Merge pull request #1100 from percona/PBM-1335-fallback-db-path-for-restore

PBM-1335: Fallback dbpath for physical restore

Author: boris-ilijic

pbm/restore/physical.go

@@ -7,12 +7,14 @@ import (
 	"fmt"
 	"io"
 	slog "log"
+	"maps"
 	"math/rand"
 	"net"
 	"os"
 	"os/exec"
 	"path"
 	"path/filepath"
+	"slices"
 	"strconv"
 	"strings"
 	"sync"
@@ -46,13 +48,16 @@ import (
 const (
 	defaultRSdbpath   = "/data/db"
 	defaultCSRSdbpath = "/data/configdb"
+	fallbackDir       = ".fallbacksync"
 
 	mongofslock = "mongod.lock"
 
 	defaultPort = 27017
 
 	tryConnCount   = 5
 	tryConnTimeout = 5 * time.Minute
+
+	internalMongodLog = "pbm.restore.log"
 )
 
 type files struct {
@@ -229,18 +234,80 @@ func (r *PhysRestore) close(noerr, cleanup bool) {
 		if err != nil && !errors.Is(err, os.ErrNotExist) {
 			r.log.Warning("remove file <%s>: %v", backup.FilelistName, err)
 		}
-	} else if cleanup { // clean-up dbpath on err if needed
+	}
+
+	r.log.Debug("wait for cluster status")
+	cStatus, err := r.waitClusterStatus()
+	if err != nil {
+		r.log.Warning("waiting for cluster status during cleanup: %v", err)
+	}
+	if cStatus == defs.StatusError {
+		r.log.Warning("apply db data from %s", fallbackDir)
+		err := r.migrateFromFallbackDirToDBDir()
+		if err != nil {
+			r.log.Error("migrate from fallback dir: %v", err)
+		}
+	} else if cleanup { // clean-up dbpath on err if needed (cluster is done or partlyDone)
 		r.log.Debug("clean-up dbpath")
-		err := removeAll(r.dbpath, r.log)
+		err := removeAll(r.dbpath, nil, r.log)
 		if err != nil {
 			r.log.Error("flush dbpath %s: %v", r.dbpath, err)
 		}
+	} else { // free space by just deleting fallback dir in any other case
+		r.log.Debug("remove fallback dir")
+		err := r.removeFallback()
+		if err != nil {
+			r.log.Error("flush fallback: %v", err)
+		}
 	}
+
 	if r.stopHB != nil {
 		close(r.stopHB)
 	}
 }
 
+// waitClusterStatus blocks until cluster status file is set on one of final statuses.
+// It also checks HB to see if cluster is stucked.
+func (r *PhysRestore) waitClusterStatus() (defs.Status, error) {
+	errF := fmt.Sprintf("%s.%s", r.syncPathCluster, defs.StatusError)
+	doneF := fmt.Sprintf("%s.%s", r.syncPathCluster, defs.StatusDone)
+	partlyDoneF := fmt.Sprintf("%s.%s", r.syncPathCluster, defs.StatusPartlyDone)
+	hbF := fmt.Sprintf("%s.%s", r.syncPathCluster, syncHbSuffix)
+
+	tk := time.NewTicker(time.Second * 5)
+	defer tk.Stop()
+
+	for range tk.C {
+		_, err := r.stg.FileStat(errF)
+		if err == nil {
+			return defs.StatusError, nil
+		} else if !errors.Is(err, storage.ErrNotExist) {
+			r.log.Error("error while reading %s file", errF)
+		}
+
+		_, err = r.stg.FileStat(doneF)
+		if err == nil {
+			return defs.StatusDone, nil
+		} else if !errors.Is(err, storage.ErrNotExist) {
+			r.log.Error("error while reading %s file", errF)
+		}
+
+		_, err = r.stg.FileStat(partlyDoneF)
+		if err == nil {
+			return defs.StatusPartlyDone, nil
+		} else if !errors.Is(err, storage.ErrNotExist) {
+			r.log.Error("error while reading %s file", errF)
+		}
+
+		err = r.checkHB(hbF)
+		if err != nil {
+			return defs.StatusError, errors.Wrap(err, "check hb for cluster")
+		}
+
+	}
+	return defs.StatusError, errors.New("wait cluster status")
+}
+
 func (r *PhysRestore) flush(ctx context.Context) error {
 	r.log.Debug("shutdown server")
 	rsStat, err := topo.GetReplsetStatus(ctx, r.node)
@@ -289,15 +356,93 @@ func (r *PhysRestore) flush(ctx context.Context) error {
 		}
 	}
 
-	r.log.Debug("remove old data")
-	err = removeAll(r.dbpath, r.log)
+	err = r.migrateDBDirToFallbackDir()
+	if err != nil {
+		return errors.Wrapf(err, "move files to fallback path")
+	}
+
+	return nil
+}
+
+// migrateDBDirToFallbackDir moves content of dbPath dir into fallback dir.
+// It also removes old fallback dir, and creates new with the same perms.
+func (r *PhysRestore) migrateDBDirToFallbackDir() error {
+	dbpath := filepath.Clean(r.dbpath)
+	fallbackPath := filepath.Join(dbpath, fallbackDir)
+	r.log.Debug("dbpath: %s, fallbackPath: %s", dbpath, fallbackPath)
+
+	err := os.RemoveAll(fallbackPath)
+	if err != nil {
+		return errors.Wrap(err, "remove fallback db path")
+	}
+
+	r.log.Debug("create %s", fallbackPath)
+	info, err := os.Stat(dbpath)
+	if err != nil {
+		return errors.Wrap(err, "stat")
+	}
+	err = os.MkdirAll(fallbackPath, info.Mode())
+	if err != nil {
+		return errors.Wrapf(err, "creating dir %s", fallbackPath)
+	}
+
+	r.log.Info("move data files from %s to %s", r.dbpath, fallbackDir)
+	err = r.moveToFallback()
 	if err != nil {
-		return errors.Wrapf(err, "flush dbpath %s", r.dbpath)
+		return errors.Wrapf(err, "fail to move to %s", fallbackPath)
 	}
 
 	return nil
 }
 
+// migrateFromFallbackDirToDBDir wipe up dbpath dir and
+// moves all content from fallback path.
+func (r *PhysRestore) migrateFromFallbackDirToDBDir() error {
+	r.log.Debug("clean-up dbpath")
+	err := removeAll(r.dbpath, []string{fallbackDir}, r.log)
+	if err != nil {
+		r.log.Error("flush dbpath %s: %v", r.dbpath, err)
+		return errors.Wrap(err, "remove all from dbpath")
+	}
+
+	r.log.Info("move data files from %s to %s", fallbackDir, r.dbpath)
+	err = r.moveFromFallback()
+	if err != nil {
+		r.log.Error("moving from %s: %v", fallbackDir, err)
+		return errors.Wrapf(err, "move from %s", fallbackDir)
+	}
+
+	return nil
+}
+
+// moveFromFallback moves all files/dirs from fallback dir to dbpath dir.
+func (r *PhysRestore) moveFromFallback() error {
+	return moveAll(
+		path.Join(r.dbpath, fallbackDir),
+		r.dbpath,
+		nil,
+		r.log,
+	)
+}
+
+// moveToFallback moves all files/dirs except fallback dir from dbpath to fallback dir
+func (r *PhysRestore) moveToFallback() error {
+	return moveAll(
+		r.dbpath,
+		path.Join(r.dbpath, fallbackDir),
+		[]string{fallbackDir},
+		r.log,
+	)
+}
+
+// removeFallback removes fallback dir
+func (r *PhysRestore) removeFallback() error {
+	dbpath := filepath.Clean(r.dbpath)
+	fallbackPath := filepath.Join(dbpath, fallbackDir)
+	err := os.RemoveAll(fallbackPath)
+	return errors.Wrap(err, "remove fallback db path")
+}
+
 func nodeShutdown(ctx context.Context, m *mongo.Client) error {
 	err := m.Database("admin").RunCommand(ctx, bson.D{{"shutdown", 1}}).Err()
 	if err == nil || strings.Contains(err.Error(), "socket was unexpectedly closed") {
@@ -441,7 +586,7 @@ func (r *PhysRestore) toState(status defs.Status) (_ defs.Status, err error) {
 
 	if r.nodeInfo.IsPrimary || status == defs.StatusDone {
 		r.log.Info("waiting for `%s` status in rs %v", status, r.syncPathPeers)
-		cstat, err := r.waitFiles(status, copyMap(r.syncPathPeers), false)
+		cstat, err := r.waitFiles(status, maps.Clone(r.syncPathPeers), false)
 		if err != nil {
 			return defs.StatusError, errors.Wrap(err, "wait for nodes in rs")
 		}
@@ -454,7 +599,7 @@ func (r *PhysRestore) toState(status defs.Status) (_ defs.Status, err error) {
 
 	if r.nodeInfo.IsClusterLeader() || status == defs.StatusDone {
 		r.log.Info("waiting for shards %v", r.syncPathShards)
-		cstat, err := r.waitFiles(status, copyMap(r.syncPathShards), true)
+		cstat, err := r.waitFiles(status, maps.Clone(r.syncPathShards), true)
 		if err != nil {
 			return defs.StatusError, errors.Wrap(err, "wait for shards")
 		}
@@ -525,13 +670,37 @@ func (n nodeError) Error() string {
 	return fmt.Sprintf("%s failed: %s", n.node, n.msg)
 }
 
-func copyMap[K comparable, V any](m map[K]V) map[K]V {
-	cp := make(map[K]V)
-	for k, v := range m {
-		cp[k] = v
+// checkForRSLevelErr checks if all nodes have an error,
+// and in that case true is returned.
+// If any node doesn't have error, false is returned.
+func (r *PhysRestore) checkForRSLevelErr() bool {
+	for f := range r.syncPathPeers {
+		errFile := f + "." + string(defs.StatusError)
+		_, err := r.stg.FileStat(errFile)
+		if errors.Is(err, storage.ErrNotExist) {
+			return false
+		}
+		if err != nil {
+			r.log.Error("error while checking file %s: %v", errFile, err)
+		}
+		// error file is found
 	}
+	return true
+}
 
-	return cp
+// checkForClusterLevelErr checks if any RS (shard) has an error.
+// It returns true if at least one RS has error, otherwise false.
+func (r *PhysRestore) checkForClusterLevelErr() bool {
+	for f := range r.syncPathShards {
+		errFile := f + "." + string(defs.StatusError)
+		_, err := r.stg.FileStat(errFile)
+		if err == nil {
+			return true
+		} else if !errors.Is(err, storage.ErrNotExist) {
+			r.log.Error("error while checking file %s: %v", errFile, err)
+		}
+	}
+	return false
 }
 
 func (r *PhysRestore) waitFiles(
@@ -756,7 +925,7 @@ func (r *PhysRestore) Snapshot(
 		// set failed status of node on error, but
 		// don't mark node as failed after the local restore succeed
 		if err != nil && !progress.is(restoreDone) && !errors.Is(err, ErrNoDataForShard) {
-			r.MarkFailed(meta, err, !progress.is(restoreStared))
+			r.MarkFailed(meta, err)
 		}
 
 		r.close(err == nil, progress.is(restoreStared) && !progress.is(restoreDone))
@@ -1650,7 +1819,7 @@ func (r *PhysRestore) agreeCommonRestoreTS() (primitive.Timestamp, error) {
 	}
 
 	if r.nodeInfo.IsClusterLeader() {
-		_, err := r.waitFiles(defs.StatusExtTS, copyMap(r.syncPathShards), true)
+		_, err := r.waitFiles(defs.StatusExtTS, maps.Clone(r.syncPathShards), true)
 		if err != nil {
 			return ts, errors.Wrap(err, "wait for shards timestamp")
 		}
@@ -1699,7 +1868,7 @@ func (r *PhysRestore) setcommittedTxn(_ context.Context, txn []phys.RestoreTxn)
 }
 
 func (r *PhysRestore) getcommittedTxn(context.Context) (map[string]primitive.Timestamp, error) {
-	shards := copyMap(r.syncPathShards)
+	shards := maps.Clone(r.syncPathShards)
 	txn := make(map[string]primitive.Timestamp)
 	for len(shards) > 0 {
 		for f := range shards {
@@ -1787,8 +1956,6 @@ func tryConn(port int, logpath string) (*mongo.Client, error) {
 	return nil, errors.Errorf("failed to  connect after %d tries: %v", tryConnCount, err)
 }
 
-const internalMongodLog = "pbm.restore.log"
-
 func (r *PhysRestore) startMongo(opts ...string) error {
 	if r.tmpConf != nil {
 		opts = append(opts, []string{"-f", r.tmpConf.Name()}...)
@@ -2309,7 +2476,7 @@ func (r *PhysRestore) checkMongod(needVersion string) (version string, err error
 }
 
 // MarkFailed sets the restore and rs state as failed with the given message
-func (r *PhysRestore) MarkFailed(meta *RestoreMeta, e error, markCluster bool) {
+func (r *PhysRestore) MarkFailed(meta *RestoreMeta, e error) {
 	var nerr nodeError
 	if errors.As(e, &nerr) {
 		e = nerr
@@ -2332,14 +2499,16 @@ func (r *PhysRestore) MarkFailed(meta *RestoreMeta, e error, markCluster bool) {
 	// At some point, every node will try to set an rs and cluster state
 	// (in `toState` method).
 	// Here we are not aware of partlyDone etc so leave it to the `toState`.
-	if r.nodeInfo.IsPrimary && markCluster {
+	if r.checkForRSLevelErr() {
+		r.log.Debug("set error on rs level")
 		serr := util.RetryableWrite(r.stg,
 			r.syncPathRS+"."+string(defs.StatusError), errStatus(e))
 		if serr != nil {
 			r.log.Error("MarkFailed: write replset error state `%v`: %v", e, serr)
 		}
 	}
-	if r.nodeInfo.IsClusterLeader() && markCluster {
+	if r.nodeInfo.IsLeader() && r.checkForClusterLevelErr() {
+		r.log.Debug("set error on cluster level")
 		serr := util.RetryableWrite(r.stg,
 			r.syncPathCluster+"."+string(defs.StatusError), errStatus(e))
 		if serr != nil {
@@ -2348,7 +2517,33 @@ func (r *PhysRestore) MarkFailed(meta *RestoreMeta, e error, markCluster bool) {
 	}
 }
 
-func removeAll(dir string, l log.LogEvent) error {
+// moveAll moves fromDir content (files and dirs) to toDir content.
+// It ignores all files/dirs specified within toIgnore slice.
+func moveAll(fromDir, toDir string, toIgnore []string, l log.LogEvent) error {
+	d, err := os.Open(fromDir)
+	if err != nil {
+		return errors.Wrap(err, "open dir")
+	}
+	defer d.Close()
+
+	names, err := d.Readdirnames(-1)
+	if err != nil {
+		return errors.Wrap(err, "read file names")
+	}
+	for _, n := range names {
+		if slices.Contains(toIgnore, n) {
+			continue
+		}
+		err = os.Rename(filepath.Join(fromDir, n), filepath.Join(toDir, n))
+		if err != nil {
+			return errors.Wrapf(err, "move %s", n)
+		}
+		l.Debug("move to %s dir: %s", toDir, n)
+	}
+	return nil
+}
+
+func removeAll(dir string, toIgnore []string, l log.LogEvent) error {
 	d, err := os.Open(dir)
 	if err != nil {
 		return errors.Wrap(err, "open dir")
@@ -2360,7 +2555,7 @@ func removeAll(dir string, l log.LogEvent) error {
 		return errors.Wrap(err, "read file names")
 	}
 	for _, n := range names {
-		if n == internalMongodLog {
+		if n == internalMongodLog || slices.Contains(toIgnore, n) {
 			continue
 		}
 		err = os.RemoveAll(filepath.Join(dir, n))