retry net.ErrClosed by default (#2949)

Author: lucix-aws

.changelog/b2213048b0dd4619bd611e41eb07bcb0.json

@@ -0,0 +1,8 @@
+{
+    "id": "b2213048-b0dd-4619-bd61-1e41eb07bcb0",
+    "type": "bugfix",
+    "description": "Retry transient net.ErrClosed errors by default.",
+    "modules": [
+        "."
+    ]
+}

aws/retry/retryable_error.go

@@ -116,7 +116,13 @@ func (r RetryableConnectionError) IsErrorRetryable(err error) aws.Ternary {
 	case errors.As(err, &conErr) && conErr.ConnectionError():
 		retryable = true
 
+	case strings.Contains(err.Error(), "use of closed network connection"):
+		fallthrough
 	case strings.Contains(err.Error(), "connection reset"):
+		// The errors "connection reset" and "use of closed network connection"
+		// are effectively the same. It appears to be the difference between
+		// sync and async read of TCP RST in the stdlib's net.Conn read loop.
+		// see #2737
 		retryable = true
 
 	case errors.As(err, &urlErr):

aws/retry/retryable_error_test.go

@@ -136,6 +136,10 @@ func TestRetryConnectionErrors(t *testing.T) {
 				Err: fmt.Errorf("some error %w", mockTimeoutError{b: true}),
 			},
 		},
+		"net.ErrClosed": {
+			Retryable: aws.TrueTernary,
+			Err:       net.ErrClosed,
+		},
 	}
 
 	for name, c := range cases {