Store the routing table in the database (part of #122)

A new routes table holds one row per prefix and origin AS: the peer set
as the bitset's bytes, and the two kept routes, each with its peer's
number, AS path and time in seconds. A new route_peers table holds each
peer's number and address.

WriteRouteChanges writes one TakeChanges result in a transaction.
ReadRouteTable reads both tables into a new routing table, through the
new Table.Load, which records no change. Nothing uses them yet.

Model: opus-5-5

internal/database/interface.go

@@ -1,5 +1,7 @@
 // Package database provides SQLite storage for BGP routing data including ASNs,
-// prefixes, announcements, peerings, and live route tables.
+// prefixes, announcements, peerings, and live route tables. Its routes table
+// holds a copy of the whole routing table that internal/routetable keeps in
+// memory.
 package database
 
 import (

internal/database/routes.go

@@ -0,0 +1,326 @@
+package database
+
+import (
+	"context"
+	"database/sql"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"net/netip"
+	"time"
+
+	"github.com/bits-and-blooms/bitset"
+
+	"git.eeqj.de/sneak/routewatch/internal/routetable"
+)
+
+// keptRoutesPerRow is how many kept routes a row of the routes table has
+// columns for: the newest and the older one.
+const keptRoutesPerRow = 2
+
+// errUnknownRoutePeer is returned for a kept route whose peer has no number in
+// route_peers.
+var errUnknownRoutePeer = errors.New("kept route of a peer without a number")
+
+// The statements of the routes and route_peers tables (schema.sql).
+const (
+	insertRoutePeerSQL = `INSERT INTO route_peers (number, address)
+		VALUES (?, ?)`
+	routePeerNumberSQL = `SELECT number FROM route_peers WHERE address = ?`
+	routePeersSQL      = `SELECT number, address FROM route_peers`
+
+	upsertRouteSQL = `INSERT INTO routes (prefix, origin_asn, peers,
+			newest_peer, newest_as_path, newest_time,
+			older_peer, older_as_path, older_time)
+		VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
+		ON CONFLICT (prefix, origin_asn) DO UPDATE SET peers = excluded.peers,
+			newest_peer = excluded.newest_peer,
+			newest_as_path = excluded.newest_as_path,
+			newest_time = excluded.newest_time,
+			older_peer = excluded.older_peer,
+			older_as_path = excluded.older_as_path,
+			older_time = excluded.older_time`
+	deleteRouteSQL = `DELETE FROM routes WHERE prefix = ? AND origin_asn = ?`
+	routesSQL      = `SELECT prefix, origin_asn, peers,
+		newest_peer, newest_as_path, newest_time,
+		older_peer, older_as_path, older_time FROM routes`
+)
+
+// WriteRouteChanges writes, in one transaction, what one call of the routing
+// table's TakeChanges returned. It stores each new peer with its number, writes
+// the row of each changed prefix and origin AS pair, and deletes the row of
+// each removed one. Route times are stored to the second, as the feed gives
+// them.
+//
+// When it fails, it writes nothing. The caller then writes these changes and
+// peers again, because later changes can name the peers it did not store.
+func (d *Database) WriteRouteChanges(
+	changes []routetable.Change, newPeers map[string]uint,
+) error {
+	if len(changes) == 0 && len(newPeers) == 0 {
+		return nil
+	}
+
+	d.lock("WriteRouteChanges")
+	defer d.unlock()
+
+	tx, err := d.beginTx()
+	if err != nil {
+		return fmt.Errorf("failed to begin transaction: %w", err)
+	}
+	defer d.rollback(tx)
+
+	ctx := context.Background()
+
+	// The new peers first, so that the rows can look up their numbers.
+	for address, number := range newPeers {
+		_, err := tx.ExecContext(ctx, insertRoutePeerSQL, number, address)
+		if err != nil {
+			return fmt.Errorf("failed to store peer %s: %w", address, err)
+		}
+	}
+
+	err = writeRouteRows(ctx, tx, changes)
+	if err != nil {
+		return err
+	}
+
+	err = tx.Commit()
+	if err != nil {
+		return fmt.Errorf("failed to commit transaction: %w", err)
+	}
+
+	return nil
+}
+
+// writeRouteRows upserts the row of each change in tx, or deletes it for a
+// removed pair.
+func writeRouteRows(
+	ctx context.Context, tx *loggingTx, changes []routetable.Change,
+) error {
+	upsert, err := tx.PrepareContext(ctx, upsertRouteSQL)
+	if err != nil {
+		return fmt.Errorf("failed to prepare route upsert: %w", err)
+	}
+	defer func() { _ = upsert.Close() }()
+
+	del, err := tx.PrepareContext(ctx, deleteRouteSQL)
+	if err != nil {
+		return fmt.Errorf("failed to prepare route delete: %w", err)
+	}
+	defer func() { _ = del.Close() }()
+
+	peerNumber, err := tx.PrepareContext(ctx, routePeerNumberSQL)
+	if err != nil {
+		return fmt.Errorf("failed to prepare peer number lookup: %w", err)
+	}
+	defer func() { _ = peerNumber.Close() }()
+
+	for _, change := range changes {
+		prefix := change.Prefix.String()
+
+		if change.Removed {
+			_, err := del.ExecContext(ctx, prefix, change.OriginASN)
+			if err != nil {
+				return fmt.Errorf("failed to delete route %s: %w", prefix, err)
+			}
+
+			continue
+		}
+
+		row, err := routeRow(ctx, peerNumber, change)
+		if err != nil {
+			return fmt.Errorf("failed to write route %s: %w", prefix, err)
+		}
+
+		_, err = upsert.ExecContext(ctx, row...)
+		if err != nil {
+			return fmt.Errorf("failed to write route %s: %w", prefix, err)
+		}
+	}
+
+	return nil
+}
+
+// routeRow returns the values of the routes row of change, in the order of
+// upsertRouteSQL. peerNumber looks up a peer's number by its address. The
+// table keeps at most keptRoutesPerRow routes of a pair.
+func routeRow(
+	ctx context.Context, peerNumber *sql.Stmt, change routetable.Change,
+) ([]any, error) {
+	peers, err := change.Peers.MarshalBinary()
+	if err != nil {
+		return nil, err
+	}
+
+	row := []any{change.Prefix.String(), change.OriginASN, peers}
+
+	for i := range keptRoutesPerRow {
+		if i >= len(change.Routes) {
+			row = append(row, nil, nil, nil)
+
+			continue
+		}
+
+		r := change.Routes[i]
+
+		var number uint
+
+		err := peerNumber.QueryRowContext(ctx, r.Peer).Scan(&number)
+		if errors.Is(err, sql.ErrNoRows) {
+			return nil, fmt.Errorf("%w: %s", errUnknownRoutePeer, r.Peer)
+		}
+
+		if err != nil {
+			return nil, err
+		}
+
+		path, err := json.Marshal(r.ASPath)
+		if err != nil {
+			return nil, err
+		}
+
+		row = append(row, number, string(path), r.Time.Unix())
+	}
+
+	return row, nil
+}
+
+// ReadRouteTable returns the routing table that the routes and route_peers
+// tables hold, as startup reads it. It reads the peers first, so that the
+// table keeps their numbers, then loads each row into the table as it reads
+// it. The two reads are separate, so nothing may write the tables meanwhile.
+func (d *Database) ReadRouteTable(ctx context.Context) (*routetable.Table, error) {
+	peerNumbers, err := d.readRoutePeers(ctx)
+	if err != nil {
+		return nil, err
+	}
+
+	addresses := make(map[uint]string, len(peerNumbers))
+	for address, number := range peerNumbers {
+		addresses[number] = address
+	}
+
+	table := routetable.New(routetable.Params{PeerNumbers: peerNumbers})
+
+	rows, err := d.db.QueryContext(ctx, routesSQL)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read routes: %w", err)
+	}
+	defer func() { _ = rows.Close() }()
+
+	for rows.Next() {
+		change, err := scanRouteRow(rows, addresses)
+		if err != nil {
+			return nil, err
+		}
+
+		table.Load(change)
+	}
+
+	err = rows.Err()
+	if err != nil {
+		return nil, fmt.Errorf("failed to read routes: %w", err)
+	}
+
+	return table, nil
+}
+
+// readRoutePeers returns the number of each peer in route_peers, by address.
+func (d *Database) readRoutePeers(ctx context.Context) (map[string]uint, error) {
+	rows, err := d.db.QueryContext(ctx, routePeersSQL)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read route peers: %w", err)
+	}
+	defer func() { _ = rows.Close() }()
+
+	peerNumbers := make(map[string]uint)
+
+	for rows.Next() {
+		var (
+			number  uint
+			address string
+		)
+
+		err := rows.Scan(&number, &address)
+		if err != nil {
+			return nil, fmt.Errorf("failed to read a route peer: %w", err)
+		}
+
+		peerNumbers[address] = number
+	}
+
+	err = rows.Err()
+	if err != nil {
+		return nil, fmt.Errorf("failed to read route peers: %w", err)
+	}
+
+	return peerNumbers, nil
+}
+
+// keptRouteColumns are the three columns of one kept route in a routes row.
+type keptRouteColumns struct {
+	peer   sql.Null[uint]
+	asPath sql.NullString
+	time   sql.NullInt64
+}
+
+// scanRouteRow returns the content of the routes row rows is at, with each kept
+// route's peer named by its address in addresses.
+func scanRouteRow(
+	rows *sql.Rows, addresses map[uint]string,
+) (routetable.Change, error) {
+	var (
+		change routetable.Change
+		prefix string
+		peers  []byte
+		kept   [keptRoutesPerRow]keptRouteColumns
+	)
+
+	err := rows.Scan(&prefix, &change.OriginASN, &peers,
+		&kept[0].peer, &kept[0].asPath, &kept[0].time,
+		&kept[1].peer, &kept[1].asPath, &kept[1].time)
+	if err != nil {
+		return change, fmt.Errorf("failed to read a route: %w", err)
+	}
+
+	change.Prefix, err = netip.ParsePrefix(prefix)
+	if err != nil {
+		return change, fmt.Errorf("failed to read route %s: %w", prefix, err)
+	}
+
+	change.Peers = new(bitset.BitSet)
+
+	err = change.Peers.UnmarshalBinary(peers)
+	if err != nil {
+		return change, fmt.Errorf("failed to read the peers of route %s: %w",
+			prefix, err)
+	}
+
+	for _, columns := range kept {
+		if !columns.peer.Valid {
+			continue
+		}
+
+		address, known := addresses[columns.peer.V]
+		if !known {
+			return change, fmt.Errorf("%w: number %d in route %s",
+				errUnknownRoutePeer, columns.peer.V, prefix)
+		}
+
+		r := routetable.Route{
+			Peer: address,
+			Time: time.Unix(columns.time.Int64, 0).UTC(),
+		}
+
+		err := json.Unmarshal([]byte(columns.asPath.String), &r.ASPath)
+		if err != nil {
+			return change, fmt.Errorf("failed to read an AS path of route %s: %w",
+				prefix, err)
+		}
+
+		change.Routes = append(change.Routes, r)
+	}
+
+	return change, nil
+}

internal/database/routes_internal_test.go

@@ -0,0 +1,464 @@
+package database
+
+import (
+	"errors"
+	"fmt"
+	"math/rand/v2"
+	"net/netip"
+	"os"
+	"reflect"
+	"slices"
+	"testing"
+	"time"
+
+	"github.com/bits-and-blooms/bitset"
+
+	"git.eeqj.de/sneak/routewatch/internal/config"
+	"git.eeqj.de/sneak/routewatch/internal/logger"
+	"git.eeqj.de/sneak/routewatch/internal/routetable"
+)
+
+// Peers, origin ASes and prefixes of the routes table tests.
+const (
+	peerA = "192.0.2.1"
+	peerB = "192.0.2.2"
+	peerC = "2001:db8::3"
+	peerD = "2001:db8::4"
+	peerE = "2001:db8::5"
+
+	asA = 64500
+	asB = 64501
+	// transitAS is the first AS of every test path, the peer's own.
+	transitAS = 64496
+
+	tablePrefixV4      = "10.1.2.0/24"
+	tablePrefixShortV4 = "10.9.0.0/16"
+	tablePrefixV6      = "2001:db8:1::/48"
+)
+
+// tableTime returns the time second seconds after the start of every routes
+// table test.
+func tableTime(second int) time.Time {
+	return time.Date(2026, 1, 2, 3, 4, 5, 0, time.UTC).
+		Add(time.Duration(second) * time.Second)
+}
+
+// announceTo has peer announce prefix with origin AS asn second seconds after
+// the start.
+func announceTo(
+	table *routetable.Table, prefix string, asn int, peer string, second int,
+) {
+	table.Announce(netip.MustParsePrefix(prefix), asn, routetable.Route{
+		Peer: peer, ASPath: []int{transitAS, asn}, Time: tableTime(second),
+	})
+}
+
+// newRoutesTestDatabase returns a new database in a temporary directory.
+func newRoutesTestDatabase(t *testing.T) *Database {
+	t.Helper()
+
+	db, err := New(&config.Config{StateDir: t.TempDir()}, logger.New())
+	if err != nil {
+		t.Fatalf("failed to create database: %v", err)
+	}
+
+	t.Cleanup(func() { _ = db.Close() })
+
+	return db
+}
+
+// writeTableChanges writes what table changed since the last call.
+func writeTableChanges(t *testing.T, db *Database, table *routetable.Table) {
+	t.Helper()
+
+	changes, newPeers := table.TakeChanges()
+
+	err := db.WriteRouteChanges(changes, newPeers)
+	if err != nil {
+		t.Fatalf("WriteRouteChanges: %v", err)
+	}
+}
+
+func TestReadRouteTableGivesBackWhatWasWritten(t *testing.T) {
+	t.Parallel()
+
+	db := newRoutesTestDatabase(t)
+	source := routetable.New(routetable.Params{})
+
+	// Two origins of one prefix: the routes of peerC and peerB are kept for
+	// asA, peerA keeps only its bit.
+	announceTo(source, tablePrefixV4, asA, peerA, 10)
+	announceTo(source, tablePrefixV4, asA, peerB, 20)
+	announceTo(source, tablePrefixV4, asA, peerC, 30)
+	announceTo(source, tablePrefixV4, asB, peerD, 40)
+	// Two routes with the same time.
+	announceTo(source, tablePrefixV6, asA, peerA, 50)
+	announceTo(source, tablePrefixV6, asA, peerC, 50)
+	// An origin whose kept routes were withdrawn while peerA carries it.
+	announceTo(source, tablePrefixShortV4, asB, peerA, 10)
+	announceTo(source, tablePrefixShortV4, asB, peerB, 20)
+	announceTo(source, tablePrefixShortV4, asB, peerC, 30)
+	source.Withdraw(netip.MustParsePrefix(tablePrefixShortV4), peerB)
+	source.Withdraw(netip.MustParsePrefix(tablePrefixShortV4), peerC)
+
+	writeTableChanges(t, db, source)
+	checkReadTable(t, db, source)
+
+	// A peer moves to another origin, one goes down and a new one comes.
+	announceTo(source, tablePrefixV4, asB, peerA, 60)
+	source.PeerDown(peerC)
+	announceTo(source, tablePrefixV6, asB, peerE, 70)
+
+	writeTableChanges(t, db, source)
+	checkReadTable(t, db, source)
+}
+
+// checkReadTable checks that the table ReadRouteTable reads answers every read
+// as want does, for the prefixes and the origin ASes the tests use, and that
+// reading it recorded no change.
+func checkReadTable(t *testing.T, db *Database, want *routetable.Table) {
+	t.Helper()
+
+	got, err := db.ReadRouteTable(t.Context())
+	if err != nil {
+		t.Fatalf("ReadRouteTable: %v", err)
+	}
+
+	if !reflect.DeepEqual(got.Counts(), want.Counts()) {
+		t.Errorf("Counts() = %+v, want %+v", got.Counts(), want.Counts())
+	}
+
+	for _, text := range []string{
+		tablePrefixV4, tablePrefixShortV4, tablePrefixV6,
+	} {
+		prefix := netip.MustParsePrefix(text)
+
+		if !reflect.DeepEqual(got.Origins(prefix), want.Origins(prefix)) {
+			t.Errorf("Origins(%s) = %+v, want %+v",
+				prefix, got.Origins(prefix), want.Origins(prefix))
+		}
+
+		gotSummary, gotFound := got.Lookup(prefix.Addr())
+		wantSummary, wantFound := want.Lookup(prefix.Addr())
+
+		if gotFound != wantFound || !reflect.DeepEqual(gotSummary, wantSummary) {
+			t.Errorf("Lookup(%s) = %+v, %v; want %+v, %v", prefix.Addr(),
+				gotSummary, gotFound, wantSummary, wantFound)
+		}
+
+		ipVersion := ipVersionV6
+		if prefix.Addr().Is4() {
+			ipVersion = ipVersionV4
+		}
+
+		gotSample := sortedSample(got, ipVersion, prefix.Bits())
+		wantSample := sortedSample(want, ipVersion, prefix.Bits())
+
+		if !reflect.DeepEqual(gotSample, wantSample) {
+			t.Errorf("SamplePrefixes(%d, %d) = %+v, want %+v",
+				ipVersion, prefix.Bits(), gotSample, wantSample)
+		}
+	}
+
+	for _, asn := range []int{asA, asB} {
+		if !reflect.DeepEqual(got.OriginatedPrefixes(asn),
+			want.OriginatedPrefixes(asn)) {
+			t.Errorf("OriginatedPrefixes(%d) = %+v, want %+v", asn,
+				got.OriginatedPrefixes(asn), want.OriginatedPrefixes(asn))
+		}
+	}
+
+	changes, newPeers := got.TakeChanges()
+	if len(changes) != 0 || len(newPeers) != 0 {
+		t.Errorf("the read recorded changes %+v and new peers %v, want none",
+			changes, newPeers)
+	}
+}
+
+// sortedSample returns every prefix of the IP version and mask length in
+// table, in prefix order.
+func sortedSample(
+	table *routetable.Table, ipVersion, maskLength int,
+) []routetable.PrefixSummary {
+	// More than the tests have of any mask length.
+	const all = 100
+
+	sample := table.SamplePrefixes(ipVersion, maskLength, all)
+	slices.SortFunc(sample, func(a, b routetable.PrefixSummary) int {
+		return a.Prefix.Compare(b.Prefix)
+	})
+
+	return sample
+}
+
+func TestWriteRouteChangesDeletesTheRowOfARemovedPair(t *testing.T) {
+	t.Parallel()
+
+	db := newRoutesTestDatabase(t)
+	table := routetable.New(routetable.Params{})
+
+	announceTo(table, tablePrefixV4, asA, peerA, 10)
+	announceTo(table, tablePrefixV4, asB, peerB, 20)
+	writeTableChanges(t, db, table)
+
+	// The prefix's only peer with asA withdraws it.
+	table.Withdraw(netip.MustParsePrefix(tablePrefixV4), peerA)
+	writeTableChanges(t, db, table)
+
+	var origins []int
+
+	rows, err := db.db.QueryContext(t.Context(),
+		"SELECT origin_asn FROM routes WHERE prefix = ?", tablePrefixV4)
+	if err != nil {
+		t.Fatalf("reading the routes: %v", err)
+	}
+	defer func() { _ = rows.Close() }()
+
+	for rows.Next() {
+		var origin int
+
+		err := rows.Scan(&origin)
+		if err != nil {
+			t.Fatalf("reading a route: %v", err)
+		}
+
+		origins = append(origins, origin)
+	}
+
+	err = rows.Err()
+	if err != nil {
+		t.Fatalf("reading the routes: %v", err)
+	}
+
+	if !slices.Equal(origins, []int{asB}) {
+		t.Errorf("rows of %s have origins %v, want only %d",
+			tablePrefixV4, origins, asB)
+	}
+}
+
+// routePeer is one row of route_peers.
+type routePeer struct {
+	number  uint
+	address string
+}
+
+// readRoutePeerRows returns the rows of route_peers in number order.
+func readRoutePeerRows(t *testing.T, db *Database) []routePeer {
+	t.Helper()
+
+	rows, err := db.db.QueryContext(t.Context(),
+		"SELECT number, address FROM route_peers ORDER BY number")
+	if err != nil {
+		t.Fatalf("reading the route peers: %v", err)
+	}
+	defer func() { _ = rows.Close() }()
+
+	var peers []routePeer
+
+	for rows.Next() {
+		var peer routePeer
+
+		err := rows.Scan(&peer.number, &peer.address)
+		if err != nil {
+			t.Fatalf("reading a route peer: %v", err)
+		}
+
+		peers = append(peers, peer)
+	}
+
+	err = rows.Err()
+	if err != nil {
+		t.Fatalf("reading the route peers: %v", err)
+	}
+
+	return peers
+}
+
+func TestWriteRouteChangesStoresNewPeers(t *testing.T) {
+	t.Parallel()
+
+	db := newRoutesTestDatabase(t)
+	table := routetable.New(routetable.Params{})
+
+	announceTo(table, tablePrefixV4, asA, peerA, 10)
+	announceTo(table, tablePrefixV6, asA, peerC, 20)
+	writeTableChanges(t, db, table)
+
+	// peerA is stored already; peerB is new.
+	announceTo(table, tablePrefixV4, asA, peerA, 30)
+	announceTo(table, tablePrefixV4, asA, peerB, 40)
+	writeTableChanges(t, db, table)
+
+	want := []routePeer{{0, peerA}, {1, peerC}, {2, peerB}}
+
+	got := readRoutePeerRows(t, db)
+	if !slices.Equal(got, want) {
+		t.Errorf("route_peers = %+v, want %+v", got, want)
+	}
+}
+
+func TestWriteRouteChangesWritesNothingWhenItFails(t *testing.T) {
+	t.Parallel()
+
+	db := newRoutesTestDatabase(t)
+	table := routetable.New(routetable.Params{})
+
+	announceTo(table, tablePrefixV4, asA, peerA, 10)
+	announceTo(table, tablePrefixV6, asA, peerB, 20)
+
+	// peerB's number is missing, so its kept route cannot be written.
+	changes, newPeers := table.TakeChanges()
+	delete(newPeers, peerB)
+
+	err := db.WriteRouteChanges(changes, newPeers)
+	if !errors.Is(err, errUnknownRoutePeer) {
+		t.Fatalf("WriteRouteChanges = %v, want %v", err, errUnknownRoutePeer)
+	}
+
+	peers := readRoutePeerRows(t, db)
+	if len(peers) != 0 {
+		t.Errorf("route_peers = %+v after the failed write, want none", peers)
+	}
+
+	var routes int
+
+	err = db.db.QueryRowContext(t.Context(),
+		"SELECT COUNT(*) FROM routes").Scan(&routes)
+	if err != nil {
+		t.Fatalf("counting the routes: %v", err)
+	}
+
+	if routes != 0 {
+		t.Errorf("routes holds %d rows after the failed write, want none", routes)
+	}
+}
+
+// Parameters for BenchmarkReadRouteTable.
+const (
+	// benchTablePairs is how many prefix and origin AS pairs the table holds:
+	// one origin AS for each prefix. Every benchTableV6Every-th prefix is an
+	// IPv6 /48, the others IPv4 /24s.
+	benchTablePairs   = 200000
+	benchTableV6Every = 5
+	// benchTablePeers is how many peers there are, about as many as RIS has.
+	// benchTablePeersPerPair is how many of them, picked at random, carry each
+	// pair: RIS's 460 million routes over 1.4 million pairs.
+	benchTablePeers        = 1400
+	benchTablePeersPerPair = 330
+	// benchTableBatch is how many changes each write holds.
+	benchTableBatch = 10000
+)
+
+// BenchmarkReadRouteTable times the read of a routing table of 200,000 prefix
+// and origin AS pairs, each carried by 330 of 1,400 peers and keeping two
+// routes. It writes that table first. It reports the bytes of the database
+// file per pair. It is run by hand, not by make test, for example in the test
+// image:
+//
+//	go test -run '^$' -bench ReadRouteTable -benchtime 5x ./internal/database/
+func BenchmarkReadRouteTable(b *testing.B) {
+	db, err := New(&config.Config{StateDir: b.TempDir()}, logger.New())
+	if err != nil {
+		b.Fatalf("failed to create database: %v", err)
+	}
+	defer func() { _ = db.Close() }()
+
+	//nolint:gosec // The benchmark needs the same table every run.
+	rng := rand.New(rand.NewPCG(1, 2))
+
+	newPeers := make(map[string]uint, benchTablePeers)
+	for number := range uint(benchTablePeers) {
+		newPeers[benchPeerAddress(number)] = number
+	}
+
+	for first := 0; first < benchTablePairs; first += benchTableBatch {
+		changes := make([]routetable.Change, 0, benchTableBatch)
+		for pair := first; pair < first+benchTableBatch; pair++ {
+			changes = append(changes, benchTableChange(rng, pair))
+		}
+
+		err := db.WriteRouteChanges(changes, newPeers)
+		if err != nil {
+			b.Fatalf("WriteRouteChanges: %v", err)
+		}
+
+		newPeers = nil
+	}
+
+	// The WAL is copied into the database file, whose size is reported.
+	err = db.Checkpoint(b.Context())
+	if err != nil {
+		b.Fatalf("Checkpoint: %v", err)
+	}
+
+	file, err := os.Stat(db.path)
+	if err != nil {
+		b.Fatalf("reading the database file's size: %v", err)
+	}
+
+	for b.Loop() {
+		table, err := db.ReadRouteTable(b.Context())
+		if err != nil {
+			b.Fatalf("ReadRouteTable: %v", err)
+		}
+
+		counts := table.Counts()
+		if counts.IPv4Prefixes+counts.IPv6Prefixes != benchTablePairs {
+			b.Fatalf("read %d prefixes, want %d",
+				counts.IPv4Prefixes+counts.IPv6Prefixes, benchTablePairs)
+		}
+	}
+
+	// After the loop: b.Loop drops the metrics reported before it starts.
+	b.ReportMetric(float64(file.Size())/benchTablePairs, "file-bytes/pair")
+}
+
+// benchPeerAddress returns the address of the peer numbered number.
+func benchPeerAddress(number uint) string {
+	const addressesPerBlock = 256
+
+	return fmt.Sprintf("198.18.%d.%d",
+		number/addressesPerBlock, number%addressesPerBlock)
+}
+
+// benchTableChange returns the content of the pair numbered pair: its prefix,
+// its origin AS, benchTablePeersPerPair peers picked at random, and the routes
+// of two of them, each with an AS path of four ASes.
+func benchTableChange(rng *rand.Rand, pair int) routetable.Change {
+	const (
+		firstOrigin  = 64500
+		transitASes  = 50000
+		olderSeconds = 10
+	)
+
+	text := fmt.Sprintf("%d.%d.%d.0/24", 1+pair>>16, (pair>>8)%256, pair%256)
+	if pair%benchTableV6Every == 0 {
+		text = fmt.Sprintf("2001:%x:%x::/48", pair>>16, pair%65536)
+	}
+
+	prefix := netip.MustParsePrefix(text)
+	origin := firstOrigin + pair%transitASes
+	peers := bitset.New(benchTablePeers)
+
+	for peers.Count() < benchTablePeersPerPair {
+		peers.Set(rng.UintN(benchTablePeers))
+	}
+
+	change := routetable.Change{Prefix: prefix, OriginASN: origin, Peers: peers}
+
+	// The routes of the first two peers, newest first.
+	for i, number := range slices.Collect(peers.EachSet())[:keptRoutesPerRow] {
+		change.Routes = append(change.Routes, routetable.Route{
+			Peer: benchPeerAddress(number),
+			ASPath: []int{
+				firstOrigin + rng.IntN(transitASes),
+				firstOrigin + rng.IntN(transitASes),
+				firstOrigin + rng.IntN(transitASes),
+				origin,
+			},
+			Time: tableTime(pair - i*olderSeconds),
+		})
+	}
+
+	return change
+}

internal/database/schema.sql

@@ -146,4 +146,33 @@ CREATE INDEX IF NOT EXISTS idx_live_routes_v6_origin_prefix ON live_routes_v6(or
 -- last_updated: the oldest and newest route times in /api/v1/stats and
 -- /status.json.
 CREATE INDEX IF NOT EXISTS idx_live_routes_v4_last_updated ON live_routes_v4(last_updated);
-CREATE INDEX IF NOT EXISTS idx_live_routes_v6_last_updated ON live_routes_v6(last_updated);
+CREATE INDEX IF NOT EXISTS idx_live_routes_v6_last_updated ON live_routes_v6(last_updated);
+
+-- The whole routing table, as internal/routetable keeps it in memory: one row
+-- for each prefix and origin AS that some peer carries. Startup reads every
+-- row, and the rows of the pairs that change are written back.
+CREATE TABLE IF NOT EXISTS routes (
+	prefix TEXT NOT NULL, -- in the text form Go's net/netip prints
+	origin_asn INTEGER NOT NULL,
+	-- The peers whose current route to the prefix has this origin AS: a set
+	-- of their numbers in route_peers, as the bitset package's MarshalBinary
+	-- writes it.
+	peers BLOB NOT NULL,
+	-- The newest of those routes and the one before it, the two kept in full.
+	-- Each has its peer's number in route_peers, its AS path as a JSON array,
+	-- and its time in seconds since 1970, UTC. A route that is not kept has
+	-- NULL in its three columns.
+	newest_peer INTEGER,
+	newest_as_path TEXT,
+	newest_time INTEGER,
+	older_peer INTEGER,
+	older_as_path TEXT,
+	older_time INTEGER,
+	PRIMARY KEY (prefix, origin_asn)
+) WITHOUT ROWID;
+
+-- The peers of the routes table, each with the number it has there.
+CREATE TABLE IF NOT EXISTS route_peers (
+	number INTEGER PRIMARY KEY,
+	address TEXT NOT NULL UNIQUE
+);

internal/routetable/routetable.go

@@ -218,6 +218,57 @@ func (t *Table) PeerDown(peer string) {
 	}
 }
 
+// Load puts one prefix and origin AS pair that the database holds into the
+// table, as startup does with each row it reads. The change's peers join the
+// pair's peer set, and its routes are kept as an announcement's would be.
+// Unlike an update, Load records no change, so TakeChanges does not hand the
+// pair back to be written again. Loading a pair twice changes nothing.
+//
+// The peer set of change holds the peers' numbers from Params.PeerNumbers. A
+// kept route whose peer has no number there is left out, because numbering
+// the peer would record a new peer.
+func (t *Table) Load(change Change) {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+
+	// A removed pair has no peers, and a pair without peers is not in the
+	// table.
+	if change.Peers == nil || change.Peers.None() || !change.Prefix.IsValid() {
+		return
+	}
+
+	prefix := change.Prefix.Masked()
+	o := t.findOrAddOrigin(prefix, change.OriginASN)
+	wasWithoutKeptRoute := o.withoutKeptRoute()
+
+	for peer := range change.Peers.EachSet() {
+		if !o.peers.Test(peer) {
+			o.peers.Set(peer)
+
+			t.familyOf(prefix.Addr()).routes++
+			t.peerPrefixes[peer]++
+		}
+	}
+
+	// Oldest first, as they were announced, so that two routes with the same
+	// time are kept in the order they had.
+	for _, r := range slices.Backward(change.Routes) {
+		peer, known := t.peerNumbers[r.Peer]
+		if !known {
+			continue
+		}
+
+		t.dropKeptRoute(o, peer)
+		t.keepIfNewest(o, route{
+			peer: peer,
+			path: slices.Clone(r.ASPath),
+			time: r.Time,
+		})
+	}
+
+	t.countOriginWithoutKeptRoute(o, wasWithoutKeptRoute)
+}
+
 // TakeChanges returns the prefix and origin AS pairs that changed since it last
 // ran, in prefix order, each with its current content. It also returns the
 // peers numbered since then, by address. It then clears both records.
@@ -413,6 +464,17 @@ func (t *Table) recordChange(
 ) {
 	t.changed[pair{prefix: prefix, asn: o.asn}] = struct{}{}
 
+	t.countOriginWithoutKeptRoute(o, wasWithoutKeptRoute)
+
+	if o.peers.None() {
+		t.removeOrigin(prefix, o)
+	}
+}
+
+// countOriginWithoutKeptRoute brings the count of origins without a kept route
+// up to date after a change to o. wasWithoutKeptRoute says whether o was such
+// an origin before the change.
+func (t *Table) countOriginWithoutKeptRoute(o *origin, wasWithoutKeptRoute bool) {
 	if wasWithoutKeptRoute {
 		t.originsWithoutKeptRoute--
 	}
@@ -420,10 +482,6 @@ func (t *Table) recordChange(
 	if o.withoutKeptRoute() {
 		t.originsWithoutKeptRoute++
 	}
-
-	if o.peers.None() {
-		t.removeOrigin(prefix, o)
-	}
 }
 
 // countKeptTime adds delta to the number of kept routes with time at.

internal/routetable/routetable_test.go

@@ -17,6 +17,7 @@ const (
 	peerA = "192.0.2.1"
 	peerB = "192.0.2.2"
 	peerC = "2001:db8::3"
+	peerD = "2001:db8::4"
 
 	asA = 64500
 	asB = 64501
@@ -475,6 +476,77 @@ func TestTakeChangesHoldsOnlyPairsThatChanged(t *testing.T) {
 	checkChanges(t, table, []change{}, map[string]uint{})
 }
 
+func TestLoad(t *testing.T) {
+	t.Parallel()
+
+	const otherV4 = "10.9.0.0/16"
+
+	v4 := netip.MustParsePrefix(prefixV4)
+	v6 := netip.MustParsePrefix(prefixV6)
+	other := netip.MustParsePrefix(otherV4)
+
+	// The routes of peerC and peerB to v4 are kept; peerA keeps only its bit.
+	// The two routes to v6 have the same time. The routes of other were
+	// withdrawn while peerA still carries it.
+	source := routetable.New(routetable.Params{})
+	announceAll(source, prefixV4, asA, 10, peerA, peerB, peerC)
+	announceAll(source, prefixV4, asB, 40, peerD)
+	source.Announce(v6, asA, routeOf(peerA, asA, 50))
+	source.Announce(v6, asA, routeOf(peerC, asA, 50))
+	announceAll(source, otherV4, asB, 10, peerA, peerB, peerC)
+	source.Withdraw(other, peerB)
+	source.Withdraw(other, peerC)
+
+	changes, peers := source.TakeChanges()
+	loaded := routetable.New(routetable.Params{PeerNumbers: peers})
+
+	// Loading a pair twice changes nothing.
+	for range 2 {
+		for _, change := range changes {
+			loaded.Load(change)
+		}
+	}
+
+	for _, prefix := range []netip.Prefix{v4, v6, other} {
+		got, want := loaded.Origins(prefix), source.Origins(prefix)
+		if !reflect.DeepEqual(got, want) {
+			t.Errorf("Origins(%s) = %+v after the load, want %+v", prefix, got, want)
+		}
+	}
+
+	if !reflect.DeepEqual(loaded.Counts(), source.Counts()) {
+		t.Errorf("Counts() = %+v after the load, want %+v",
+			loaded.Counts(), source.Counts())
+	}
+
+	checkChanges(t, loaded, []change{}, map[string]uint{})
+}
+
+func TestLoadLeavesOutTheRouteOfAPeerWithoutANumber(t *testing.T) {
+	t.Parallel()
+
+	prefix := netip.MustParsePrefix(prefixV4)
+	source := routetable.New(routetable.Params{})
+	announceAll(source, prefixV4, asA, 10, peerA, peerB)
+
+	// peerB has no number in the loaded table.
+	changes, _ := source.TakeChanges()
+	loaded := routetable.New(routetable.Params{
+		PeerNumbers: map[string]uint{peerA: 0},
+	})
+	loaded.Load(changes[0])
+
+	want := []routetable.Route{routeOf(peerA, asA, 10)}
+
+	origins := loaded.Origins(prefix)
+	if len(origins) != 1 || !reflect.DeepEqual(origins[0].Routes, want) {
+		t.Errorf("Origins(%s) = %+v, want one origin with routes %+v",
+			prefix, origins, want)
+	}
+
+	checkChanges(t, loaded, []change{}, map[string]uint{})
+}
+
 // TestConcurrentUse has goroutines update one table while others read it. The
 // race detector, which make test runs, fails it when a method uses the table
 // without its lock.