Implementing Regex optimization on the MatchNotRegexp matcher type

pkg/store/bucket.go

@@ -2282,28 +2282,30 @@ func checkNilPosting(l labels.Label, p index.Postings) index.Postings {
 func toPostingGroup(ctx context.Context, lvalsFn func(name string) ([]string, error), m *labels.Matcher) (*postingGroup, error) {
  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, labelsFromSetMatchers(m.Name, vals), nil), nil
  }
  }
 
  // If the matcher selects an empty value, it selects all the series which don't
  // have the label name set too. See: https://github.com/prometheus/prometheus/issues/3575
  // and https://github.com/prometheus/prometheus/pull/3578#issuecomment-351653555.
  if m.Matches("") {
+ var toRemove []labels.Label
+
+ // Inverse of a MatchNotRegexp is MatchRegexp (double negation).
+ // Fast-path for set matching.
+ if m.Type == labels.MatchNotRegexp {
+ if vals := findSetMatches(m.Value); len(vals) > 0 {
+ toRemove = labelsFromSetMatchers(m.Name, vals)
+ return newPostingGroup(true, nil, toRemove), nil
+ }
+ }
+
  vals, err := lvalsFn(m.Name)
  if err != nil {
  return nil, err
  }
 
- var toRemove []labels.Label
  for _, val := range vals {
  if ctx.Err() != nil {
  return nil, ctx.Err()
@@ -2339,6 +2341,18 @@ func toPostingGroup(ctx context.Context, lvalsFn func(name string) ([]string, er
  return newPostingGroup(false, toAdd, nil), nil
 }
 
+func labelsFromSetMatchers(name string, vals []string) []labels.Label {
+ // 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: name, Value: val})
+ }
+ return toAdd
+}
+
 type postingPtr struct {
  keyID int
  ptr index.Range