scatter: check pending peers before scattering (#31691)

close pingcap/tidb#31034

pkg/restore/range.go

@@ -75,8 +75,10 @@ func SortRanges(ranges []rtree.Range, rewriteRules *RewriteRules) ([]rtree.Range
 
 // RegionInfo includes a region and the leader of the region.
 type RegionInfo struct {
-	Region *metapb.Region
-	Leader *metapb.Peer
+	Region       *metapb.Region
+	Leader       *metapb.Peer
+	PendingPeers []*metapb.Peer
+	DownPeers    []*metapb.Peer
 }
 
 // ContainsInterior returns whether the region contains the given key, and also

pkg/restore/split.go

@@ -185,12 +185,27 @@ SplitRegions:
 	return nil
 }
 
-func (rs *RegionSplitter) hasRegion(ctx context.Context, regionID uint64) (bool, error) {
+func (rs *RegionSplitter) hasHealthyRegion(ctx context.Context, regionID uint64) (bool, error) {
 	regionInfo, err := rs.client.GetRegionByID(ctx, regionID)
 	if err != nil {
 		return false, errors.Trace(err)
 	}
-	return regionInfo != nil, nil
+	// the region hasn't get ready.
+	if regionInfo == nil {
+		return false, nil
+	}
+
+	// check whether the region is healthy and report.
+	// TODO: the log may be too verbose. we should use Prometheus metrics once it get ready for BR.
+	for _, peer := range regionInfo.PendingPeers {
+		log.Debug("unhealthy region detected", logutil.Peer(peer), zap.String("type", "pending"))
+	}
+	for _, peer := range regionInfo.DownPeers {
+		log.Debug("unhealthy region detected", logutil.Peer(peer), zap.String("type", "down"))
+	}
+	// we ignore down peers for they are (normally) hard to be fixed in reasonable time.
+	// (or once there is a peer down, we may get stuck at waiting region get ready.)
+	return len(regionInfo.PendingPeers) == 0, nil
 }
 
 func (rs *RegionSplitter) isScatterRegionFinished(ctx context.Context, regionID uint64) (bool, error) {
@@ -218,7 +233,7 @@ func (rs *RegionSplitter) isScatterRegionFinished(ctx context.Context, regionID
 func (rs *RegionSplitter) waitForSplit(ctx context.Context, regionID uint64) {
 	interval := SplitCheckInterval
 	for i := 0; i < SplitCheckMaxRetryTimes; i++ {
-		ok, err := rs.hasRegion(ctx, regionID)
+		ok, err := rs.hasHealthyRegion(ctx, regionID)
 		if err != nil {
 			log.Warn("wait for split failed", zap.Error(err))
 			return

pkg/restore/split_client.go

@@ -172,8 +172,10 @@ func (c *pdClient) GetRegionByID(ctx context.Context, regionID uint64) (*RegionI
 		return nil, nil
 	}
 	return &RegionInfo{
-		Region: region.Meta,
-		Leader: region.Leader,
+		Region:       region.Meta,
+		Leader:       region.Leader,
+		PendingPeers: region.PendingPeers,
+		DownPeers:    region.DownPeers,
 	}, nil
 }