perf(validation): optimize overlap checks (alias grouping + identical hash)

Author: pavelnikolov

fragment_queries_bench_test.go

@@ -0,0 +1,182 @@
+//go:build bench
+// +build bench
+
+package graphql_test
+
+// Benchmark inspired by historical PR #102 ("Control memory explosion on large list of queries").
+// The validation phase used to exhibit O(n^2) memory/time growth for large
+// numbers of root selections due to pairwise overlap checks. This benchmark
+// recreates a simplified version focusing on many aliased root fields to
+// exercise overlap validation. Two scenarios are measured:
+//   * overlapping aliases   (all fields share the same alias name)
+//   * non-overlapping aliases (each field has a unique alias)
+// The overlapping case forces the validator to consider merging, potentially
+// increasing work. Non-overlapping aliases avoid that specific merge path.
+//
+// Larger counts (5000, 10000) are included only when BIG_FRAGMENT_BENCH=1 is
+// set in the environment to keep default benchmarks fast.
+
+import (
+	"bytes"
+	"context"
+	"fmt"
+	"os"
+	"testing"
+
+	"github.com/graph-gophers/graphql-go"
+)
+
+const fragmentBenchSchema = `
+    type Query { a: Int! }
+`
+
+type fragmentBenchResolver struct{}
+
+func (fragmentBenchResolver) A() int32 { return 1 }
+
+func buildAliasQuery(count int, nonOverlap bool) string {
+	// Build a single operation with many aliased uses of the same field.
+	// Example (nonOverlap, count=3):
+	// query Q { f0: a f1: a f2: a }
+	// Example (overlap, count=3):
+	// query Q { x: a x: a x: a }
+	q := "query Q {"
+	if nonOverlap {
+		for i := 0; i < count; i++ {
+			q += fmt.Sprintf(" f%d: a", i)
+		}
+	} else {
+		for i := 0; i < count; i++ {
+			q += " x: a" // same alias each time
+		}
+	}
+	q += " }"
+	return q
+}
+
+// BenchmarkSimpleRootAlias measures validation/exec cost for large flat selection sets
+// with and without overlapping aliases.
+func BenchmarkSimpleRootAlias(b *testing.B) {
+	schema := graphql.MustParseSchema(fragmentBenchSchema, &fragmentBenchResolver{})
+
+	counts := []int{1, 10, 100, 500, 1000}
+	if os.Getenv("BIG_FRAGMENT_BENCH") == "1" {
+		counts = append(counts, 5000, 10000)
+	}
+
+	ctx := context.Background()
+
+	for _, c := range counts {
+		for _, nonOverlap := range []bool{false, true} {
+			aliasMode := "overlapping"
+			if nonOverlap {
+				aliasMode = "non-overlapping"
+			}
+			queryStr := buildAliasQuery(c, nonOverlap)
+			// Warm-up single execution (outside timing) to catch schema issues early.
+			if resp := schema.Exec(ctx, queryStr, "", nil); len(resp.Errors) > 0 {
+				b.Fatalf("unexpected exec errors preparing benchmark: %v", resp.Errors)
+			}
+			b.Run(fmt.Sprintf("%d_queries_%s_aliases", c, aliasMode), func(b *testing.B) {
+				b.ReportAllocs()
+				for b.Loop() {
+					if resp := schema.Exec(ctx, queryStr, "", nil); len(resp.Errors) > 0 {
+						b.Fatalf("exec errors: %v", resp.Errors)
+					}
+				}
+			})
+		}
+	}
+}
+
+func TestOverlappingAlias(t *testing.T) {
+	query := `
+		{
+			hero(episode: EMPIRE) {
+				a: name
+				a: id
+			}
+		}
+	`
+	result := starwarsSchema.Exec(context.Background(), query, "", nil)
+	if len(result.Errors) == 0 {
+		t.Fatal("Expected error from overlapping alias")
+	}
+}
+
+// go test -bench=FragmentQueries -benchmem
+// BenchmarkStarWarsFragmentAlias expands the scenario with fragment spreads on the
+// canonical StarWars schema exercising deeper selection hashing paths.
+func BenchmarkStarWarsFragmentAlias(b *testing.B) {
+	singleQuery := `
+		composed_%d: hero(episode: EMPIRE) {
+			name
+			...friendsNames
+			...friendsIds
+		}
+	`
+
+	queryTemplate := `
+		{
+			%s
+		}
+		fragment friendsNames on Character {
+			friends {
+				name
+			}
+		}
+		fragment friendsIds on Character {
+			friends {
+				id
+			}
+		}
+	`
+
+	testCases := []int{
+		1,
+		10,
+		100,
+		1000,
+		10000,
+	}
+
+	for _, c := range testCases {
+		// for each count, add a case for overlapping aliases vs non-overlapping aliases
+		for _, o := range []bool{true} {
+			var buffer bytes.Buffer
+			for i := 0; i < c; i++ {
+				idx := 0
+				if o {
+					idx = i
+				}
+				buffer.WriteString(fmt.Sprintf(singleQuery, idx))
+			}
+
+			query := fmt.Sprintf(queryTemplate, buffer.String())
+			a := "overlapping"
+			if o {
+				a = "non-overlapping"
+			}
+			b.Run(fmt.Sprintf("%d queries %s aliases", c, a), func(b *testing.B) {
+				for n := 0; n < b.N; n++ {
+					result := starwarsSchema.Exec(context.Background(), query, "", nil)
+					if len(result.Errors) != 0 {
+						b.Fatal(result.Errors[0])
+					}
+				}
+			})
+		}
+	}
+}
+
+// Performance mitigation roadmap (see discussion):
+// 1. Implement "early exit after first conflict" in field overlap validation.
+//    - Spec allows returning after first conflict; minimizes quadratic blow-up.
+//    - Lowest risk change; preserves correctness, may slightly reduce number of
+//      reported sibling conflicts (acceptable trade-off for widely used lib).
+// 2. (Optional, behind internal constant) Threshold-based fallback: if conflicts
+//    in a single comparison group exceed a limit, short‑circuit with one
+//    aggregate error to cap worst-case cost.
+// 3. Defer heavier changes (fragment-aware hashing, canonical structural
+//    deduplication) until profiling shows residual hot spots; these add
+//    complexity and require careful determinism/error messaging review.

internal/validation/pathological_overlap_bench_test.go

@@ -0,0 +1,64 @@
+//go:build bench
+// +build bench
+
+package validation_test
+
+import (
+	"fmt"
+	"testing"
+
+	"github.com/graph-gophers/graphql-go/ast"
+	"github.com/graph-gophers/graphql-go/internal/query"
+	"github.com/graph-gophers/graphql-go/internal/schema"
+	"github.com/graph-gophers/graphql-go/internal/validation"
+)
+
+// Builds a query with n repetitions of the same response key whose underlying field differs
+// every other occurrence, forcing overlap validation to perform many pairwise comparisons if
+// optimization is ineffective. Used to evaluate the adaptive identical-field fast path.
+func buildPathologicalQuery(n int) string {
+	// Alternate between a: realField and a: otherField to trigger conflicts only in half pairs.
+	body := "{\n"
+	for i := 0; i < n; i++ {
+		if i%2 == 0 {
+			body += "  x: fieldA { leaf }\n"
+		} else {
+			body += "  x: fieldB { leaf }\n"
+		}
+	}
+	body += "}\n"
+	return body
+}
+
+var pathologicalSchemaSDL = `
+  type Query { fieldA: Obj fieldB: Obj }
+  type Obj { leaf: Int }
+`
+
+func preparePathologicalSchema(b *testing.B) *ast.Schema {
+	s := schema.New()
+	if err := schema.Parse(s, pathologicalSchemaSDL, false); err != nil {
+		b.Fatalf("schema parse: %v", err)
+	}
+	return s
+}
+
+// BenchmarkPathologicalOverlap stresses the worst-case alias group: repeated identical response key
+// with alternating underlying field definitions. Without the identical-fields fast path, this trends
+// toward O(k^2) pairwise comparisons; with optimization it aims for near O(k) (shallow pass + hashes).
+func BenchmarkPathologicalOverlap(b *testing.B) {
+	s := preparePathologicalSchema(b)
+	cases := []int{10, 50, 100, 250, 500}
+	for _, n := range cases {
+		b.Run(fmt.Sprintf("n=%d", n), func(b *testing.B) {
+			queryStr := buildPathologicalQuery(n)
+			for b.Loop() {
+				doc, err := query.Parse(queryStr)
+				if err != nil {
+					b.Fatalf("parse: %v", err)
+				}
+				_ = validation.Validate(s, doc, nil, 0)
+			}
+		})
+	}
+}