Avoid double compaction by cleaning partition files in 2 cycles

CHANGELOG.md

@@ -9,6 +9,7 @@
 * [ENHANCEMENT] Distributor: Add a label references validation for remote write v2 request. #7074
 * [ENHANCEMENT] Distributor: Add count, spans, and buckets validations for native histogram. #7072
 * [ENHANCEMENT] Ruler: Add DecodingConcurrency config flag for Thanos Engine. #7118
+* [ENHANCEMENT] Compactor: Avoid double compaction by cleaning partition files in 2 cycles. #7129
 * [BUGFIX] Ring: Change DynamoDB KV to retry indefinitely for WatchKey. #7088
 * [BUGFIX] Ruler: Add XFunctions validation support. #7111
 

pkg/compactor/blocks_cleaner.go

@@ -609,6 +609,12 @@ func (c *BlocksCleaner) cleanUser(ctx context.Context, userLogger log.Logger, us
 		c.tenantCleanDuration.WithLabelValues(userID).Set(time.Since(startTime).Seconds())
 	}()
 
+	if c.cfg.ShardingStrategy == util.ShardingStrategyShuffle && c.cfg.CompactionStrategy == util.CompactionStrategyPartitioning {
+		begin := time.Now()
+		c.cleanPartitionedGroupInfo(ctx, userBucket, userLogger, userID)
+		level.Info(userLogger).Log("msg", "finish cleaning partitioned group info files", "duration", time.Since(begin), "duration_ms", time.Since(begin).Milliseconds())
+	}
+
 	// Migrate block deletion marks to the global markers location. This operation is a best-effort.
 	if firstRun && c.cfg.BlockDeletionMarksMigrationEnabled {
 		if err := bucketindex.MigrateBlockDeletionMarksToGlobalLocation(ctx, c.bucketClient, userID, c.cfgProvider); err != nil {
@@ -753,12 +759,6 @@ func (c *BlocksCleaner) cleanUser(ctx context.Context, userLogger log.Logger, us
 		level.Info(userLogger).Log("msg", "finish writing new index", "duration", time.Since(begin), "duration_ms", time.Since(begin).Milliseconds())
 	}
 	c.updateBucketMetrics(userID, parquetEnabled, idx, float64(len(partials)), float64(totalBlocksBlocksMarkedForNoCompaction))
-
-	if c.cfg.ShardingStrategy == util.ShardingStrategyShuffle && c.cfg.CompactionStrategy == util.CompactionStrategyPartitioning {
-		begin = time.Now()
-		c.cleanPartitionedGroupInfo(ctx, userBucket, userLogger, userID)
-		level.Info(userLogger).Log("msg", "finish cleaning partitioned group info files", "duration", time.Since(begin), "duration_ms", time.Since(begin).Milliseconds())
-	}
 	return nil
 }
 
@@ -787,28 +787,35 @@ func (c *BlocksCleaner) cleanPartitionedGroupInfo(ctx context.Context, userBucke
 	}
 
 	for partitionedGroupInfo, extraInfo := range existentPartitionedGroupInfo {
+		isPartitionGroupInfoDeleted := false
 		partitionedGroupInfoFile := extraInfo.path
 
 		if extraInfo.status.CanDelete {
 			if extraInfo.status.IsCompleted {
 				// Try to remove all blocks included in partitioned group info
-				if err := partitionedGroupInfo.markAllBlocksForDeletion(ctx, userBucket, userLogger, c.blocksMarkedForDeletion, userID); err != nil {
+				deletedBlocksCount, err := partitionedGroupInfo.markAllBlocksForDeletion(ctx, userBucket, userLogger, c.blocksMarkedForDeletion, userID)
+				if err != nil {
 					level.Warn(userLogger).Log("msg", "unable to mark all blocks in partitioned group info for deletion", "partitioned_group_id", partitionedGroupInfo.PartitionedGroupID)
 					// if one block can not be marked for deletion, we should
 					// skip delete this partitioned group. next iteration
 					// would try it again.
 					continue
 				}
-			}
-
-			if err := userBucket.Delete(ctx, partitionedGroupInfoFile); err != nil {
-				level.Warn(userLogger).Log("msg", "failed to delete partitioned group info", "partitioned_group_info", partitionedGroupInfoFile, "err", err)
-			} else {
-				level.Info(userLogger).Log("msg", "deleted partitioned group info", "partitioned_group_info", partitionedGroupInfoFile)
+				if deletedBlocksCount > 0 {
+					level.Info(userLogger).Log("msg", "parent blocks deleted, will delete partition group file in next cleaning cycle", "partitioned_group_id", partitionedGroupInfo.PartitionedGroupID)
+				} else {
+					level.Info(userLogger).Log("msg", "deleting partition group now that all associated blocks have been deleted", "partitioned_group_id", partitionedGroupInfo.PartitionedGroupID)
+					if err := userBucket.Delete(ctx, partitionedGroupInfoFile); err != nil {
+						level.Warn(userLogger).Log("msg", "failed to delete partitioned group info", "partitioned_group_info", partitionedGroupInfoFile, "err", err)
+					} else {
+						level.Info(userLogger).Log("msg", "deleted partitioned group info", "partitioned_group_info", partitionedGroupInfoFile)
+						isPartitionGroupInfoDeleted = true
+					}
+				}
 			}
 		}
 
-		if extraInfo.status.CanDelete || extraInfo.status.DeleteVisitMarker {
+		if isPartitionGroupInfoDeleted && (extraInfo.status.CanDelete || extraInfo.status.DeleteVisitMarker) {
 			// Remove partition visit markers
 			if _, err := bucket.DeletePrefix(ctx, userBucket, GetPartitionVisitMarkerDirectoryPath(partitionedGroupInfo.PartitionedGroupID), userLogger, defaultDeleteBlocksConcurrency); err != nil {
 				level.Warn(userLogger).Log("msg", "failed to delete partition visit markers for partitioned group", "partitioned_group_info", partitionedGroupInfoFile, "err", err)

pkg/compactor/blocks_cleaner_test.go

@@ -925,7 +925,24 @@ func TestBlocksCleaner_CleanPartitionedGroupInfo(t *testing.T) {
 	dummyGaugeVec := prometheus.NewGaugeVec(prometheus.GaugeOpts{}, []string{"test"})
 
 	cleaner := NewBlocksCleaner(cfg, bucketClient, scanner, 60*time.Second, cfgProvider, logger, "test-cleaner", reg, time.Minute, 30*time.Second, blocksMarkedForDeletion, dummyGaugeVec)
-
+	idx := &bucketindex.Index{
+		Blocks: bucketindex.Blocks{
+			{
+				ID:      block1,
+				MinTime: startTime,
+				MaxTime: endTime,
+				Parquet: &parquet.ConverterMarkMeta{},
+			},
+			{
+				ID:      block2,
+				MinTime: startTime,
+				MaxTime: endTime,
+				Parquet: &parquet.ConverterMarkMeta{},
+			},
+		},
+	}
+	err = bucketindex.WriteIndex(ctx, bucketClient, userID, nil, idx)
+	require.NoError(t, err)
 	userBucket := bucket.NewUserBucketClient(userID, bucketClient, cfgProvider)
 
 	partitionedGroupInfo := PartitionedGroupInfo{
@@ -955,11 +972,17 @@ func TestBlocksCleaner_CleanPartitionedGroupInfo(t *testing.T) {
 	err = visitMarkerManager.updateVisitMarker(ctx)
 	require.NoError(t, err)
 
-	cleaner.cleanPartitionedGroupInfo(ctx, userBucket, logger, userID)
+	// first cleaning cycle deletes only the blocks
+	err = cleaner.cleanUser(ctx, logger, userBucket, userID, false)
+	require.NoError(t, err)
+
+	idx, err = bucketindex.ReadIndex(ctx, bucketClient, userID, cfgProvider, logger)
+	require.NoError(t, err)
+	require.Equal(t, []ulid.ULID{block1}, idx.BlockDeletionMarks.GetULIDs())
 
 	partitionedGroupFileExists, err := userBucket.Exists(ctx, GetPartitionedGroupFile(partitionedGroupID))
 	require.NoError(t, err)
-	require.False(t, partitionedGroupFileExists)
+	require.True(t, partitionedGroupFileExists)
 
 	block1DeletionMarkerExists, err := userBucket.Exists(ctx, path.Join(block1.String(), metadata.DeletionMarkFilename))
 	require.NoError(t, err)
@@ -968,6 +991,14 @@ func TestBlocksCleaner_CleanPartitionedGroupInfo(t *testing.T) {
 	block2DeletionMarkerExists, err := userBucket.Exists(ctx, path.Join(block2.String(), metadata.DeletionMarkFilename))
 	require.NoError(t, err)
 	require.False(t, block2DeletionMarkerExists)
+
+	// second cleaning cycle deletes the partition group info after all blocks are deleted
+	err = cleaner.cleanUser(ctx, logger, userBucket, userID, false)
+	require.NoError(t, err)
+
+	partitionedGroupFileExists, err = userBucket.Exists(ctx, GetPartitionedGroupFile(partitionedGroupID))
+	require.NoError(t, err)
+	require.False(t, partitionedGroupFileExists)
 }
 
 func TestBlocksCleaner_DeleteEmptyBucketIndex(t *testing.T) {

pkg/compactor/partitioned_group_info.go

@@ -234,7 +234,7 @@ func (p *PartitionedGroupInfo) isBlockNoCompact(ctx context.Context, userBucket
 	return noCompactMarkerExists
 }
 
-func (p *PartitionedGroupInfo) markAllBlocksForDeletion(ctx context.Context, userBucket objstore.InstrumentedBucket, userLogger log.Logger, blocksMarkedForDeletion *prometheus.CounterVec, userID string) error {
+func (p *PartitionedGroupInfo) markAllBlocksForDeletion(ctx context.Context, userBucket objstore.InstrumentedBucket, userLogger log.Logger, blocksMarkedForDeletion *prometheus.CounterVec, userID string) (int, error) {
 	blocks := p.getAllBlocks()
 	deleteBlocksCount := 0
 	defer func() {
@@ -244,13 +244,13 @@ func (p *PartitionedGroupInfo) markAllBlocksForDeletion(ctx context.Context, use
 		if p.doesBlockExist(ctx, userBucket, userLogger, blockID) && !p.isBlockDeleted(ctx, userBucket, userLogger, blockID) && !p.isBlockNoCompact(ctx, userBucket, userLogger, blockID) {
 			if err := block.MarkForDeletion(ctx, userLogger, userBucket, blockID, "delete block during partitioned group completion check", blocksMarkedForDeletion.WithLabelValues(userID, reasonValueRetention)); err != nil {
 				level.Warn(userLogger).Log("msg", "unable to mark block for deletion", "partitioned_group_id", p.PartitionedGroupID, "block", blockID.String())
-				return err
+				return deleteBlocksCount, err
 			}
 			deleteBlocksCount++
 			level.Debug(userLogger).Log("msg", "marked block for deletion during partitioned group info clean up", "partitioned_group_id", p.PartitionedGroupID, "block", blockID.String())
 		}
 	}
-	return nil
+	return deleteBlocksCount, nil
 }
 
 func (p *PartitionedGroupInfo) String() string {