func testNodeRestartInner(t *testing.T, c cluster.Cluster, cfg cluster.TestConfig) {
num := c.NumNodes()
if num <= 0 {
t.Fatalf("%d nodes in cluster", num)
}
// One client for each node.
initBank(t, c.PGUrl(0))
start := timeutil.Now()
state := testState{
t: t,
errChan: make(chan error, 1),
teardown: make(chan struct{}),
deadline: start.Add(cfg.Duration),
clients: make([]testClient, 1),
}
client := &state.clients[0]
client.Lock()
client.db = makePGClient(t, c.PGUrl(num-1))
client.Unlock()
go transferMoneyLoop(0, &state, *numAccounts, *maxTransfer)
defer func() {
<-state.teardown
}()
// Chaos monkey.
rnd, seed := randutil.NewPseudoRand()
log.Warningf("monkey starts (seed %d)", seed)
pickNodes := func() []int {
return []int{rnd.Intn(num - 1)}
}
go chaosMonkey(&state, c, false, pickNodes)
waitClientsStop(1, &state, cfg.Stall)
// Verify accounts.
verifyAccounts(t, client)
elapsed := time.Since(start)
count := atomic.LoadUint64(&client.count)
log.Infof("%d %.1f/sec", count, float64(count)/elapsed.Seconds())
kvClient, kvStopper := c.NewClient(t, num-1)
defer kvStopper.Stop()
if pErr := kvClient.CheckConsistency(keys.TableDataMin, keys.TableDataMax); pErr != nil {
// TODO(.*): change back to t.Fatal after #5051.
log.Error(pErr)
}
}
func testClusterRecoveryInner(t *testing.T, c cluster.Cluster, cfg cluster.TestConfig) {
num := c.NumNodes()
if num <= 0 {
t.Fatalf("%d nodes in cluster", num)
}
// One client for each node.
initBank(t, c.PGUrl(0))
start := timeutil.Now()
state := testState{
t: t,
errChan: make(chan error, num),
teardown: make(chan struct{}),
deadline: start.Add(cfg.Duration),
clients: make([]testClient, num),
}
for i := 0; i < num; i++ {
state.clients[i].Lock()
state.initClient(t, c, i)
state.clients[i].Unlock()
go transferMoneyLoop(i, &state, *numAccounts, *maxTransfer)
}
defer func() {
<-state.teardown
}()
// Chaos monkey.
rnd, seed := randutil.NewPseudoRand()
log.Warningf("monkey starts (seed %d)", seed)
pickNodes := func() []int {
return rnd.Perm(num)[:rnd.Intn(num)+1]
}
go chaosMonkey(&state, c, true, pickNodes)
waitClientsStop(num, &state, cfg.Stall)
// Verify accounts.
verifyAccounts(t, &state.clients[0])
elapsed := time.Since(start)
var count uint64
counts := state.counts()
for _, c := range counts {
count += c
}
log.Infof("%d %.1f/sec", count, float64(count)/elapsed.Seconds())
}
func testNodeRestartInner(t *testing.T, c cluster.Cluster, cfg cluster.TestConfig) {
num := c.NumNodes()
if minNum := 3; num < minNum {
t.Skipf("need at least %d nodes, got %d", minNum, num)
}
// One client for each node.
initBank(t, c.PGUrl(0))
start := timeutil.Now()
state := testState{
t: t,
errChan: make(chan error, 1),
teardown: make(chan struct{}),
deadline: start.Add(cfg.Duration),
clients: make([]testClient, 1),
}
client := &state.clients[0]
client.Lock()
client.db = makePGClient(t, c.PGUrl(num-1))
client.Unlock()
go transferMoneyLoop(0, &state, *numAccounts, *maxTransfer)
defer func() {
<-state.teardown
}()
// Chaos monkey.
rnd, seed := randutil.NewPseudoRand()
log.Warningf(context.Background(), "monkey starts (seed %d)", seed)
pickNodes := func() []int {
return []int{rnd.Intn(num - 1)}
}
go chaosMonkey(&state, c, false, pickNodes)
waitClientsStop(1, &state, stall)
// Verify accounts.
verifyAccounts(t, client)
elapsed := timeutil.Since(start)
count := atomic.LoadUint64(&client.count)
log.Infof(context.Background(), "%d %.1f/sec", count, float64(count)/elapsed.Seconds())
}
// initClient initializes the client talking to node "i".
// It requires that the caller hold the client's write lock.
func (state *testState) initClient(t *testing.T, c cluster.Cluster, i int) {
state.clients[i].db = makePGClient(t, c.PGUrl(i))
}
func testMonotonicInsertsInner(t *testing.T, c cluster.Cluster, cfg cluster.TestConfig) {
var clients []mtClient
for i := 0; i < c.NumNodes(); i++ {
clients = append(clients, mtClient{ID: i, DB: makePGClient(t, c.PGUrl(i))})
}
// We will insert into this table by selecting MAX(val) and increasing by
// one and expect that val and sts (the commit timestamp) are both
// simultaneously increasing.
if _, err := clients[0].Exec(`
CREATE DATABASE mono;
CREATE TABLE IF NOT EXISTS mono.mono (val INT, sts STRING, node INT, tb INT);
INSERT INTO mono.mono VALUES(-1, '0', -1, -1)`); err != nil {
t.Fatal(err)
}
var idGen uint64
invoke := func(client mtClient) {
logPrefix := fmt.Sprintf("%03d.%03d: ", atomic.AddUint64(&idGen, 1), client.ID)
l := func(msg string, args ...interface{}) {
t.Logf(logPrefix+msg, args...)
}
l("begin")
defer l("done")
var exRow, insRow mtRow
var attempt int
if err := crdb.ExecuteTx(client.DB, func(tx *gosql.Tx) error {
attempt++
l("attempt %d", attempt)
if err := tx.QueryRow(`SELECT cluster_logical_timestamp()`).Scan(
&insRow.sts,
); err != nil {
l(err.Error())
return err
}
l("read max val")
if err := tx.QueryRow(`SELECT MAX(val) AS m FROM mono.mono`).Scan(
&exRow.val,
); err != nil {
l(err.Error())
return err
}
l("read max row for val=%d", exRow.val)
if err := tx.QueryRow(`SELECT sts, node, tb FROM mono.mono WHERE val = $1`,
exRow.val,
).Scan(
&exRow.sts, &exRow.node, &exRow.tb,
); err != nil {
l(err.Error())
return err
}
l("insert")
if err := tx.QueryRow(`
INSERT INTO mono.mono (val, sts, node, tb) VALUES($1, $2, $3, $4)
RETURNING val, sts, node, tb`,
exRow.val+1, insRow.sts, client.ID, 0,
).Scan(
&insRow.val, &insRow.sts, &insRow.node, &insRow.tb,
); err != nil {
l(err.Error())
return err
}
l("commit")
return nil
}); err != nil {
t.Errorf("%T: %v", err, err)
}
}
verify := func() {
client := clients[0]
var numDistinct int
if err := client.QueryRow("SELECT COUNT(DISTINCT(val)) FROM mono.mono").Scan(
&numDistinct,
); err != nil {
t.Fatal(err)
}
rows, err := client.Query("SELECT val, sts, node, tb FROM mono.mono ORDER BY val ASC, sts ASC")
if err != nil {
t.Fatal(err)
}
var results mtRows
for rows.Next() {
var row mtRow
if err := rows.Scan(&row.val, &row.sts, &row.node, &row.tb); err != nil {
t.Fatal(err)
}
results = append(results, row)
}
if !sort.IsSorted(results) {
t.Errorf("results are not sorted:\n%s", results)
}
if numDistinct != len(results) {
t.Errorf("'val' column is not unique: %d results, but %d distinct:\n%s",
len(results), numDistinct, results)
}
}
concurrency := 2 * c.NumNodes()
sem := make(chan struct{}, concurrency)
timer := time.After(cfg.Duration)
defer verify()
defer func() {
// Now that consuming has stopped, fill up the semaphore (i.e. wait for
// still-running goroutines to stop)
for i := 0; i < concurrency; i++ {
sem <- struct{}{}
}
}()
for {
select {
case sem <- struct{}{}:
case <-stopper:
return
case <-timer:
return
}
go func(client mtClient) {
invoke(client)
<-sem
}(clients[rand.Intn(c.NumNodes())])
}
}
func testEventLogInner(t *testing.T, c cluster.Cluster, cfg cluster.TestConfig) {
num := c.NumNodes()
if num <= 0 {
t.Fatalf("%d nodes in cluster", num)
}
var confirmedClusterID uuid.UUID
type nodeEventInfo struct {
Descriptor roachpb.NodeDescriptor
ClusterID uuid.UUID
}
// Verify that a node_join message was logged for each node in the cluster.
// We expect there to eventually be one such message for each node in the
// cluster, and each message must be correctly formatted.
util.SucceedsSoon(t, func() error {
db := makePGClient(t, c.PGUrl(0))
defer db.Close()
// Query all node join events. There should be one for each node in the
// cluster.
rows, err := db.Query(
"SELECT targetID, info FROM system.eventlog WHERE eventType = $1",
string(csql.EventLogNodeJoin))
if err != nil {
return err
}
seenIds := make(map[int64]struct{})
var clusterID uuid.UUID
for rows.Next() {
var targetID int64
var infoStr gosql.NullString
if err := rows.Scan(&targetID, &infoStr); err != nil {
t.Fatal(err)
}
// Verify the stored node descriptor.
if !infoStr.Valid {
t.Fatalf("info not recorded for node join, target node %d", targetID)
}
var info nodeEventInfo
if err := json.Unmarshal([]byte(infoStr.String), &info); err != nil {
t.Fatal(err)
}
if a, e := int64(info.Descriptor.NodeID), targetID; a != e {
t.Fatalf("Node join with targetID %d had descriptor for wrong node %d", e, a)
}
// Verify cluster ID is recorded, and is the same for all nodes.
if uuid.Equal(info.ClusterID, *uuid.EmptyUUID) {
t.Fatalf("Node join recorded nil cluster id, info: %v", info)
}
if uuid.Equal(clusterID, *uuid.EmptyUUID) {
clusterID = info.ClusterID
} else if !uuid.Equal(clusterID, info.ClusterID) {
t.Fatalf(
"Node join recorded different cluster ID than earlier node. Expected %s, got %s. Info: %v",
clusterID, info.ClusterID, info)
}
// Verify that all NodeIDs are different.
if _, ok := seenIds[targetID]; ok {
t.Fatalf("Node ID %d seen in two different node join messages", targetID)
}
seenIds[targetID] = struct{}{}
}
if err := rows.Err(); err != nil {
return err
}
if a, e := len(seenIds), c.NumNodes(); a != e {
return errors.Errorf("expected %d node join messages, found %d: %v", e, a, seenIds)
}
confirmedClusterID = clusterID
return nil
})
// Stop and Start Node 0, and verify the node restart message.
if err := c.Kill(0); err != nil {
t.Fatal(err)
}
if err := c.Restart(0); err != nil {
t.Fatal(err)
}
util.SucceedsSoon(t, func() error {
db := makePGClient(t, c.PGUrl(0))
defer db.Close()
// Query all node restart events. There should only be one.
rows, err := db.Query(
"SELECT targetID, info FROM system.eventlog WHERE eventType = $1",
string(csql.EventLogNodeRestart))
if err != nil {
return err
}
seenCount := 0
for rows.Next() {
var targetID int64
var infoStr gosql.NullString
if err := rows.Scan(&targetID, &infoStr); err != nil {
t.Fatal(err)
}
// Verify the stored node descriptor.
if !infoStr.Valid {
t.Fatalf("info not recorded for node join, target node %d", targetID)
}
var info nodeEventInfo
if err := json.Unmarshal([]byte(infoStr.String), &info); err != nil {
t.Fatal(err)
}
if a, e := int64(info.Descriptor.NodeID), targetID; a != e {
t.Fatalf("Node join with targetID %d had descriptor for wrong node %d", e, a)
}
// Verify cluster ID is recorded, and is the same for all nodes.
if !uuid.Equal(confirmedClusterID, info.ClusterID) {
t.Fatalf(
"Node restart recorded different cluster ID than earlier join. Expected %s, got %s. Info: %v",
confirmedClusterID, info.ClusterID, info)
}
seenCount++
}
if err := rows.Err(); err != nil {
return err
}
if seenCount != 1 {
return errors.Errorf("Expected only one node restart event, found %d", seenCount)
}
return nil
})
}