提交 6a0b3657 编写于 作者: H helinwang 提交者: GitHub

Merge pull request #3062 from helinwang/grace

gracefully shutdown master, pserver, fix gometalinter errors
......@@ -19,6 +19,8 @@ import (
"net"
"net/http"
"net/rpc"
"os"
"os/signal"
"strconv"
"strings"
"time"
......@@ -68,6 +70,20 @@ func main() {
store = &master.InMemStore{}
}
shutdown := func() {
log.Infoln("shutting down gracefully")
err := store.Shutdown()
if err != nil {
log.Errorln(err)
}
}
// Guaranteed to run even panic happens.
defer shutdown()
c := make(chan os.Signal, 1)
signal.Notify(c, os.Interrupt)
s, err := master.NewService(store, *chunkPerTask, *taskTimeoutDur, *taskTimeoutMax)
if err != nil {
log.Fatal(err)
......@@ -84,8 +100,12 @@ func main() {
log.Fatal(err)
}
go func() {
err = http.Serve(l, nil)
if err != nil {
log.Fatal(err)
}
}()
<-c
}
......@@ -18,6 +18,8 @@ import (
"net"
"net/http"
"net/rpc"
"os"
"os/signal"
"strconv"
"time"
......@@ -33,7 +35,8 @@ func main() {
index := flag.Int("index", -1, "index of this pserver, should be larger or equal than 0")
etcdEndpoint := flag.String("etcd-endpoint", "http://127.0.0.1:2379",
"comma separated endpoint string for pserver to connect to etcd")
etcdTimeout := flag.Duration("etcd-timeout", 5*time.Second, "timeout for etcd calls")
dialTimeout := flag.Duration("dial-timeout", 5*time.Second, "dial timeout")
etcdTTL := flag.Int("etcd-ttl", 5, "etcd time to live in seconds")
numPservers := flag.Int("num-pservers", 1, "total pserver count in a training job")
checkpointPath := flag.String("checkpoint-path", "/checkpoints/", "save checkpoint path")
checkpointInterval := flag.Duration("checkpoint-interval", 600*time.Second, "save checkpoint per interval seconds")
......@@ -53,7 +56,7 @@ func main() {
if *index >= 0 {
idx = *index
} else {
e = pserver.NewEtcdClient(*etcdEndpoint, *numPservers, *etcdTimeout)
e = pserver.NewEtcdClient(*etcdEndpoint, *numPservers, *dialTimeout, *etcdTTL)
idx, err = e.Register(*port)
candy.Must(err)
......@@ -67,6 +70,20 @@ func main() {
}
}
shutdown := func() {
log.Infoln("shutting down gracefully")
sErr := e.Shutdown()
if sErr != nil {
log.Errorln(sErr)
}
}
// Guaranteed to run even panic happens.
defer shutdown()
c := make(chan os.Signal, 1)
signal.Notify(c, os.Interrupt)
s, err := pserver.NewService(idx, *checkpointInterval, *checkpointPath, e, cp)
candy.Must(err)
......@@ -77,7 +94,11 @@ func main() {
l, err := net.Listen("tcp", ":"+strconv.Itoa(*port))
candy.Must(err)
go func() {
log.Infof("start pserver at port %d", *port)
err = http.Serve(l, nil)
candy.Must(err)
}()
<-c
}
......@@ -39,15 +39,12 @@ type EtcdClient struct {
statePath string
client *clientv3.Client
lock *concurrency.Mutex
sess *concurrency.Session
}
// NewEtcdClient creates a new EtcdClient.
func NewEtcdClient(endpoints []string, addr string, lockPath, addrPath, statePath string, ttlSec int) (*EtcdClient, error) {
log.Debugf("Connecting to etcd at %v", endpoints)
// TODO(helin): gracefully shutdown etcd store. Because etcd
// store holds a etcd lock, even though the lock will expire
// when the lease timeout, we need to implement graceful
// shutdown to release the lock.
cli, err := clientv3.New(clientv3.Config{
Endpoints: endpoints,
DialTimeout: dialTimeout,
......@@ -67,12 +64,12 @@ func NewEtcdClient(endpoints []string, addr string, lockPath, addrPath, statePat
// one master running, but split-brain problem may cause
// multiple master servers running), and the cluster management
// software will kill one of them.
log.Debugf("Trying to acquire lock at %s.", lockPath)
log.Infof("Trying to acquire lock at %s.", lockPath)
err = lock.Lock(context.TODO())
if err != nil {
return nil, err
}
log.Debugf("Successfully acquired lock at %s.", lockPath)
log.Infof("Successfully acquired lock at %s.", lockPath)
put := clientv3.OpPut(addrPath, addr)
resp, err := cli.Txn(context.Background()).If(lock.IsOwner()).Then(put).Commit()
......@@ -89,6 +86,7 @@ func NewEtcdClient(endpoints []string, addr string, lockPath, addrPath, statePat
statePath: statePath,
client: cli,
lock: lock,
sess: sess,
}
return e, nil
......@@ -157,6 +155,21 @@ func (e *EtcdClient) Load() ([]byte, error) {
return state, nil
}
// Shutdown shuts down the etcd client gracefully.
func (e *EtcdClient) Shutdown() error {
err := e.sess.Close()
newErr := e.client.Close()
if newErr != nil {
if err == nil {
err = newErr
} else {
log.Errorln(newErr)
}
}
return err
}
// GetKey gets the value by the specify key.
func GetKey(c *clientv3.Client, key string, timeout time.Duration) (string, error) {
ctx, cancel := context.WithTimeout(context.Background(), timeout)
......
......@@ -40,3 +40,8 @@ func (m *InMemStore) Load() ([]byte, error) {
return m.buf, nil
}
// Shutdown shuts down the in mem store.
func (m *InMemStore) Shutdown() error {
return nil
}
......@@ -50,6 +50,7 @@ var ErrPassAfter = errors.New("pass number larger than master")
type Store interface {
Save([]byte) error
Load() ([]byte, error)
Shutdown() error
}
// Chunk is a chunk of data consisted of several data instances.
......
......@@ -55,10 +55,10 @@ var curHandle C.paddle_pserver_client
func add(c *client.Client) C.paddle_pserver_client {
mu.Lock()
defer mu.Unlock()
client := curHandle
cli := curHandle
curHandle++
handleMap[client] = c
return client
handleMap[cli] = c
return cli
}
func get(client C.paddle_pserver_client) *client.Client {
......
......@@ -34,16 +34,19 @@ const (
PsPath = "/ps/"
// PsCheckpoint is the etcd path for store checkpoints information
PsCheckpoint = "/checkpoints/"
retryTimeout = 5 * time.Second
)
// EtcdClient is the etcd client that the pserver uses for fault
// tolerance, service registry and coordination.
type EtcdClient struct {
numPservers int
etcdEndpoints string
etcdClient *clientv3.Client
// etcdTimeout is also used as retry intervals.
etcdTimeout time.Duration
endpoints string
client *clientv3.Client
sess *concurrency.Session
dialTimeout time.Duration
ttlSec int
// FIXME: ensure GetExternalIP gets the correct ip for trainers to connect.
externalIP string
// desired number of pservers in the job.
......@@ -52,11 +55,12 @@ type EtcdClient struct {
}
// NewEtcdClient creates an EtcdClient
func NewEtcdClient(endpoints string, numPservers int, timeout time.Duration) *EtcdClient {
func NewEtcdClient(endpoints string, numPservers int, dialtimeout time.Duration, ttlSec int) *EtcdClient {
return &EtcdClient{
etcdTimeout: timeout,
dialTimeout: dialtimeout,
ttlSec: ttlSec,
numPservers: numPservers,
etcdEndpoints: endpoints,
endpoints: endpoints,
}
}
......@@ -64,7 +68,6 @@ func NewEtcdClient(endpoints string, numPservers int, timeout time.Duration) *Et
//
// Register returns the index of the current pserver.
func (e *EtcdClient) Register(port int) (int, error) {
var err error
e.externalIP, err = networkhelper.GetExternalIP()
if err != nil {
......@@ -72,19 +75,26 @@ func (e *EtcdClient) Register(port int) (int, error) {
}
// initialize connection to etcd.
ep := strings.Split(e.etcdEndpoints, ",")
ep := strings.Split(e.endpoints, ",")
for {
cli, err := clientv3.New(clientv3.Config{
Endpoints: ep,
DialTimeout: e.etcdTimeout,
DialTimeout: e.dialTimeout,
})
if err != nil {
log.Errorf("connect to etcd error: %v", err)
time.Sleep(e.etcdTimeout)
time.Sleep(retryTimeout)
continue
}
e.client = cli
sess, err := concurrency.NewSession(cli, concurrency.WithTTL(e.ttlSec))
if err != nil {
log.Errorf("create etcd session error: %v", err)
time.Sleep(retryTimeout)
continue
}
e.etcdClient = cli
log.Debugf("inited client to %s", e.etcdEndpoints)
e.sess = sess
log.Debugf("inited client to %s", e.endpoints)
break
}
// init /ps_desired using transaction, for multiple pservers may want to write
......@@ -95,7 +105,7 @@ func (e *EtcdClient) Register(port int) (int, error) {
cancel()
if err != nil {
log.Warn(err)
time.Sleep(e.etcdTimeout)
time.Sleep(retryTimeout)
continue
}
break
......@@ -106,18 +116,18 @@ func (e *EtcdClient) Register(port int) (int, error) {
// wait and set s.desired init value
for {
ctx, cancel := context.WithTimeout(context.Background(), time.Second)
resp, err := e.etcdClient.Get(ctx, PsDesired)
resp, err := e.client.Get(ctx, PsDesired)
cancel()
if err != nil {
log.Errorf("getting %s error: %v", PsDesired, err)
time.Sleep(e.etcdTimeout)
time.Sleep(retryTimeout)
continue
}
if len(resp.Kvs) != 0 {
e.desired, err = strconv.Atoi(string(resp.Kvs[0].Value))
if err != nil {
log.Errorf("value of %s invalid %v\n", PsDesired, err)
time.Sleep(e.etcdTimeout)
time.Sleep(retryTimeout)
// NOTE: wait util ps_desired value change
continue
}
......@@ -134,7 +144,7 @@ func (e *EtcdClient) Register(port int) (int, error) {
cancel()
if err != nil {
log.Warn(err)
time.Sleep(e.etcdTimeout)
time.Sleep(retryTimeout)
continue
}
break
......@@ -144,10 +154,10 @@ func (e *EtcdClient) Register(port int) (int, error) {
}
func (e *EtcdClient) initDesiredPservers(ctx context.Context, numPservers int) (*clientv3.TxnResponse, error) {
return concurrency.NewSTM(e.etcdClient, func(c concurrency.STM) error {
return concurrency.NewSTM(e.client, func(c concurrency.STM) error {
dsStr := c.Get(PsDesired)
if dsStr == "" {
c.Put(PsDesired, strconv.Itoa(numPservers))
c.Put(PsDesired, strconv.Itoa(numPservers), clientv3.WithLease(e.sess.Lease()))
}
return nil
}, concurrency.WithAbortContext(ctx), concurrency.WithIsolation(concurrency.RepeatableReads))
......@@ -156,7 +166,7 @@ func (e *EtcdClient) initDesiredPservers(ctx context.Context, numPservers int) (
// registerPserverEtcd registers pserver node on etcd using transaction.
func (e *EtcdClient) registerPserverEtcd(ctx context.Context, port int) (int, error) {
var idx int
_, err := concurrency.NewSTM(e.etcdClient, func(c concurrency.STM) error {
_, err := concurrency.NewSTM(e.client, func(c concurrency.STM) error {
registered := false
for i := 0; i < e.desired; i++ {
psKey := PsPath + strconv.Itoa(i)
......@@ -165,26 +175,10 @@ func (e *EtcdClient) registerPserverEtcd(ctx context.Context, port int) (int, er
log.Debugf("got value (%s) for key: %s", ps, psKey)
if ps == "" {
resp, err := e.etcdClient.Grant(context.TODO(), 5)
if err != nil {
log.Fatal(err)
}
// find the first id and write info
pserverAddr := e.externalIP + ":" + strconv.Itoa(port)
c.Put(psKey, pserverAddr, clientv3.WithLease(resp.ID))
c.Put(psKey, pserverAddr, clientv3.WithLease(e.sess.Lease()))
log.Debugf("set pserver node %s with value %s", psKey, pserverAddr)
ch, kaerr := e.etcdClient.KeepAlive(context.TODO(), resp.ID)
if kaerr != nil {
log.Errorf("keepalive etcd node error: %v", kaerr)
return kaerr
}
// Eat the keep alive message so etcd
// will not expire the lease.
go func(ch <-chan *clientv3.LeaseKeepAliveResponse) {
ka := <-ch
log.Debugf("keepalive: %d\n", ka.TTL)
}(ch)
log.Debug("register finished")
idx = i
registered = true
......@@ -207,7 +201,7 @@ func (e *EtcdClient) registerPserverEtcd(ctx context.Context, port int) (int, er
// GetKey gets the value by the specified key
func (e *EtcdClient) GetKey(key string, timeout time.Duration) ([]byte, error) {
ctx, cancel := context.WithTimeout(context.Background(), timeout)
resp, err := e.etcdClient.Get(ctx, key)
resp, err := e.client.Get(ctx, key)
cancel()
if err != nil {
return []byte{}, err
......@@ -223,7 +217,27 @@ func (e *EtcdClient) GetKey(key string, timeout time.Duration) ([]byte, error) {
// PutKey put into etcd with value by key specified
func (e *EtcdClient) PutKey(key string, value []byte, timeout time.Duration) error {
ctx, cancel := context.WithTimeout(context.Background(), timeout)
_, err := e.etcdClient.Put(ctx, key, string(value))
_, err := e.client.Put(ctx, key, string(value), clientv3.WithLease(e.sess.Lease()))
cancel()
return err
}
// Shutdown shuts down the etcd client gracefully.
func (e *EtcdClient) Shutdown() error {
var err error
if e.sess != nil {
err = e.sess.Close()
}
if e.client != nil {
newErr := e.client.Close()
if newErr != nil {
if err != nil {
log.Errorln(newErr)
} else {
err = newErr
}
}
}
return err
}
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册