Sample the prefix-length pages with bounded lookups, each prefix once (closes #66)

/prefixlength/{length} and /prefixlength6/{length} grouped every live route
of that length by prefix, shuffled all the prefixes and joined the routes
back, so a request grew with the number of routes and listed a prefix once
for each peer that shared its newest time.

Now a length with at most 5000 prefixes has them all read and shuffled, and
the page lists 500 of them. A longer one is sampled with random row-id
lookups on the mask length index, at most four per prefix asked for, filled
up from the shuffled prefixes if the lookups keep repeating. Each prefix is
shown once, with its newest route.

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
 

internal/database/database.go

@@ -8,6 +8,7 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"math/rand/v2"
 	"net"
 	"net/netip"
 	"os"
@@ -37,6 +38,14 @@ 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 (
@@ -53,6 +62,17 @@ const (
 	sqliteHardHeapLimitBytes = 1610612736
 )
 
+// Bounds on the work GetRandomPrefixesByLength does to pick its sample.
+const (
+	// maxPrefixesToShuffle is the most prefixes of one mask length that are
+	// read in full and shuffled to pick the sample. A mask length with more is
+	// sampled with random lookups.
+	maxPrefixesToShuffle = 5000
+	// sampleLookupsPerPrefix is the most random lookups made for each prefix
+	// asked for.
+	sampleLookupsPerPrefix = 4
+)
+
 // Common errors
 var (
 	// ErrInvalidIP is returned when an IP address is malformed
@@ -1696,45 +1716,183 @@ func (d *Database) GetRandomPrefixesByLength(maskLength, ipVersion, limit int) (
 	return d.GetRandomPrefixesByLengthContext(context.Background(), maskLength, ipVersion, limit)
 }
 
-// GetRandomPrefixesByLengthContext returns a random sample of prefixes with context support
+// GetRandomPrefixesByLengthContext returns limit prefixes with the given mask
+// length picked at random, or all of them when there are fewer, each once with
+// its most recently updated live route, for a limit up to
+// maxPrefixesToShuffle. However many routes there are, it reads at most
+// maxPrefixesToShuffle+1 prefixes and makes at most sampleLookupsPerPrefix
+// lookups for each prefix asked for.
 func (d *Database) GetRandomPrefixesByLengthContext(
 	ctx context.Context, maskLength, ipVersion, limit int) ([]LiveRoute, error) {
-	// Select unique prefixes with their most recent route information
 	tableName := "live_routes_v4"
 	if ipVersion == ipVersionV6 {
 		tableName = "live_routes_v6"
 	}
 
+	// Reading one prefix more than maxPrefixesToShuffle tells whether these
+	// are all the prefixes with this mask length.
+	prefixes, err := d.firstPrefixesByLength(ctx, tableName, maskLength, maxPrefixesToShuffle+1)
+	if err != nil {
+		return nil, err
+	}
+	rand.Shuffle(len(prefixes), func(i, j int) {
+		prefixes[i], prefixes[j] = prefixes[j], prefixes[i]
+	})
+	if len(prefixes) > maxPrefixesToShuffle {
+		prefixes, err = d.samplePrefixesByLength(ctx, tableName, maskLength, limit, prefixes)
+		if err != nil {
+			return nil, err
+		}
+	}
+	prefixes = prefixes[:min(limit, len(prefixes))]
+
+	return d.newestRouteOfEachPrefix(ctx, tableName, ipVersion, prefixes)
+}
+
+// firstPrefixesByLength returns the first n distinct prefixes, in text order,
+// with the given mask length. It reads only their entries in the mask length
+// and prefix index.
+func (d *Database) firstPrefixesByLength(
+	ctx context.Context, tableName string, maskLength, n int) ([]string, error) {
 	//nolint:gosec // Table name is hardcoded based on IP version
 	query := fmt.Sprintf(`
-		WITH unique_prefixes AS (
-			SELECT prefix, MAX(last_updated) as max_updated
-			FROM %s
-			WHERE mask_length = ?
-			GROUP BY prefix
-			ORDER BY RANDOM()
-			LIMIT ?
-		)
-		SELECT lr.prefix, lr.mask_length, lr.origin_asn, lr.as_path, 
-			lr.peer_ip, lr.last_updated
-		FROM %s lr
-		INNER JOIN unique_prefixes up ON lr.prefix = up.prefix AND lr.last_updated = up.max_updated
-		WHERE lr.mask_length = ?
-	`, tableName, tableName)
+		SELECT DISTINCT prefix FROM %s
+		WHERE mask_length = ?
+		ORDER BY prefix
+		LIMIT ?
+	`, tableName)
 
-	rows, err := d.db.QueryContext(ctx, query, maskLength, limit, maskLength)
+	rows, err := d.db.QueryContext(ctx, query, maskLength, n)
 	if err != nil {
-		return nil, fmt.Errorf("failed to query random prefixes: %w", err)
+		return nil, fmt.Errorf("failed to list prefixes: %w", err)
 	}
-	defer func() {
-		_ = rows.Close()
-	}()
+	defer func() { _ = rows.Close() }()
 
-	var routes []LiveRoute
+	var prefixes []string
 	for rows.Next() {
-		var route LiveRoute
+		var prefix string
+		if err := rows.Scan(&prefix); err != nil {
+			return nil, fmt.Errorf("failed to scan prefix: %w", err)
+		}
+		prefixes = append(prefixes, prefix)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, fmt.Errorf("failed to list prefixes: %w", err)
+	}
+
+	return prefixes, nil
+}
+
+// samplePrefixesByLength picks limit distinct prefixes with the given mask
+// length at random, in at most sampleLookupsPerPrefix lookups for each prefix
+// asked for. ORDER BY RANDOM() would read every route with that mask length.
+// Instead each lookup draws a row id between those of the first and last route
+// with that mask length and takes the prefix of the first such route at or
+// after it. SQLite keeps the row id at the end of every index entry, so the
+// mask length index is ordered by mask length, then row id, and the lookup is
+// one search of it.
+//
+// The sample is uneven, as the page needs a varied sample, not an even one: a
+// route is picked in proportion to the row ids between it and the previous
+// route with that mask length, including the row ids of routes with other mask
+// lengths and of withdrawn routes. A route that follows a long run of
+// withdrawn routes is picked far more often than one that directly follows
+// another route with that mask length.
+//
+// Should the lookups keep landing on prefixes already picked, the rest are
+// taken from shuffled, prefixes of that length in random order.
+func (d *Database) samplePrefixesByLength(
+	ctx context.Context, tableName string, maskLength, limit int, shuffled []string) ([]string, error) {
+	//nolint:gosec // Table name is hardcoded based on IP version
+	rangeQuery := fmt.Sprintf(`
+		SELECT
+			(SELECT MIN(rowid) FROM %[1]s WHERE mask_length = ?),
+			(SELECT MAX(rowid) FROM %[1]s WHERE mask_length = ?)
+	`, tableName)
+
+	var first, last sql.NullInt64
+	if err := d.db.QueryRowContext(ctx, rangeQuery, maskLength, maskLength).Scan(&first, &last); err != nil {
+		return nil, fmt.Errorf("failed to read row id range: %w", err)
+	}
+	if !first.Valid {
+		// Every route with this mask length was withdrawn since they were listed.
+		return nil, nil
+	}
+
+	//nolint:gosec // Table name is hardcoded based on IP version
+	lookupQuery := fmt.Sprintf(`
+		SELECT prefix FROM %s
+		WHERE mask_length = ? AND rowid >= ?
+		ORDER BY rowid
+		LIMIT 1
+	`, tableName)
+
+	stmt, err := d.db.PrepareContext(ctx, lookupQuery)
+	if err != nil {
+		return nil, fmt.Errorf("failed to prepare prefix lookup: %w", err)
+	}
+	defer func() { _ = stmt.Close() }()
+
+	picked := make(map[string]bool, limit)
+	prefixes := make([]string, 0, limit)
+	pick := func(prefix string) {
+		if !picked[prefix] {
+			picked[prefix] = true
+			prefixes = append(prefixes, prefix)
+		}
+	}
+
+	for lookups := 0; lookups < limit*sampleLookupsPerPrefix && len(prefixes) < limit; lookups++ {
+		//nolint:gosec // The sample needs variety, not unpredictability.
+		rowID := first.Int64 + rand.Int64N(last.Int64-first.Int64+1)
+
+		var prefix string
+		err := stmt.QueryRowContext(ctx, maskLength, rowID).Scan(&prefix)
+		if errors.Is(err, sql.ErrNoRows) {
+			// The routes from rowID on were withdrawn since the range was read.
+			continue
+		}
+		if err != nil {
+			return nil, fmt.Errorf("failed to look up prefix: %w", err)
+		}
+		pick(prefix)
+	}
+
+	for _, prefix := range shuffled {
+		if len(prefixes) == limit {
+			break
+		}
+		pick(prefix)
+	}
+
+	return prefixes, nil
+}
+
+// newestRouteOfEachPrefix returns the most recently updated live route of each
+// prefix, one lookup on the prefix index per prefix. A prefix whose routes were
+// all withdrawn since it was picked is left out.
+func (d *Database) newestRouteOfEachPrefix(
+	ctx context.Context, tableName string, ipVersion int, prefixes []string) ([]LiveRoute, error) {
+	//nolint:gosec // Table name is hardcoded based on IP version
+	query := fmt.Sprintf(`
+		SELECT prefix, mask_length, origin_asn, as_path, peer_ip, last_updated
+		FROM %s
+		WHERE prefix = ?
+		ORDER BY last_updated DESC
+		LIMIT 1
+	`, tableName)
+
+	stmt, err := d.db.PrepareContext(ctx, query)
+	if err != nil {
+		return nil, fmt.Errorf("failed to prepare route query: %w", err)
+	}
+	defer func() { _ = stmt.Close() }()
+
+	routes := make([]LiveRoute, 0, len(prefixes))
+	for _, prefix := range prefixes {
+		route := LiveRoute{IPVersion: ipVersion}
 		var pathJSON string
-		err := rows.Scan(
+		err := stmt.QueryRowContext(ctx, prefix).Scan(
 			&route.Prefix,
 			&route.MaskLength,
 			&route.OriginASN,
@@ -1742,17 +1900,16 @@ func (d *Database) GetRandomPrefixesByLengthContext(
 			&route.PeerIP,
 			&route.LastUpdated,
 		)
-		// Set IP version based on which table we queried
-		route.IPVersion = ipVersion
-		if err != nil {
+		if errors.Is(err, sql.ErrNoRows) {
 			continue
 		}
+		if err != nil {
+			return nil, fmt.Errorf("failed to query route of %s: %w", prefix, err)
+		}
 
-		// Decode AS path
 		if err := json.Unmarshal([]byte(pathJSON), &route.ASPath); err != nil {
 			route.ASPath = []int{}
 		}
-
 		routes = append(routes, route)
 	}
 
@@ -1876,11 +2033,12 @@ 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 +2049,11 @@ 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 +2094,7 @@ 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

@@ -6,6 +6,7 @@ import (
 	"errors"
 	"fmt"
 	"net/netip"
+	"slices"
 	"sync"
 	"testing"
 	"time"
@@ -45,6 +46,32 @@ const (
 	asnSecondBand = 100
 )
 
+// Parameters for the prefix-length sample tests.
+const (
+	// sampleLimit is how many prefixes the tests ask the sample for.
+	sampleLimit = 10
+	// sampledPrefixCount is how many prefixes the sampling test stores, more
+	// than sampleLimit so the sample has to choose among them.
+	sampledPrefixCount = 30
+	// secondV4PeerIP and secondV6PeerIP are the second peer each prefix is
+	// learned from, besides the one mkV4Route and mkV6Route set.
+	secondV4PeerIP = "192.0.2.2"
+	secondV6PeerIP = "2001:db8::2"
+	// largeSamplePrefixCount is how many prefixes the test of a mask length
+	// with more than maxPrefixesToShuffle prefixes stores: twice that, so about
+	// half of them come after the first maxPrefixesToShuffle+1 in text order.
+	largeSamplePrefixCount = 2 * maxPrefixesToShuffle
+	// largeSampleLimit is how many prefixes that test asks for. Each random
+	// lookup lands after the first maxPrefixesToShuffle+1 in text order about
+	// half the time, so none of them doing so is about as likely as
+	// largeSampleLimit coin tosses all coming up heads.
+	largeSampleLimit = 30
+	// liveRoutesV4Table and liveRoutesV6Table are the live route tables the
+	// tests pass to the sampling steps.
+	liveRoutesV4Table = "live_routes_v4"
+	liveRoutesV6Table = "live_routes_v6"
+)
+
 // TestGetIPInfoFindsMostSpecificLiveRoute stores nested live prefixes for both
 // families and checks that a lookup returns the most specific one covering the
 // address, ErrNoRoute when none covers it, and the next less specific prefix
@@ -133,6 +160,78 @@ 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
@@ -317,3 +416,281 @@ func TestBatchWriteDuringCheckpoint(t *testing.T) {
 	cancel()
 	wg.Wait()
 }
+
+// TestGetRandomPrefixesByLengthListsPrefixOnce stores one prefix of each
+// family learned from two peers in the same second and checks that the
+// prefix-length sample lists each of them once.
+func TestGetRandomPrefixesByLengthListsPrefixOnce(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)
+	v4First := mkV4Route(t, "198.51.100.0/24", 64500, ts)
+	v4Second := mkV4Route(t, "198.51.100.0/24", 64500, ts)
+	v4Second.PeerIP = secondV4PeerIP
+	v6First := mkV6Route("2001:db8::/32", 64500, ts)
+	v6Second := mkV6Route("2001:db8::/32", 64500, ts)
+	v6Second.PeerIP = secondV6PeerIP
+	if err := db.UpsertLiveRouteBatch([]*LiveRoute{v4First, v4Second, v6First, v6Second}); err != nil {
+		t.Fatalf("UpsertLiveRouteBatch: %v", err)
+	}
+
+	for _, want := range []*LiveRoute{v4First, v6First} {
+		routes, err := db.GetRandomPrefixesByLength(want.MaskLength, want.IPVersion, sampleLimit)
+		if err != nil {
+			t.Fatalf("GetRandomPrefixesByLength(%d, %d): %v", want.MaskLength, want.IPVersion, err)
+		}
+		if len(routes) != 1 || routes[0].Prefix != want.Prefix {
+			t.Errorf("GetRandomPrefixesByLength(%d, %d) = %+v, want %s listed once",
+				want.MaskLength, want.IPVersion, routes, want.Prefix)
+		}
+	}
+}
+
+// TestGetRandomPrefixesByLengthListsNewestRoute stores a prefix learned from
+// two peers at different times, the older route first, and checks that the
+// prefix-length sample lists the newer route.
+func TestGetRandomPrefixesByLengthListsNewestRoute(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() }()
+
+	older := time.Date(2026, 1, 2, 3, 4, 5, 0, time.UTC)
+	olderRoute := mkV4Route(t, "198.51.100.0/24", 64500, older)
+	newerRoute := mkV4Route(t, "198.51.100.0/24", 64500, older.Add(time.Minute))
+	newerRoute.PeerIP = secondV4PeerIP
+	if err := db.UpsertLiveRouteBatch([]*LiveRoute{olderRoute, newerRoute}); err != nil {
+		t.Fatalf("UpsertLiveRouteBatch: %v", err)
+	}
+
+	routes, err := db.GetRandomPrefixesByLength(newerRoute.MaskLength, ipVersionV4, sampleLimit)
+	if err != nil {
+		t.Fatalf("GetRandomPrefixesByLength: %v", err)
+	}
+	if len(routes) != 1 || routes[0].PeerIP != newerRoute.PeerIP {
+		t.Errorf("GetRandomPrefixesByLength = %+v, want only the newer route, from %s",
+			routes, newerRoute.PeerIP)
+	}
+}
+
+// TestGetRandomPrefixesByLengthSamplesManyPrefixes stores more /24 prefixes
+// than the sample asks for, each followed by a /16 so that routes of another
+// mask length sit between the /24 rows, and each prefix learned from two peers
+// in the same second. It checks both the sample the pages get, picked from the
+// shuffled prefixes, and the random lookups used for a mask length with more
+// than maxPrefixesToShuffle prefixes.
+func TestGetRandomPrefixesByLengthSamplesManyPrefixes(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)
+	var routes []*LiveRoute
+	var stored []string
+	for i := range sampledPrefixCount {
+		prefix := fmt.Sprintf("10.0.%d.0/24", i)
+		for _, p := range []string{prefix, fmt.Sprintf("172.%d.0.0/16", i)} {
+			first := mkV4Route(t, p, 64500, ts)
+			second := mkV4Route(t, p, 64500, ts)
+			second.PeerIP = secondV4PeerIP
+			routes = append(routes, first, second)
+		}
+		stored = append(stored, prefix)
+	}
+	if err := db.UpsertLiveRouteBatch(routes); err != nil {
+		t.Fatalf("UpsertLiveRouteBatch: %v", err)
+	}
+	slices.Sort(stored)
+
+	maskLength := routes[0].MaskLength
+	sample, err := db.GetRandomPrefixesByLength(maskLength, ipVersionV4, sampleLimit)
+	if err != nil {
+		t.Fatalf("GetRandomPrefixesByLength: %v", err)
+	}
+	var listed []string
+	for _, route := range sample {
+		listed = append(listed, route.Prefix)
+	}
+	checkPrefixSample(t, "GetRandomPrefixesByLength", listed, stored)
+
+	looked, err := db.samplePrefixesByLength(
+		context.Background(), liveRoutesV4Table, maskLength, sampleLimit, nil)
+	if err != nil {
+		t.Fatalf("samplePrefixesByLength: %v", err)
+	}
+	checkPrefixSample(t, "samplePrefixesByLength", looked, stored)
+}
+
+// checkPrefixSample checks that sample lists sampleLimit prefixes, each once
+// and each one of stored, and that it is not just the first sampleLimit of
+// stored, which is in text order.
+func checkPrefixSample(t *testing.T, name string, sample, stored []string) {
+	t.Helper()
+
+	if len(sample) != sampleLimit {
+		t.Errorf("%s listed %d prefixes, want %d", name, len(sample), sampleLimit)
+	}
+	listed := make(map[string]bool)
+	for _, prefix := range sample {
+		if listed[prefix] || !slices.Contains(stored, prefix) {
+			t.Errorf("%s lists %s twice or it is not one of the /24s", name, prefix)
+		}
+		listed[prefix] = true
+	}
+	if slices.Equal(slices.Sorted(slices.Values(sample)), stored[:sampleLimit]) {
+		t.Errorf("%s lists just the first %d prefixes in text order", name, sampleLimit)
+	}
+}
+
+// TestGetRandomPrefixesByLengthSamplesPastFirstPrefixes stores more prefixes
+// of one mask length than maxPrefixesToShuffle. It checks that
+// firstPrefixesByLength reads just the first maxPrefixesToShuffle+1 in text
+// order, and that the sample lists the prefixes asked for, each once, some of
+// them after those first ones, which only the random lookups reach.
+func TestGetRandomPrefixesByLengthSamplesPastFirstPrefixes(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)
+	var routes []*LiveRoute
+	var stored []string
+	for i := range largeSamplePrefixCount {
+		route := mkV6Route(fmt.Sprintf("2001:%x::/32", i), 64500, ts)
+		routes = append(routes, route)
+		stored = append(stored, route.Prefix)
+	}
+	if err := db.UpsertLiveRouteBatch(routes); err != nil {
+		t.Fatalf("UpsertLiveRouteBatch: %v", err)
+	}
+	slices.Sort(stored)
+	firstPrefixes := stored[:maxPrefixesToShuffle+1]
+
+	maskLength := routes[0].MaskLength
+	read, err := db.firstPrefixesByLength(
+		context.Background(), liveRoutesV6Table, maskLength, maxPrefixesToShuffle+1)
+	if err != nil {
+		t.Fatalf("firstPrefixesByLength: %v", err)
+	}
+	if !slices.Equal(read, firstPrefixes) {
+		t.Errorf("firstPrefixesByLength read %d prefixes, want the first %d in text order",
+			len(read), len(firstPrefixes))
+	}
+
+	sample, err := db.GetRandomPrefixesByLength(maskLength, ipVersionV6, largeSampleLimit)
+	if err != nil {
+		t.Fatalf("GetRandomPrefixesByLength: %v", err)
+	}
+	if len(sample) != largeSampleLimit {
+		t.Errorf("GetRandomPrefixesByLength listed %d prefixes, want %d", len(sample), largeSampleLimit)
+	}
+	listed := make(map[string]bool)
+	pastFirstPrefixes := 0
+	for _, route := range sample {
+		if listed[route.Prefix] {
+			t.Errorf("GetRandomPrefixesByLength lists %s twice", route.Prefix)
+		}
+		listed[route.Prefix] = true
+		if !slices.Contains(firstPrefixes, route.Prefix) {
+			pastFirstPrefixes++
+		}
+	}
+	if pastFirstPrefixes == 0 {
+		t.Errorf("GetRandomPrefixesByLength lists only prefixes among the first %d in text order",
+			len(firstPrefixes))
+	}
+}
+
+// TestSamplePrefixesByLengthFillsUpFromShuffled stores a single prefix, so
+// every random lookup lands on it, and checks that the sample still lists
+// sampleLimit prefixes: the stored one, then the shuffled ones in order. The
+// stored prefix is the second shuffled one, so the fill-up must skip it.
+func TestSamplePrefixesByLengthFillsUpFromShuffled(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() }()
+
+	var shuffled []string
+	for i := range sampledPrefixCount {
+		shuffled = append(shuffled, fmt.Sprintf("10.0.%d.0/24", i))
+	}
+	route := mkV4Route(t, shuffled[1], 64500, time.Date(2026, 1, 2, 3, 4, 5, 0, time.UTC))
+	if err := db.UpsertLiveRouteBatch([]*LiveRoute{route}); err != nil {
+		t.Fatalf("UpsertLiveRouteBatch: %v", err)
+	}
+
+	sample, err := db.samplePrefixesByLength(
+		context.Background(), liveRoutesV4Table, route.MaskLength, sampleLimit, shuffled)
+	if err != nil {
+		t.Fatalf("samplePrefixesByLength: %v", err)
+	}
+	want := append([]string{shuffled[1], shuffled[0]}, shuffled[2:sampleLimit]...)
+	if !slices.Equal(sample, want) {
+		t.Errorf("samplePrefixesByLength = %v, want %v", sample, want)
+	}
+}
+
+// TestGetRandomPrefixesByLengthSkipsWithdrawnRoutes withdraws a route after
+// storing it, as can happen between the steps of GetRandomPrefixesByLength. It
+// checks that the random lookups pick nothing for a mask length with no route
+// left, and that only the prefixes that still have a route get their newest
+// route listed.
+func TestGetRandomPrefixesByLengthSkipsWithdrawnRoutes(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)
+	kept := mkV4Route(t, "198.51.100.0/24", 64500, ts)
+	withdrawn := mkV4Route(t, "172.16.0.0/16", 64500, ts)
+	if err := db.UpsertLiveRouteBatch([]*LiveRoute{kept, withdrawn}); err != nil {
+		t.Fatalf("UpsertLiveRouteBatch: %v", err)
+	}
+	if err := db.DeleteLiveRoute(withdrawn.Prefix, withdrawn.OriginASN, withdrawn.PeerIP); err != nil {
+		t.Fatalf("DeleteLiveRoute: %v", err)
+	}
+
+	ctx := context.Background()
+	sample, err := db.samplePrefixesByLength(
+		ctx, liveRoutesV4Table, withdrawn.MaskLength, sampleLimit, []string{withdrawn.Prefix})
+	if err != nil {
+		t.Fatalf("samplePrefixesByLength: %v", err)
+	}
+	if len(sample) != 0 {
+		t.Errorf("samplePrefixesByLength = %v, want none for a mask length with no route left", sample)
+	}
+
+	routes, err := db.newestRouteOfEachPrefix(
+		ctx, liveRoutesV4Table, ipVersionV4, []string{withdrawn.Prefix, kept.Prefix})
+	if err != nil {
+		t.Fatalf("newestRouteOfEachPrefix: %v", err)
+	}
+	if len(routes) != 1 || routes[0].Prefix != kept.Prefix {
+		t.Errorf("newestRouteOfEachPrefix = %+v, want only the route of %s", routes, kept.Prefix)
+	}
+}

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/cli.go

@@ -20,6 +20,8 @@ const (
 	debugInterval = 60 * time.Second
 	// bytesPerMB is bytes per megabyte
 	bytesPerMB = 1024 * 1024
+	// failureExitCode is the exit status when RouteWatch stops on an error
+	failureExitCode = 1
 )
 
 // logDebugStats logs goroutine count and memory usage
@@ -83,6 +85,9 @@ func CLIEntry() {
 					go func() {
 						if err := rw.Run(ctx); err != nil {
 							logger.Error("RouteWatch error", "error", err)
+							if err := shutdowner.Shutdown(fx.ExitCode(failureExitCode)); err != nil {
+								logger.Error("Failed to shutdown gracefully", "error", err)
+							}
 						}
 					}()
 

internal/routewatch/prefixhandler.go

@@ -105,7 +105,24 @@ 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{
@@ -123,22 +140,6 @@ 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
 	if len(h.batch) >= prefixBatchSize {
 		h.flushBatchLocked()
@@ -179,14 +180,14 @@ func (h *PrefixHandler) flushBatchLocked() {
 	startTime := time.Now()
 	batchSize := len(h.batch)
 
-	// Group updates by prefix to deduplicate
-	// For each prefix, keep the latest update
-	prefixMap := make(map[string]prefixUpdate)
+	// 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 h.batch {
-		key := update.prefix
-		if existing, ok := prefixMap[key]; !ok || update.timestamp.After(existing.timestamp) {
-			prefixMap[key] = update
-		}
+		lastUpdates[prefixAndPeer{update.prefix, update.peer}] = update
 	}
 
 	// Collect routes to upsert and delete
@@ -196,7 +197,7 @@ 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
@@ -255,7 +256,7 @@ 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(),
 	)

internal/routewatch/prefixhandler_test.go

@@ -1,7 +1,9 @@
 package routewatch
 
 import (
+	"context"
 	"errors"
+	"slices"
 	"testing"
 	"time"
 
@@ -72,3 +74,140 @@ 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.mu.Lock()
+				h.flushBatchLocked()
+				h.mu.Unlock()
+			}
+
+			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 {

internal/server/server.go

@@ -3,6 +3,8 @@ package server
 
 import (
 	"context"
+	"fmt"
+	"net"
 	"net/http"
 	"strconv"
 	"time"
@@ -53,9 +55,12 @@ func New(db database.Store, streamer *streamer.Streamer, logger *logger.Logger,
 	return s
 }
 
-// Start starts the HTTP server
+// Start listens on the configured port and serves HTTP in the background.
+// It returns an error if it cannot listen, for example because the port is
+// already in use.
 func (s *Server) Start() error {
 	const (
+		network           = "tcp"
 		readHeaderTimeout = 40 * time.Second
 		readTimeout       = 60 * time.Second
 		writeTimeout      = 60 * time.Second
@@ -73,10 +78,15 @@ func (s *Server) Start() error {
 
 	s.logger.Info("Starting HTTP server", "port", s.port, "addr", s.srv.Addr)
 
-	// Start in goroutine but log when actually listening
+	listener, err := net.Listen(network, s.srv.Addr)
+	if err != nil {
+		return fmt.Errorf("starting HTTP server: %w", err)
+	}
+
+	s.logger.Info("HTTP server listening", "addr", listener.Addr().String())
+
 	go func() {
-		s.logger.Info("HTTP server listening", "addr", s.srv.Addr)
-		if err := s.srv.ListenAndServe(); err != nil && err != http.ErrServerClosed {
+		if err := s.srv.Serve(listener); err != nil && err != http.ErrServerClosed {
 			s.logger.Error("HTTP server error", "error", err)
 		}
 	}()

internal/server/server_test.go

@@ -0,0 +1,33 @@
+package server
+
+import (
+	"context"
+	"net"
+	"testing"
+
+	"git.eeqj.de/sneak/routewatch/internal/config"
+	"git.eeqj.de/sneak/routewatch/internal/logger"
+	"git.eeqj.de/sneak/routewatch/internal/metrics"
+	"git.eeqj.de/sneak/routewatch/internal/streamer"
+)
+
+// TestStartFailsWhenPortIsHeld holds a port with a listener of its own and
+// checks that Start on that port returns an error. Before the fix Start
+// listened in a goroutine, only logged the failure and returned nil, so the
+// daemon kept running without its web server.
+func TestStartFailsWhenPortIsHeld(t *testing.T) {
+	held, err := net.Listen("tcp", ":0")
+	if err != nil {
+		t.Fatalf("failed to open the test listener: %v", err)
+	}
+	defer func() { _ = held.Close() }()
+
+	port := held.Addr().(*net.TCPAddr).Port
+	log := logger.New()
+	s := New(nil, streamer.New(log, metrics.New()), log, &config.Config{Port: port})
+	defer func() { _ = s.Stop(context.Background()) }()
+
+	if err := s.Start(); err == nil {
+		t.Fatalf("Start returned nil while port %d was held by another listener", port)
+	}
+}