertLiveRouteBatch([]*database.LiveRoute{
+		{
+			ID: uuid.New(), Prefix: "192.0.2.0/24", MaskLength: 24, IPVersion: 4,
+			OriginASN: 64500, PeerIP: "198.51.100.1", ASPath: []int{64500}, NextHop: "198.51.100.1",
+			LastUpdated: ts,
+		},
+		{
+			ID: uuid.New(), Prefix: "2001:db8::/32", MaskLength: 32, IPVersion: 6,
+			OriginASN: 64501, PeerIP: "2001:db8::1", ASPath: []int{64501}, NextHop: "2001:db8::1",
+			LastUpdated: ts,
+		},
+	}); err != nil {
+		t.Fatalf("UpsertLiveRouteBatch: %v", err)
+	}
+
+	s := New(db, streamer.New(logger.New(), metrics.New()), logger.New(), &config.Config{})
+
+	tests := []struct {
+		path string
+		want int
+	}{
+		{"/api/v1/prefix/foo/24", http.StatusBadRequest},
+		{"/prefix/foo/24", http.StatusBadRequest},
+		{"/api/v1/prefix/2001:DB8:0:0::/32", http.StatusOK},
+		{"/prefix/2001:DB8:0:0::/32", http.StatusOK},
+		{"/api/v1/prefix/192.0.2.1/24", http.StatusOK},
+		{"/prefix/192.0.2.1/24", http.StatusOK},
+	}
+
+	for _, tt := range tests {
+		rec := httptest.NewRecorder()
+		s.router.ServeHTTP(rec, httptest.NewRequest(http.MethodGet, tt.path, nil))
+		if rec.Code != tt.want {
+			t.Errorf("%s: status %d, want %d; body %s", tt.path, rec.Code, tt.want, rec.Body)
+		}
+	}
+}
+
 // settledGoroutineCount lets transient goroutines finish, then reports the
 // current count.
 func settledGoroutineCount() int {