Merge pull request #430 from stevvooe/address-s3-timeout-error

Attempt to address intermittent s3 RequestTimeout error

registry/storage/driver/s3/s3.go

@@ -28,6 +28,7 @@ import (
 
 	"github.com/AdRoll/goamz/aws"
 	"github.com/AdRoll/goamz/s3"
+	"github.com/Sirupsen/logrus"
 	storagedriver "github.com/docker/distribution/registry/storage/driver"
 	"github.com/docker/distribution/registry/storage/driver/base"
 	"github.com/docker/distribution/registry/storage/driver/factory"
@@ -398,18 +399,64 @@ func (d *driver) WriteStream(path string, offset int64, reader io.Reader) (total
 		go func(bytesRead int, from int64, buf []byte) {
 			defer d.putbuf(buf) // this buffer gets dropped after this call
 
-			// parts and partNumber are safe, because this function is the only one modifying them and we
-			// force it to be executed serially.
-			if bytesRead > 0 {
-				part, putErr := multi.PutPart(int(partNumber), bytes.NewReader(buf[0:int64(bytesRead)+from]))
-				if putErr != nil {
-					putErrChan <- putErr
+			// DRAGONS(stevvooe): There are few things one might want to know
+			// about this section. First, the putErrChan is expecting an error
+			// and a nil or just a nil to come through the channel. This is
+			// covered by the silly defer below. The other aspect is the s3
+			// retry backoff to deal with RequestTimeout errors. Even though
+			// the underlying s3 library should handle it, it doesn't seem to
+			// be part of the shouldRetry function (see AdRoll/goamz/s3).
+			defer func() {
+				putErrChan <- nil // for some reason, we do this no matter what.
+			}()
+
+			if bytesRead <= 0 {
+				return
+			}
+
+			var err error
+			var part s3.Part
+
+		loop:
+			for retries := 0; retries < 5; retries++ {
+				part, err = multi.PutPart(int(partNumber), bytes.NewReader(buf[0:int64(bytesRead)+from]))
+				if err == nil {
+					break // success!
 				}
 
-				parts = append(parts, part)
-				partNumber++
+				// NOTE(stevvooe): This retry code tries to only retry under
+				// conditions where the s3 package does not. We may add s3
+				// error codes to the below if we see others bubble up in the
+				// application. Right now, the most troubling is
+				// RequestTimeout, which seems to only triggered when a tcp
+				// connection to s3 slows to a crawl. If the RequestTimeout
+				// ends up getting added to the s3 library and we don't see
+				// other errors, this retry loop can be removed.
+				switch err := err.(type) {
+				case *s3.Error:
+					switch err.Code {
+					case "RequestTimeout":
+						// allow retries on only this error.
+					default:
+						break loop
+					}
+				}
+
+				backoff := 100 * time.Millisecond * time.Duration(retries+1)
+				logrus.Errorf("error putting part, retrying after %v: %v", err, backoff.String())
+				time.Sleep(backoff)
 			}
-			putErrChan <- nil
+
+			if err != nil {
+				logrus.Errorf("error putting part, aborting: %v", err)
+				putErrChan <- err
+			}
+
+			// parts and partNumber are safe, because this function is the
+			// only one modifying them and we force it to be executed
+			// serially.
+			parts = append(parts, part)
+			partNumber++
 		}(bytesRead, from, buf)
 
 		buf = d.getbuf() // use a new buffer for the next call