Problem description

@grpc/grpc-js-xds can crash when an xDS route uses weighted_clusters, omits the deprecated total_weight field, and the cluster weights do not add up to 100.

For example, a route containing one cluster with weight 1 fails with:

TypeError: Cannot read properties of undefined (reading 'ref')
    at Object.invoke (.../node_modules/@grpc/grpc-js-xds/build/src/resolver-xds.js:343:36)

The same xDS route works with grpcurl/gRPC Go.

The resolver currently defaults an absent total_weight to 100:

routeAction = new WeightedClusterRouteAction(
  weightedClusters,
  route.route!.weighted_clusters!.total_weight?.value ?? 100,
  // ...
);

WeightedClusterRouteAction therefore generates a random value in [0, 100). With one cluster of weight 1, only values in [0, 1) select that cluster. The remaining ~99% return an empty cluster name.

The resolver then looks up that empty name and dereferences undefined:

const clusterResult = action.getCluster();
const clusterRef = this.clusterRefs.get(clusterResult.name)!;
clusterRef.ref();

The generated Envoy API documentation says that total_weight is deprecated and clients should use the sum of all cluster weights.

Relevant code:

• https://github.com/grpc/grpc-node/blob/master/packages/grpc-js-xds/src/resolver-xds.ts
• https://github.com/grpc/grpc-node/blob/master/packages/grpc-js-xds/src/route-action.ts
• https://github.com/grpc/grpc-node/blob/master/packages/grpc-js-xds/src/generated/envoy/config/route/v3/WeightedCluster.ts

Reproduction steps

1. Configure an RDS route similar to:

route:
  weighted_clusters:
    clusters:
      - name: test-cluster
        weight:
          value: 1
    # total_weight intentionally omitted

2. Register @grpc/grpc-js-xds and create an xDS client:

const grpc = require('@grpc/grpc-js');
require('@grpc/grpc-js-xds').register();

const client = new grpc.Client(
  'xds:///test-service',
  grpc.credentials.createInsecure()
);

3. Make repeated RPCs through the route.

4. Most calls fail with:

TypeError: Cannot read properties of undefined (reading 'ref')

Setting the cluster weight to 100, explicitly setting total_weight: 1, or using a non-weighted single-cluster route avoids the failure.

Expected behavior

When total_weight is omitted, the client should use the sum of the supplied cluster weights.

For one cluster with weight 1, that cluster should receive 100% of requests.

The resolver should also return a controlled error instead of throwing if cluster selection somehow produces an unknown or empty cluster name.

Environment

• OS: Linux x86_64, Kubernetes
• Node version: [fill in]
• @grpc/grpc-js version: [fill in]
• @grpc/grpc-js-xds version: [fill in]
• Installation method: npm

Suggested fix

Calculate the denominator from the cluster weights:

const totalWeight = weightedClusters.reduce(
  (sum, cluster) => sum + cluster.weight,
  0
);

Use that value when creating WeightedClusterRouteAction, rather than defaulting an omitted total_weight to 100.

It would also be helpful to guard the clusterRefs.get() result so an invalid selection produces a gRPC status error instead of an uncaught TypeError.

Additional context

A configuration with weights 1 and 99 works because the weights happen to total 100. The failure occurs when the supplied weights have a nonzero sum other than 100 and total_weight is omitted.