xds: skip DiscoveryRequest for unsubscribed types on stream ready

[paulmurhy123]

TL;DR

When the bootstrap declares more than one xDS server (e.g. a default server for LDS/CDS plus an authority-specific EDS-only server), grpc-java was sending CDS/LDS DiscoveryRequests to the EDS-only server too. That server replies UNIMPLEMENTED, which tears down the stream and EDS data never arrives. Fix: skip DiscoveryRequests for resource types we don't actually subscribe to on a given server.

Problem

Under A47 — xDS Federation, authorities can declare their own xds_servers block in the bootstrap. When an ADS stream is opened to an authority-specific server, ControlPlaneClient.adjustResourceSubscription sends a DiscoveryRequest for every globally-subscribed resource type — even types that have no subscription for this server. The empty request still carries a type_url, and an authority-specific server (e.g. an EDS-only control plane) may reject it with UNIMPLEMENTED, which tears down the entire stream before the legitimate request that follows can complete.

Reproduce

Bootstrap declares two servers — call them control-plane-A (handles LDS/CDS for authority A) and control-plane-B (handles EDS only for authority B).

A grpc-java channel that resolves through LDS → CDS in authority A and EDS in authority B opens an ADS stream to control-plane-B. When that stream becomes ready, sendDiscoveryRequests iterates resourceStore.getSubscribedResourceTypesWithTypeUrl() — which returns all three types (Listener, Cluster, Endpoint) — and calls adjustResourceSubscription for each. For Listener and Cluster, getSubscribedResources(serverInfo, type) returns null/empty, but the request is still sent on the wire:

io.grpc.StatusRuntimeException: UNAVAILABLE: Error retrieving EDS resource ...:
  UNIMPLEMENTED. Details: Watches for type type.googleapis.com/envoy.config.cluster.v3.Cluster
  are not supported in this service

The Cluster (CDS) request reaches control-plane-B, gets rejected, and the stream goes into backoff with no EDS data ever delivered. grpc-go on the same bootstrap works fine against the same server, which pointed at the asymmetry.

Fix

In adjustResourceSubscription, return early when both:

1. resources is null (the store reports no subscription for this type on this server), and
2. No DiscoveryRequest of this type has been sent on the current stream (!adsStream.sentTypes.contains(resourceType)).

Per the ResourceStore contract in XdsClient.java, a null return means "no subscription", while an empty collection means a wildcard subscription — a real subscription that must still emit the empty resource_names request and start its missing-resource timers.

The "DiscoveryRequest of this type has been sent on the current stream" discriminator is tracked in a per-stream sentTypes set on AdsStream, populated wherever a request is actually transmitted (initial sends, ACKs, NACKs). Per-stream is the right scope because the server's view of our subscriptions is per-stream — on stream replacement the set is cleared implicitly along with the AdsStream instance.

Gating on !versions.containsKey(type) instead would be incorrect because versions is only populated on ACK. If a watch is canceled after the initial DiscoveryRequest goes out but before any response is ACKed, that guard would suppress the empty unsubscribe — leaving the server with a stale subscription until the stream resets. Tracking actual sends per-stream closes that window.

The unsubscribe-all path (had-version, now-empty/null) is preserved: we still send the empty request and clear the version, telling the server to drop our subscription for that type.

Mirrors grpc-go

This brings grpc-java in line with grpc-go's behavior. The Go equivalent is adsStreamImpl.sendExisting:

for typ, state := range s.resourceTypeState {
    state.nonce = ""

    if len(state.subscribedResources) == 0 {
        continue                              // <-- explicit skip
    }

    names := resourceNames(state.subscribedResources)
    if err := s.sendMessageLocked(stream, names, typ.TypeURL, state.version, state.nonce, nil); err != nil {
        return err
    }
    s.startWatchTimersLocked(typ, names)
}

Two structural reasons grpc-go avoids this bug today:

1. The iteration domain is per-stream (s.resourceTypeState), populated only when subscribe is called for a resource on this stream — so a Cluster type never even appears in the iteration of an EDS-only stream.
2. Even within that per-stream iteration, the explicit if len(state.subscribedResources) == 0 { continue } covers the case where the type has no subscription on this stream.

The grpc-java fix is the equivalent of (2). The !adsStream.sentTypes.contains guard is needed because Java's iteration domain is global (getSubscribedResourceTypesWithTypeUrl is xds-client-wide), so we may see types we never subscribed to on this stream.

Note that grpc-go physically separates two paths: sendNewLocked handles runtime sub/unsub and sends every queued request unconditionally (so an empty unsubscribe always goes on the wire, ACK or no ACK), while sendExisting handles stream re-establishment and applies the len == 0 skip. grpc-java has a single adjustResourceSubscription function that serves both paths — the per-stream sentTypes set is what lets the same guard distinguish them: on a fresh stream sentTypes is empty so the guard reduces to "no subscription → skip" (mirroring sendExisting), while at runtime after the initial request sentTypes.contains(type) is true so the guard does not trigger and the empty unsubscribe is sent (mirroring sendNewLocked).

Test plan

Unit tests in ControlPlaneClientTest:

• streamReady_skipsEmptyDiscoveryRequestForUnsubscribedType — the federation case, asserts CDS request is suppressed and EDS still goes through
• streamReady_sendsRequestForAllTypesWhenAllHaveResources — guards against over-eager skip
• streamReady_skipsTypeWithNoSubscription — null return skips
• streamReady_sendsWildcardRequestAndStartsTimers — empty collection (wildcard) still sends and starts timers
• cancelBeforeAck_sendsEmptyUnsubscribe — cancel-before-ACK timing window still emits the unsubscribe

Spec note

xDS SoTW spec (Envoy xDS protocol) treats an empty resource_names for LDS/CDS as a wildcard subscription ("send me everything of this type"). The previous grpc-java behavior would unintentionally trigger wildcard CDS subscriptions on every authority-specific stream — which an EDS-only server is right to refuse. Skipping when no subscription exists side-steps that misinterpretation; legitimate wildcard subscriptions (empty collection from a real subscription) still go on the wire as intended, and the existing unsubscribe-all path (with a prior version) continues to work.

[linux-foundation-easycla]

The committers listed above are authorized under a signed CLA.

• ✅ login: paulmurhy123 / name: paul.murphy (bbbdab1)

[Copilot]

Pull request overview

Fixes xDS federation behavior in ControlPlaneClient so that when an ADS stream becomes ready, grpc-java does not send DiscoveryRequests for resource types that have no subscription on that specific server, avoiding UNIMPLEMENTED errors from authority-specific (e.g., EDS-only) control planes that would otherwise tear down the stream.

Changes:

• Add a per-ADS-stream sentTypes tracker and use it to skip sending initial empty DiscoveryRequests for unsubscribed types on stream ready.
• Preserve correct unsubscribe behavior by still sending empty requests when a type was previously sent on the stream (including the cancel-before-ACK window).
• Add unit tests covering federation, wildcard subscription, and unsubscribe timing behavior.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
xds/src/main/java/io/grpc/xds/client/ControlPlaneClient.java	Skips initial requests for types with null subscription on a server unless that type was previously sent on the current stream; tracks per-stream sent types.
xds/src/test/java/io/grpc/xds/client/ControlPlaneClientTest.java	Adds focused unit tests validating the new skip behavior, wildcard handling, and cancel-before-ACK unsubscribe correctness.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[kannanjgithub]

/gcbrun

[paulmurhy123]

@kannanjgithub
Thank you for taking a look at this.
What is the ETA for this being merged?