Move series fetches out of inner loop of SortedPostings. (#485)

With 1M series:

Before:
BenchmarkHeadPostingForMatchers-8              1        3501996117 ns/op 61311520 B/op         78 allocs/op

After:
BenchmarkHeadPostingForMatchers-8              1        1403072952 ns/op 69261568 B/op         72 allocs/op

This works out as 3X faster, as the above time includes other things.

Signed-off-by: Brian Brazil <brian.brazil@robustperception.io>

head.go

@@ -1096,25 +1096,30 @@ func (h *headIndexReader) Postings(name, value string) (index.Postings, error) {
 }
 
 func (h *headIndexReader) SortedPostings(p index.Postings) index.Postings {
-	ep := make([]uint64, 0, 128)
+	series := make([]*memSeries, 0, 128)
 
+	// Fetch all the series only once.
 	for p.Next() {
-		ep = append(ep, p.At())
+		s := h.head.series.getByID(p.At())
+		if s == nil {
+			level.Debug(h.head.logger).Log("msg", "looked up series not found")
+		} else {
+			series = append(series, s)
+		}
 	}
 	if err := p.Err(); err != nil {
 		return index.ErrPostings(errors.Wrap(err, "expand postings"))
 	}
 
-	sort.Slice(ep, func(i, j int) bool {
-		a := h.head.series.getByID(ep[i])
-		b := h.head.series.getByID(ep[j])
-
-		if a == nil || b == nil {
-			level.Debug(h.head.logger).Log("msg", "looked up series not found")
-			return false
-		}
-		return labels.Compare(a.lset, b.lset) < 0
+	sort.Slice(series, func(i, j int) bool {
+		return labels.Compare(series[i].lset, series[j].lset) < 0
 	})
+
+	// Convert back to list.
+	ep := make([]uint64, 0, len(series))
+	for _, p := range series {
+		ep = append(ep, p.ref)
+	}
 	return index.NewListPostings(ep)
 }
 

head_bench_test.go

@@ -57,7 +57,7 @@ func BenchmarkHeadPostingForMatchers(b *testing.B) {
 	defer h.Close()
 
 	// TODO: vary number of series
-	for i := 0; i < 100; i++ {
+	for i := 0; i < 1000000; i++ {
 		h.getOrCreate(uint64(i), labels.FromStrings("a", strconv.Itoa(i)))
 	}