Solidify runc create retries (#18709)

Author: csweichel

components/docker-up/runc-facade/main.go

@@ -9,13 +9,14 @@ import (
 	"os"
 	"os/exec"
 	"syscall"
+	"time"
 
 	"github.com/opencontainers/runtime-spec/specs-go"
 	"github.com/sirupsen/logrus"
 	"golang.org/x/xerrors"
 )
 
-const RETRY = 3
+const RETRY = 10
 
 var (
 	defaultOOMScoreAdj = 1000
@@ -81,22 +82,28 @@ func createAndRunc(runcPath string, log *logrus.Logger) error {
 	if err != nil {
 		return xerrors.Errorf("cannot encode config.json: %w", err)
 	}
-	for _, fn := range []string{"config.json", "/tmp/debug.json"} {
-		err = os.WriteFile(fn, fc, 0644)
-		if err != nil {
-			return xerrors.Errorf("cannot encode config.json: %w", err)
-		}
+	err = os.WriteFile("config.json", fc, 0644)
+	if err != nil {
+		return xerrors.Errorf("cannot encode config.json: %w", err)
 	}
 
 	// See here for more details on why retries are necessary.
 	// https://github.com/gitpod-io/gitpod/issues/12365
 	for i := 0; i <= RETRY; i++ {
-		err = syscall.Exec(runcPath, os.Args, os.Environ())
-		if err == nil {
-			return err
-		} else {
+		err = exec.Command(runcPath, os.Args[1:]...).Run()
+
+		if err != nil {
 			log.WithError(err).Warn("runc failed")
+
+			// runc creation failures can be caused by timing issues with workspacekit/seccomp notify under load.
+			// Easing of on the pressure here lowers the likelihood of that error.
+			// NOTE(cw): glossing over races with delays is bad style, but also pragmatic.
+			//
+			// Context: https://linear.app/gitpod/issue/ENG-797/docker-containers-sometimes-fail-to-start
+			time.Sleep(100 * time.Millisecond)
+			continue
 		}
+		return nil
 	}
 	return xerrors.Errorf("exec %s: %w", runcPath, err)
 }