diff --git a/internal/querycoordv2/job/job_load.go b/internal/querycoordv2/job/job_load.go
index 51598171ce55f2525e5d0a9c5415273134580cd1..4031b76dfe09fcf4b9ae4334a8038a320dedc0fd 100644
--- a/internal/querycoordv2/job/job_load.go
+++ b/internal/querycoordv2/job/job_load.go
@@ -204,9 +204,6 @@ func (job *LoadCollectionJob) Execute() error {
 		return utils.WrapError(msg, err)
 	}
 
-	if !colExisted {
-		metrics.QueryCoordNumCollections.WithLabelValues().Inc()
-	}
 	metrics.QueryCoordNumPartitions.WithLabelValues().Add(float64(len(partitions)))
 	return nil
 }
@@ -380,7 +377,6 @@ func (job *LoadPartitionJob) Execute() error {
 			log.Error(msg, zap.Error(err))
 			return utils.WrapError(msg, err)
 		}
-		metrics.QueryCoordNumCollections.WithLabelValues().Inc()
 	} else { // collection exists, put partitions only
 		err = job.meta.CollectionManager.PutPartition(partitions...)
 		if err != nil {
diff --git a/internal/querycoordv2/meta/collection_manager.go b/internal/querycoordv2/meta/collection_manager.go
index f6194dd952f2de65513e113b9e5bf15be5b0e1b8..8f40f6b4275ebcd3faee55d64f30fb0de2006083 100644
--- a/internal/querycoordv2/meta/collection_manager.go
+++ b/internal/querycoordv2/meta/collection_manager.go
@@ -491,6 +491,11 @@ func (m *CollectionManager) UpdateLoadPercent(partitionID int64, loadPercent int
 		saveCollection = true
 		newCollection.Status = querypb.LoadStatus_Loaded
 		elapsed := time.Since(newCollection.CreatedAt)
+
+		// TODO: what if part of the collection has been unloaded? Now we decrease the metric only after
+		// 	`ReleaseCollection` is triggered. Maybe it's hard to make this metric really accurate.
+		metrics.QueryCoordNumCollections.WithLabelValues().Inc()
+
 		metrics.QueryCoordLoadLatency.WithLabelValues().Observe(float64(elapsed.Milliseconds()))
 	}
 	return collectionPercent, m.putCollection(saveCollection, newCollection)
diff --git a/internal/querycoordv2/server.go b/internal/querycoordv2/server.go
index 8a36f86fd395882ba35a85f8388694ba47ed0345..daf0b03d71bef4bb7340de315dd11fe0b5ac780c 100644
--- a/internal/querycoordv2/server.go
+++ b/internal/querycoordv2/server.go
@@ -311,7 +311,10 @@ func (s *Server) initMeta() error {
 	}
 	collections := s.meta.GetAll()
 	log.Info("recovering collections...", zap.Int64s("collections", collections))
-	metrics.QueryCoordNumCollections.WithLabelValues().Set(float64(len(collections)))
+
+	// We really update the metric after observers think the collection loaded.
+	metrics.QueryCoordNumCollections.WithLabelValues().Set(0)
+
 	metrics.QueryCoordNumPartitions.WithLabelValues().Set(float64(len(s.meta.GetAllPartitions())))
 
 	err = s.meta.ReplicaManager.Recover(collections)