diff --git a/block.go b/block.go
index b5b2f3ee5..1d5ebfb10 100644
--- a/block.go
+++ b/block.go
@@ -178,6 +178,8 @@ type BlockMetaCompaction struct {
 	Level int `json:"level"`
 	// ULIDs of all source head blocks that went into the block.
 	Sources []ulid.ULID `json:"sources,omitempty"`
+	// Short descriptions of the direct blocks that were used to create
+	// this block.
 	Parents []BlockDesc `json:"parents,omitempty"`
 	Failed  bool        `json:"failed,omitempty"`
 }
diff --git a/db.go b/db.go
index f81c2e8fa..394181089 100644
--- a/db.go
+++ b/db.go
@@ -326,6 +326,13 @@ func (a dbAppender) Commit() error {
 	return err
 }
 
+// Compact data if possible. After successful compaction blocks are reloaded
+// which will also trigger blocks to be deleted that fall out of the retention
+// window.
+// If no blocks are compacted, the retention window state doesn't change. Thus,
+// this is sufficient to reliably delete old data.
+// Old blocks are only deleted on reload based on the new block's parent information.
+// See DB.reload documentation for further information.
 func (db *DB) compact() (changes bool, err error) {
 	db.cmtx.Lock()
 	defer db.cmtx.Unlock()
@@ -426,8 +433,13 @@ func (db *DB) reload() (err error) {
 	// We delete old blocks that have been superseded by new ones by gathering all parents
 	// from existing blocks. Those parents all have newer replacements and can be safely deleted
 	// after we loaded the other blocks.
+	// This makes us resilient against the process crashing towards the end of a compaction.
+	// Creation of a new block and deletion of its parents cannot happen atomically. By creating
+	// blocks atomically with their parents, we can pick up the deletion where it left off during
+	// a crash.
 	var (
 		blocks     []*Block
+		corrupted  = map[ulid.ULID]error{}
 		opened     = map[ulid.ULID]struct{}{}
 		deleteable = map[ulid.ULID]struct{}{}
 	)
@@ -437,6 +449,13 @@ func (db *DB) reload() (err error) {
 			// The block was potentially in the middle of being deleted during a crash.
 			// Skip it since we may delete it properly further down again.
 			level.Warn(db.logger).Log("msg", "read meta information", "err", err, "dir", dir)
+
+			ulid, err2 := ulid.Parse(filepath.Base(dir))
+			if err2 != nil {
+				level.Error(db.logger).Log("msg", "not a block dir", "dir", dir)
+				continue
+			}
+			corrupted[ulid] = err
 			continue
 		}
 		if db.beyondRetention(meta) {
@@ -447,11 +466,17 @@ func (db *DB) reload() (err error) {
 			deleteable[b.ULID] = struct{}{}
 		}
 	}
+	// Blocks we failed to open should all be those we are want to delete anyway.
+	for c, err := range corrupted {
+		if _, ok := deleteable[c]; !ok {
+			return errors.Wrapf(err, "unexpected corrupted block %s", c)
+		}
+	}
+	// Load new blocks into memory.
 	for _, dir := range dirs {
 		meta, err := readMetaFile(dir)
 		if err != nil {
-			level.Warn(db.logger).Log("msg", "read meta information", "err", err, "dir", dir)
-			continue
+			return errors.Wrapf(err, "read meta information %s", dir)
 		}
 		// Don't load blocks that are scheduled for deletion.
 		if _, ok := deleteable[meta.ULID]; ok {