Store: improve index header reading performance by sorting values first

Signed-off-by: Xiaochao Dong (@damnever) <the.xcdong@gmail.com>

CHANGELOG.md

@@ -40,6 +40,7 @@ We use *breaking :warning:* to mark changes that are not backward compatible (re
 - [#5451](https://github.com/thanos-io/thanos/pull/5451) Azure: Reduce memory usage by not buffering file downloads entirely in memory.
 - [#5484](https://github.com/thanos-io/thanos/pull/5484) Update Prometheus deps to v2.36.2.
 - [#5511](https://github.com/thanos-io/thanos/pull/5511) Update Prometheus deps to v2.37.0.
+- [#5588](https://github.com/thanos-io/thanos/pull/5588) Store: improve index header reading performance by sorting values first.
 
 ### Removed
 

pkg/store/bucket.go

@@ -1934,13 +1934,18 @@ func checkNilPosting(l labels.Label, p index.Postings) index.Postings {
 
 // NOTE: Derived from tsdb.postingsForMatcher. index.Merge is equivalent to map duplication.
 func toPostingGroup(lvalsFn func(name string) ([]string, error), m *labels.Matcher) (*postingGroup, error) {
-	if m.Type == labels.MatchRegexp && len(findSetMatches(m.Value)) > 0 {
-		vals := findSetMatches(m.Value)
-		toAdd := make([]labels.Label, 0, len(vals))
-		for _, val := range vals {
-			toAdd = append(toAdd, labels.Label{Name: m.Name, Value: val})
+	if m.Type == labels.MatchRegexp {
+		if vals := findSetMatches(m.Value); len(vals) > 0 {
+			// Sorting will improve the performance dramatically if the dataset is relatively large
+			// since entries in the postings offset table was sorted by label name and value,
+			// the sequential reading is much faster.
+			sort.Strings(vals)
+			toAdd := make([]labels.Label, 0, len(vals))
+			for _, val := range vals {
+				toAdd = append(toAdd, labels.Label{Name: m.Name, Value: val})
+			}
+			return newPostingGroup(false, toAdd, nil), nil
 		}
-		return newPostingGroup(false, toAdd, nil), nil
 	}
 
 	// If the matcher selects an empty value, it selects all the series which don't