Implement chunking for gzip encoder in deferredResponseWriter

Signed-off-by: z1cheng <[email protected]>

Author: z1cheng

staging/src/k8s.io/apiserver/pkg/endpoints/handlers/responsewriters/writers.go

@@ -157,6 +157,9 @@ const (
 	// (usually the entire object), and if the size is smaller no gzipping will be performed
 	// if the client requests it.
 	defaultGzipThresholdBytes = 128 * 1024
+	// Use the length of the first write of streaming implementations.
+	// TODO: Update when streaming proto is implemented
+	firstWriteStreamingThresholdBytes = 1
 )
 
 // negotiateContentEncoding returns a supported client-requested content encoding for the
@@ -192,14 +195,53 @@ type deferredResponseWriter struct {
 	statusCode      int
 	contentEncoding string
 
-	hasWritten bool
-	hw         http.ResponseWriter
-	w          io.Writer
+	hasBuffered bool
+	buffer      []byte
+	hasWritten  bool
+	hw          http.ResponseWriter
+	w           io.Writer
 
 	ctx context.Context
 }
 
 func (w *deferredResponseWriter) Write(p []byte) (n int, err error) {
+	switch {
+	case w.hasWritten:
+		// already written, cannot buffer
+		return w.unbufferedWrite(p)
+
+	case w.contentEncoding != "gzip":
+		// non-gzip, no need to buffer
+		return w.unbufferedWrite(p)
+
+	case !w.hasBuffered && len(p) > defaultGzipThresholdBytes:
+		// not yet buffered, first write is long enough to trigger gzip, no need to buffer
+		return w.unbufferedWrite(p)
+
+	case !w.hasBuffered && len(p) > firstWriteStreamingThresholdBytes:
+		// not yet buffered, first write is longer than expected for streaming scenarios that would require buffering, no need to buffer
+		return w.unbufferedWrite(p)
+
+	default:
+		if !w.hasBuffered {
+			w.hasBuffered = true
+			// Start at 80 bytes to avoid rapid reallocation of the buffer.
+			// The minimum size of a 0-item serialized list object is 80 bytes:
+			// {"kind":"List","apiVersion":"v1","metadata":{"resourceVersion":"1"},"items":[]}\n
+			w.buffer = make([]byte, 0, max(80, len(p)))
+		}
+		w.buffer = append(w.buffer, p...)
+		var err error
+		if len(w.buffer) > defaultGzipThresholdBytes {
+			// we've accumulated enough to trigger gzip, write and clear buffer
+			_, err = w.unbufferedWrite(w.buffer)
+			w.buffer = nil
+		}
+		return len(p), err
+	}
+}
+
+func (w *deferredResponseWriter) unbufferedWrite(p []byte) (n int, err error) {
 	ctx := w.ctx
 	span := tracing.SpanFromContext(ctx)
 	// This Step usually wraps in-memory object serialization.
@@ -245,11 +287,17 @@ func (w *deferredResponseWriter) Write(p []byte) (n int, err error) {
 	return w.w.Write(p)
 }
 
-func (w *deferredResponseWriter) Close() error {
+func (w *deferredResponseWriter) Close() (err error) {
 	if !w.hasWritten {
-		return nil
+		if !w.hasBuffered {
+			return nil
+		}
+		// never reached defaultGzipThresholdBytes, no need to do the gzip writer cleanup
+		_, err := w.unbufferedWrite(w.buffer)
+		w.buffer = nil
+		return err
 	}
-	var err error
+
 	switch t := w.w.(type) {
 	case *gzip.Writer:
 		err = t.Close()

staging/src/k8s.io/apiserver/pkg/endpoints/handlers/responsewriters/writers_test.go

@@ -33,7 +33,6 @@ import (
 	"os"
 	"reflect"
 	"strconv"
-	"strings"
 	"testing"
 	"time"
 
@@ -42,6 +41,7 @@ import (
 	kerrors "k8s.io/apimachinery/pkg/api/errors"
 	"k8s.io/apimachinery/pkg/runtime"
 	"k8s.io/apimachinery/pkg/runtime/schema"
+	rand2 "k8s.io/apimachinery/pkg/util/rand"
 	"k8s.io/apimachinery/pkg/util/uuid"
 	"k8s.io/apiserver/pkg/features"
 	utilfeature "k8s.io/apiserver/pkg/util/feature"
@@ -378,29 +378,94 @@ func TestDeferredResponseWriter_Write(t *testing.T) {
 	largeChunk := bytes.Repeat([]byte("b"), defaultGzipThresholdBytes+1)
 
 	tests := []struct {
-		name       string
-		chunks     [][]byte
-		expectGzip bool
+		name          string
+		chunks        [][]byte
+		expectGzip    bool
+		expectHeaders http.Header
 	}{
+		{
+			name:          "no writes",
+			chunks:        nil,
+			expectGzip:    false,
+			expectHeaders: http.Header{},
+		},
+		{
+			name:       "one empty write",
+			chunks:     [][]byte{{}},
+			expectGzip: false,
+			expectHeaders: http.Header{
+				"Content-Type": []string{"text/plain"},
+			},
+		},
+		{
+			name:       "one single byte write",
+			chunks:     [][]byte{{'{'}},
+			expectGzip: false,
+			expectHeaders: http.Header{
+				"Content-Type": []string{"text/plain"},
+			},
+		},
 		{
 			name:       "one small chunk write",
 			chunks:     [][]byte{smallChunk},
 			expectGzip: false,
+			expectHeaders: http.Header{
+				"Content-Type": []string{"text/plain"},
+			},
 		},
 		{
 			name:       "two small chunk writes",
 			chunks:     [][]byte{smallChunk, smallChunk},
 			expectGzip: false,
+			expectHeaders: http.Header{
+				"Content-Type": []string{"text/plain"},
+			},
+		},
+		{
+			name:       "one single byte and one small chunk write",
+			chunks:     [][]byte{{'{'}, smallChunk},
+			expectGzip: false,
+			expectHeaders: http.Header{
+				"Content-Type": []string{"text/plain"},
+			},
+		},
+		{
+			name:       "two single bytes and one small chunk write",
+			chunks:     [][]byte{{'{'}, {'{'}, smallChunk},
+			expectGzip: true,
+			expectHeaders: http.Header{
+				"Content-Type":     []string{"text/plain"},
+				"Content-Encoding": []string{"gzip"},
+				"Vary":             []string{"Accept-Encoding"},
+			},
 		},
 		{
 			name:       "one large chunk writes",
 			chunks:     [][]byte{largeChunk},
 			expectGzip: true,
+			expectHeaders: http.Header{
+				"Content-Type":     []string{"text/plain"},
+				"Content-Encoding": []string{"gzip"},
+				"Vary":             []string{"Accept-Encoding"},
+			},
 		},
 		{
 			name:       "two large chunk writes",
 			chunks:     [][]byte{largeChunk, largeChunk},
 			expectGzip: true,
+			expectHeaders: http.Header{
+				"Content-Type":     []string{"text/plain"},
+				"Content-Encoding": []string{"gzip"},
+				"Vary":             []string{"Accept-Encoding"},
+			},
+		},
+		{
+			name:       "one small chunk and one large chunk write",
+			chunks:     [][]byte{smallChunk, largeChunk},
+			expectGzip: false,
+			expectHeaders: http.Header{
+				"Content-Type": []string{"text/plain"},
+			},
 		},
 	}
 
@@ -441,23 +506,16 @@ func TestDeferredResponseWriter_Write(t *testing.T) {
 			if res.StatusCode != http.StatusOK {
 				t.Fatalf("status code is not writtend properly, expected: 200, got: %d", res.StatusCode)
 			}
-			contentEncoding := res.Header.Get("Content-Encoding")
-			varyHeader := res.Header.Get("Vary")
+			if !reflect.DeepEqual(res.Header, tt.expectHeaders) {
+				t.Fatal(cmp.Diff(tt.expectHeaders, res.Header))
+			}
 
 			resBytes, err := io.ReadAll(res.Body)
 			if err != nil {
 				t.Fatalf("unexpected error occurred while reading response body: %v", err)
 			}
 
 			if tt.expectGzip {
-				if contentEncoding != "gzip" {
-					t.Fatalf("content-encoding is not set properly, expected: gzip, got: %s", contentEncoding)
-				}
-
-				if !strings.Contains(varyHeader, "Accept-Encoding") {
-					t.Errorf("vary header doesn't have Accept-Encoding")
-				}
-
 				gr, err := gzip.NewReader(bytes.NewReader(resBytes))
 				if err != nil {
 					t.Fatalf("failed to create gzip reader: %v", err)
@@ -471,22 +529,101 @@ func TestDeferredResponseWriter_Write(t *testing.T) {
 				if !bytes.Equal(fullPayload, decompressed) {
 					t.Errorf("payload mismatch, expected: %s, got: %s", fullPayload, decompressed)
 				}
-
 			} else {
-				if contentEncoding != "" {
-					t.Errorf("content-encoding is set unexpectedly")
-				}
-
-				if strings.Contains(varyHeader, "Accept-Encoding") {
-					t.Errorf("accept encoding is set unexpectedly")
-				}
-
 				if !bytes.Equal(fullPayload, resBytes) {
 					t.Errorf("payload mismatch, expected: %s, got: %s", fullPayload, resBytes)
 				}
-
 			}
+		})
+	}
+}
+
+func benchmarkChunkingGzip(b *testing.B, count int, chunk []byte) {
+	mockResponseWriter := httptest.NewRecorder()
+	mockResponseWriter.Body = nil
+
+	drw := &deferredResponseWriter{
+		mediaType:       "text/plain",
+		statusCode:      200,
+		contentEncoding: "gzip",
+		hw:              mockResponseWriter,
+		ctx:             context.Background(),
+	}
+	b.ResetTimer()
+	for i := 0; i < count; i++ {
+		n, err := drw.Write(chunk)
+		if err != nil {
+			b.Fatalf("unexpected error while writing chunk: %v", err)
+		}
+		if n != len(chunk) {
+			b.Errorf("write is not complete, expected: %d bytes, written: %d bytes", len(chunk), n)
+		}
+	}
+	err := drw.Close()
+	if err != nil {
+		b.Fatalf("unexpected error when closing deferredResponseWriter: %v", err)
+	}
+	res := mockResponseWriter.Result()
+	if res.StatusCode != http.StatusOK {
+		b.Fatalf("status code is not writtend properly, expected: 200, got: %d", res.StatusCode)
+	}
+}
+
+func BenchmarkChunkingGzip(b *testing.B) {
+	tests := []struct {
+		count int
+		size  int
+	}{
+		{
+			count: 100,
+			size:  1_000,
+		},
+		{
+			count: 100,
+			size:  100_000,
+		},
+		{
+			count: 1_000,
+			size:  100_000,
+		},
+		{
+			count: 1_000,
+			size:  1_000_000,
+		},
+		{
+			count: 10_000,
+			size:  100_000,
+		},
+		{
+			count: 100_000,
+			size:  10_000,
+		},
+		{
+			count: 1,
+			size:  100_000,
+		},
+		{
+			count: 1,
+			size:  1_000_000,
+		},
+		{
+			count: 1,
+			size:  10_000_000,
+		},
+		{
+			count: 1,
+			size:  100_000_000,
+		},
+		{
+			count: 1,
+			size:  1_000_000_000,
+		},
+	}
 
+	for _, t := range tests {
+		b.Run(fmt.Sprintf("Count=%d/Size=%d", t.count, t.size), func(b *testing.B) {
+			chunk := []byte(rand2.String(t.size))
+			benchmarkChunkingGzip(b, t.count, chunk)
 		})
 	}
 }