backup: check the store state by last heartbeat (#43099)

close #42973

br/pkg/backup/push.go

@@ -77,16 +77,16 @@ func (push *pushDown) pushBackup(
 		store := s
 		storeID := s.GetId()
 		lctx := logutil.ContextWithField(ctx, zap.Uint64("store-id", storeID))
-		if s.GetState() != metapb.StoreState_Up {
-			logutil.CL(lctx).Warn("skip store", zap.Stringer("State", s.GetState()))
+		if err := utils.CheckStoreLiveness(s); err != nil {
+			logutil.CL(lctx).Warn("skip store", logutil.ShortError(err))
 			continue
 		}
 		client, err := push.mgr.GetBackupClient(lctx, storeID)
 		if err != nil {
 			// BR should be able to backup even some of stores disconnected.
 			// The regions managed by this store can be retried at fine-grained backup then.
 			logutil.CL(lctx).Warn("fail to connect store, skipping", zap.Error(err))
-			return nil
+			continue
 		}
 		wg.Add(1)
 		go func() {

br/pkg/utils/misc.go

@@ -19,13 +19,24 @@ import (
 	"time"
 
 	"github.com/pingcap/errors"
+	"github.com/pingcap/kvproto/pkg/metapb"
+	berrors "github.com/pingcap/tidb/br/pkg/errors"
 	"github.com/pingcap/tidb/parser/mysql"
 	"github.com/pingcap/tidb/parser/types"
 	"google.golang.org/grpc"
 	"google.golang.org/grpc/credentials"
 	"google.golang.org/grpc/credentials/insecure"
 )
 
+const (
+	// storeDisconnectionDuration is the max duration of a store to be treated as living.
+	// when a store doesn't send heartbeat for 100s, it is probably offline, and most of leaders should be transformed.
+	// (How about network partition between TiKV and PD? Even that is rare.)
+	// Also note that the offline threshold in PD is 20s, see
+	// https://github.com/tikv/pd/blob/c40e319f50822678cda71ae62ee2fd70a9cac010/pkg/core/store.go#L523
+	storeDisconnectionDuration = 100 * time.Second
+)
+
 // IsTypeCompatible checks whether type target is compatible with type src
 // they're compatible if
 // - same null/not null and unsigned flag(maybe we can allow src not null flag, target null flag later)
@@ -103,3 +114,20 @@ func GRPCConn(ctx context.Context, storeAddr string, tlsConf *tls.Config, opts .
 	}
 	return connection, nil
 }
+
+// CheckStoreLiveness checks whether a store is still alive.
+// Some versions of PD may not set the store state in the gRPC response.
+// We need to check it manually.
+func CheckStoreLiveness(s *metapb.Store) error {
+	if s.State != metapb.StoreState_Up {
+		return errors.Annotatef(berrors.ErrKVStorage, "the store state isn't up, it is %s", s.State)
+	}
+	// If the field isn't present (the default value), skip this check.
+	if s.GetLastHeartbeat() > 0 {
+		lastHeartBeat := time.Unix(0, s.GetLastHeartbeat())
+		if sinceLastHB := time.Since(lastHeartBeat); sinceLastHB > storeDisconnectionDuration {
+			return errors.Annotatef(berrors.ErrKVStorage, "the store last heartbeat is too far, at %s", sinceLastHB)
+		}
+	}
+	return nil
+}