Avoid buffering to tempfile when pushing with V2

The practice of buffering to a tempfile during a pushing contributes massively
to slow V2 push performance perception. The protocol was actually designed to
avoid precalculation, supporting cut-through data push. This means we can
assemble the layer, calculate its digest and push to the remote endpoint, all
at the same time.

This should increase performance massively on systems with slow disks or IO
bottlenecks.

Signed-off-by: Stephen J Day <stephen.day@docker.com>

graph/graph.go

@@ -2,7 +2,6 @@ package graph
 
 import (
 	"compress/gzip"
-	"crypto/sha256"
 	"encoding/json"
 	"errors"
 	"fmt"
@@ -340,26 +339,6 @@ func (graph *Graph) newTempFile() (*os.File, error) {
 	return ioutil.TempFile(tmp, "")
 }
 
-func bufferToFile(f *os.File, src io.Reader) (int64, digest.Digest, error) {
-	var (
-		h = sha256.New()
-		w = gzip.NewWriter(io.MultiWriter(f, h))
-	)
-	_, err := io.Copy(w, src)
-	w.Close()
-	if err != nil {
-		return 0, "", err
-	}
-	n, err := f.Seek(0, os.SEEK_CUR)
-	if err != nil {
-		return 0, "", err
-	}
-	if _, err := f.Seek(0, 0); err != nil {
-		return 0, "", err
-	}
-	return n, digest.NewDigest("sha256", h), nil
-}
-
 // Delete atomically removes an image from the graph.
 func (graph *Graph) Delete(name string) error {
 	id, err := graph.idIndex.Get(name)

graph/push_v2.go

@@ -2,8 +2,8 @@ package graph
 
 import (
 	"fmt"
+	"io"
 	"io/ioutil"
-	"os"
 
 	"github.com/Sirupsen/logrus"
 	"github.com/docker/distribution"
@@ -199,7 +199,7 @@ func (p *v2Pusher) pushV2Tag(tag string) error {
 func (p *v2Pusher) pushV2Image(bs distribution.BlobService, img *image.Image) (digest.Digest, error) {
 	out := p.config.OutStream
 
-	out.Write(p.sf.FormatProgress(stringid.TruncateID(img.ID), "Buffering to Disk", nil))
+	out.Write(p.sf.FormatProgress(stringid.TruncateID(img.ID), "Preparing", nil))
 
 	image, err := p.graph.Get(img.ID)
 	if err != nil {
@@ -209,52 +209,46 @@ func (p *v2Pusher) pushV2Image(bs distribution.BlobService, img *image.Image) (d
 	if err != nil {
 		return "", err
 	}
-
+	defer arch.Close()
-	tf, err := p.graph.newTempFile()
-	if err != nil {
-		return "", err
-	}
-	defer func() {
-		tf.Close()
-		os.Remove(tf.Name())
-	}()
-
-	size, dgst, err := bufferToFile(tf, arch)
-	if err != nil {
-		return "", err
-	}
 
 	// Send the layer
-	logrus.Debugf("rendered layer for %s of [%d] size", img.ID, size)
 	layerUpload, err := bs.Create(context.Background())
 	if err != nil {
 		return "", err
 	}
 	defer layerUpload.Close()
 
+	digester := digest.Canonical.New()
+	tee := io.TeeReader(arch, digester.Hash())
+
 	reader := progressreader.New(progressreader.Config{
-		In:        ioutil.NopCloser(tf),
+		In:        ioutil.NopCloser(tee), // we'll take care of close here.
 		Out:       out,
 		Formatter: p.sf,
-		Size:      size,
+
-		NewLines:  false,
+		// TODO(stevvooe): This may cause a size reporting error. Try to get
-		ID:        stringid.TruncateID(img.ID),
+		// this from tar-split or elsewhere. The main issue here is that we
-		Action:    "Pushing",
+		// don't want to buffer to disk *just* to calculate the size.
+		Size: img.Size,
+
+		NewLines: false,
+		ID:       stringid.TruncateID(img.ID),
+		Action:   "Pushing",
 	})
-	n, err := layerUpload.ReadFrom(reader)
+
+	out.Write(p.sf.FormatProgress(stringid.TruncateID(img.ID), "Pushing", nil))
+	nn, err := io.Copy(layerUpload, reader)
 	if err != nil {
 		return "", err
 	}
-	if n != size {
-		return "", fmt.Errorf("short upload: only wrote %d of %d", n, size)
-	}
 
-	desc := distribution.Descriptor{Digest: dgst}
+	dgst := digester.Digest()
-	if _, err := layerUpload.Commit(context.Background(), desc); err != nil {
+	if _, err := layerUpload.Commit(context.Background(), distribution.Descriptor{Digest: dgst}); err != nil {
 		return "", err
 	}
 
-	out.Write(p.sf.FormatProgress(stringid.TruncateID(img.ID), "Image successfully pushed", nil))
+	logrus.Debugf("uploaded layer %s (%s), %d bytes", img.ID, dgst, nn)
+	out.Write(p.sf.FormatProgress(stringid.TruncateID(img.ID), "Pushed", nil))
 
 	return dgst, nil
 }

integration-cli/docker_cli_push_test.go

@@ -108,7 +108,7 @@ func (s *DockerRegistrySuite) TestPushInterrupt(c *check.C) {
 	}
 
 	// Interrupt push (yes, we have no idea at what point it will get killed).
-	time.Sleep(200 * time.Millisecond)
+	time.Sleep(50 * time.Millisecond) // dependent on race condition.
 	if err := pushCmd.Process.Kill(); err != nil {
 		c.Fatalf("Failed to kill push process: %v", err)
 	}

pkg/jsonmessage/jsonmessage.go

@@ -67,8 +67,14 @@ func (p *JSONProgress) String() string {
 		}
 		pbBox = fmt.Sprintf("[%s>%s] ", strings.Repeat("=", percentage), strings.Repeat(" ", numSpaces))
 	}
+
 	numbersBox = fmt.Sprintf("%8v/%v", current, total)
 
+	if p.Current > p.Total {
+		// remove total display if the reported current is wonky.
+		numbersBox = fmt.Sprintf("%8v", current)
+	}
+
 	if p.Current > 0 && p.Start > 0 && percentage < 50 {
 		fromStart := time.Now().UTC().Sub(time.Unix(p.Start, 0))
 		perEntry := fromStart / time.Duration(p.Current)

pkg/jsonmessage/jsonmessage_test.go

@@ -3,12 +3,12 @@ package jsonmessage
 import (
 	"bytes"
 	"fmt"
+	"strings"
 	"testing"
 	"time"
 
 	"github.com/docker/docker/pkg/term"
 	"github.com/docker/docker/pkg/timeutils"
-	"strings"
 )
 
 func TestError(t *testing.T) {
@@ -45,7 +45,7 @@ func TestProgress(t *testing.T) {
 	}
 
 	// this number can't be negative gh#7136
-	expected = "[==================================================>]     50 B/40 B"
+	expected = "[==================================================>]     50 B"
 	jp5 := JSONProgress{Current: 50, Total: 40}
 	if jp5.String() != expected {
 		t.Fatalf("Expected %q, got %q", expected, jp5.String())