Merge pull request #601 from jgehrcke/jp/perm-error-in-prepare-path

Return strict-decode errors in NodePrepareResources() as permanentError

Author: jgehrcke

cmd/compute-domain-kubelet-plugin/device_state.go

@@ -509,7 +509,7 @@ func (s *DeviceState) getConfigResultsMap(rcs *resourceapi.ResourceClaimStatus,
 		rcs.Allocation.Devices.Config,
 	)
 	if err != nil {
-		return nil, fmt.Errorf("error getting opaque device configs: %v", err)
+		return nil, fmt.Errorf("error getting opaque device configs: %w", err)
 	}
 
 	// Add the default ComputeDomainConfig to the front of the config list with the
@@ -613,7 +613,11 @@ func GetOpaqueDeviceConfigs(
 
 		decodedConfig, err := runtime.Decode(decoder, config.Opaque.Parameters.Raw)
 		if err != nil {
-			return nil, fmt.Errorf("error decoding config parameters: %w", err)
+			// Bad opaque config: i) do not retry preparing this resource
+			// internally and ii) return notion of permanent error to kubelet,
+			// to give it an opportunity to play this error back to the user so
+			// that it becomes actionable.
+			return nil, permanentError{fmt.Errorf("error decoding config parameters: %w", err)}
 		}
 
 		resultConfig := &OpaqueDeviceConfig{

cmd/compute-domain-kubelet-plugin/driver.go

@@ -221,6 +221,10 @@ func (d *driver) HandleError(ctx context.Context, err error, msg string) {
 	runtime.HandleErrorWithContext(ctx, err, msg)
 }
 
+// nodePrepareResource() returns a 2-tuple; the first value is a boolean
+// indicating whether the work is 'done', the second value is a result which can
+// also reflect an error. Set the boolean to `true` for any result wrapping a
+// non-retryable error.
 func (d *driver) nodePrepareResource(ctx context.Context, claim *resourceapi.ResourceClaim) (bool, kubeletplugin.PrepareResult) {
 	release, err := d.pulock.Acquire(ctx, flock.WithTimeout(10*time.Second))
 	if err != nil {