feat: support decentralized router with NATs core (#4921)

Signed-off-by: PeaBrane <yanrpei@gmail.com>

docs/router/kv_cache_routing.md

@@ -42,7 +42,10 @@ The main KV-aware routing arguments:
 - `--router-prune-target-ratio`: Target size ratio to prune down to when `--router-max-tree-size` is exceeded. For example, with a value of 0.8 (default) and max tree size of 1048576, the router will prune down to approximately 838860 blocks when the threshold is exceeded. Defaults to 0.8 when `--no-kv-events` is used. This creates headroom before the next pruning cycle.
 
 >[!Note]
-> State persistence is only available when KV events are enabled (default). When using `--no-kv-events`, state persistence is not currently supported.
+> **State persistence** depends on the event transport mode:
+> - **JetStream mode** (default): State persists across router restarts via JetStream and NATS object store snapshots.
+> - **NATS Core with Local Indexer mode** (`--enable-local-indexer` on workers): State persists on workers—router rebuilds state by querying workers on startup.
+> - **No KV events** (`--no-kv-events`): State persistence is not supported.
 >
 > When `--kv-overlap-score-weight` is set to 0 or `--no-kv-events` is set, no KvIndexer will be launched to drain and process KV events. It's recommended to disable your backend workers from relaying events through `KvEventPublisher` to avoid event accumulation in JetStream. WIP to enable disabling publishing of KV events completely in these cases.
 >
@@ -110,9 +113,16 @@ await register_llm(
 
 The KV-aware router operates on two key principles to optimize request routing:
 
-### Global KV Cache State via JetStream
+### Global KV Cache State Synchronization
 
-First, KV events from engines are sent to a persistent NATS JetStream. Each KV router/indexer replica acts as a durable consumer, pulling messages from this shared stream to maintain a global view of cached blocks across all engines. This architecture ensures consistency across router replicas and persistence across restarts.
+KV events from engines are collected by the router to maintain a global view of cached blocks across all workers. The router supports two event transport modes:
+
+#### Mode 1: JetStream (Default)
+
+KV events are sent to a persistent NATS JetStream. Each KV router/indexer replica acts as a durable consumer, pulling messages from this shared stream. This architecture ensures consistency across router replicas and persistence across restarts.
+
+- **Best for**: Production deployments requiring durability and multi-replica router consistency
+- **Tradeoffs**: Requires JetStream setup; slightly higher latency due to persistence guarantees
 
 ```mermaid
 graph TD
@@ -152,6 +162,60 @@ graph TD
     style R2 fill:#f3e5f5,color:#5a850f
 ```
 
+#### Mode 2: NATS Core with Local Indexer
+
+When workers are started with `--enable-local-indexer`, each worker maintains its own local radix tree (local indexer) and publishes events over NATS Core (fire-and-forget pub/sub) instead of JetStream. Each worker assigns monotonically increasing event IDs to its events. The router detects gaps in event sequences and recovers missed events by querying the worker's local indexer directly.
+
+- **Best for**: Lower-latency setups; simpler deployments without JetStream; single-router scenarios
+- **Tradeoffs**: State persists on workers (not centralized); recovery depends on workers being available
+- **Enable with**: `--enable-local-indexer` flag on workers (vLLM, mocker)
+
+```mermaid
+graph TD
+    subgraph Engines
+        E1[Engine 1<br/>LocalKvIndexer]
+        E2[Engine 2<br/>LocalKvIndexer]
+        E3[Engine 3<br/>LocalKvIndexer]
+    end
+
+    subgraph "NATS Core"
+        NC[KV Events Pub/Sub<br/>- Block created<br/>- Block removed]
+    end
+
+    subgraph "Router Replicas"
+        R1[Router 1<br/>KVIndexer]
+        R2[Router 2<br/>KVIndexer]
+    end
+
+    E1 -->|Publish Events| NC
+    E2 -->|Publish Events| NC
+    E3 -->|Publish Events| NC
+
+    NC -->|Subscribe| R1
+    NC -->|Subscribe| R2
+
+    style NC fill:#e1f5fe,color:#5a850f
+    style E1 fill:#fff3e0,color:#5a850f
+    style E2 fill:#fff3e0,color:#5a850f
+    style E3 fill:#fff3e0,color:#5a850f
+    style R1 fill:#f3e5f5,color:#5a850f
+    style R2 fill:#f3e5f5,color:#5a850f
+```
+
+**How gap detection works:**
+1. Each worker assigns monotonically increasing event IDs starting from 0
+2. The router tracks the last received event ID per worker
+3. If an event arrives with `event_id > last_id + 1`, the router detects a gap
+4. The router queries the worker's local indexer for the missing event range `[last_id+1, event_id-1]`
+5. On worker discovery (Added event), the router dumps the worker's entire local indexer state
+
+**Startup behavior:**
+- When a worker is discovered, the router queries and ingests its full local indexer state
+- When a worker is removed, the router removes all its blocks from the global radix tree
+
+>[!Note]
+> The router automatically selects the transport mode based on worker configuration. If all connected workers have `enable_local_indexer=true`, the router uses NATS Core mode. Otherwise, it uses JetStream mode.
+
 ### Local Active Block Management with Replica Sync
 
 Second, in addition to cached blocks, each router replica needs to track active blocks (blocks being used for ongoing generation) as load metrics. Since this information is highly time-sensitive, it should be predicted immediately when:
@@ -249,19 +313,26 @@ Without this flag, each replica maintains its own isolated view of active blocks
 
 ### Persistence and Recovery
 
-**Prefix blocks persist by default:**
-- Stored in NATS JetStream with 1-hour retention
+Persistence behavior depends on which event transport mode is active:
+
+**JetStream Mode (default):**
+- Prefix blocks are stored in NATS JetStream with 1-hour retention
 - Snapshots saved to NATS object store at configurable thresholds
 - New replicas automatically restore this state on startup
-
-You can a launch a third Router replica even if the first two Router replicas are down, and it will recover the full prefix state. (As mentioned above, the tracking of active blocks will not persist, but will become eventually consistent through request handling.)
+- You can launch a third Router replica even if the first two are down, and it will recover the full prefix state
 
 ```bash
 python -m dynamo.frontend --router-mode kv --port 8002 --router-replica-sync
 ```
 
+**NATS Core with Local Indexer Mode:**
+- State persists on workers—events are fire-and-forget but workers retain their local indexer state
+- On startup, the router queries each worker's local indexer to rebuild state
+- Recovery depends on workers being available; if a worker is down, its blocks cannot be recovered
+- Simpler infrastructure (no JetStream required) but less resilient
+
 >[!Note]
-> If you need to start with a fresh state, you have two options:
+> If you need to start with a fresh state in JetStream mode, you have two options:
 > 1. **Recommended**: Use a different namespace/component (see [Distributed Runtime](/docs/design_docs/distributed_runtime.md)) which will start a new stream and NATS object store path
 > 2. **Use with caution**: Launch a router with the `--router-reset-states` flag, which will purge the entire stream and radix snapshot. This should only be done when launching the first router replica in a component, as it can bring existing router replicas into an inconsistent state.
 
@@ -375,13 +446,6 @@ In distributed deployments with multiple routers, each router maintains visibili
 
 Each event carries a unique router ID to prevent self-event processing. This asynchronous communication system ensures optimal routing decisions by maintaining consistent KV cache state across all routers, even as they handle different request streams.
 
-### Event Persistence and Recovery
-
-KV cache events are persisted in NATS JetStream, allowing router replicas to maintain their global view of KV blocks across restarts. By default, routers persist their state - they download any available snapshot from NATS object store and continue consuming events from their last acknowledged position in the stream. This default behavior ensures KV cache awareness is maintained across router restarts without any additional configuration.
-
-To manage stream growth, when the message count exceeds `--router-snapshot-threshold`, a router acquires an etcd-based distributed lock, purges acknowledged messages from the stream, and uploads the current radix tree state to NATS object store. This snapshot serves as a checkpoint for faster initialization of future router instances.
-
-
 ## Using KvPushRouter Python API
 
 Instead of launching the KV Router via command line, you can create a `KvPushRouter` object directly in Python. This allows per-request routing configuration overrides.

lib/llm/src/kv_router.rs

@@ -53,7 +53,7 @@ use crate::{
         },
         scheduler::{KvScheduler, KvSchedulerError, PotentialLoad, SchedulingRequest},
         sequence::SequenceError,
-        subscriber::{recover_from_all_workers, start_kv_router_background},
+        subscriber::{start_kv_router_background, start_kv_router_background_nats_core},
     },
     local_model::runtime_config::ModelRuntimeConfig,
     model_card::ModelDeploymentCard,
@@ -69,7 +69,7 @@ use crate::{
 pub const KV_METRICS_ENDPOINT: &str = "load_metrics";
 
 // for metric publishing (push-based)
-pub const KV_EVENT_SUBJECT: &str = "kv_events";
+pub const KV_EVENT_SUBJECT: &str = "kv-events";
 pub const KV_HIT_RATE_SUBJECT: &str = "kv-hit-rate";
 pub const KV_METRICS_SUBJECT: &str = "kv_metrics";
 
@@ -357,66 +357,90 @@ impl KvRouter {
         tracing::info!("Worker query client initialized");
 
         // Start KV event subscriber background process (only when use_kv_events is enabled)
+        // This is spawned as a background task to avoid blocking router startup.
+        // The task waits for runtime_configs to determine whether to use NATS Core or JetStream.
         if kv_router_config.use_kv_events
             && let Indexer::KvIndexer(ref kv_indexer) = indexer
         {
-            start_kv_router_background(
-                component.clone(),
-                consumer_id,
-                kv_indexer.event_sender(),
-                kv_indexer.remove_worker_sender(),
-                kv_router_config
-                    .router_snapshot_threshold
-                    .map(|_| kv_indexer.get_workers_sender()),
-                kv_router_config
-                    .router_snapshot_threshold
-                    .map(|_| kv_indexer.snapshot_event_sender()),
-                cancellation_token.clone(),
-                kv_router_config.router_snapshot_threshold,
-                kv_router_config.router_reset_states,
-            )
-            .await?;
+            // Clone everything needed for the background task
+            let component_clone = component.clone();
+            let kv_indexer_clone = kv_indexer.clone();
+            let cancellation_token_clone = cancellation_token.clone();
+            let mut runtime_configs_rx_clone = runtime_configs_rx.clone();
+            let worker_query_client_clone =
+                worker_query::WorkerQueryClient::new(component.clone(), runtime_configs_rx.clone());
 
-            // Perform startup recovery from workers with local indexers
-            // This catches up on any events missed while the router was offline
-            let last_event_ids = kv_indexer
-                .get_last_received_event_ids()
-                .await
-                .unwrap_or_default();
-            let instances = client.instance_source.as_ref().borrow().clone();
-            let worker_ids: Vec<WorkerId> = instances.iter().map(|i| i.instance_id).collect();
+            tokio::spawn(async move {
+                // Wait for runtime_configs to have at least one entry
+                let (all_local_indexer, count) = loop {
+                    {
+                        let configs = runtime_configs_rx_clone.borrow();
+                        if !configs.is_empty() {
+                            let all_local_indexer =
+                                configs.values().all(|c| c.enable_local_indexer);
+                            break (all_local_indexer, configs.len());
+                        }
+                    }
+
+                    // Wait for changes to runtime_configs
+                    tokio::select! {
+                        _ = cancellation_token_clone.cancelled() => {
+                            tracing::debug!("Subscriber selection task cancelled");
+                            return;
+                        }
+                        result = runtime_configs_rx_clone.changed() => {
+                            if result.is_err() {
+                                tracing::debug!("Runtime configs channel closed");
+                                return;
+                            }
+                        }
+                    }
+                };
 
-            if !worker_ids.is_empty() {
+                if all_local_indexer {
+                    // All workers have local_indexer enabled - use NATS Core
                     tracing::info!(
-                    worker_count = worker_ids.len(),
-                    "Starting recovery from workers with local indexers"
+                        "All {count} workers have local_indexer enabled, using NATS Core subscription"
                     );
 
-                // NOTE: recover_from_all_workers() is a no-op if
-                // Worker with worker_id is not associated with a
-                // local indexer instance.
-                let recovered = recover_from_all_workers(
-                    &worker_query_client,
-                    &last_event_ids,
-                    &worker_ids,
-                    &kv_indexer.event_sender(),
+                    if let Err(e) = start_kv_router_background_nats_core(
+                        component_clone.clone(),
+                        kv_indexer_clone.event_sender(),
+                        kv_indexer_clone.remove_worker_sender(),
+                        cancellation_token_clone.clone(),
+                        worker_query_client_clone,
                     )
-                .await;
-
-                if recovered > 0 {
-                    tracing::info!(
-                        recovered_events = recovered,
-                        "KV Router startup: Recovered {} KV events from workers {:?}",
-                        recovered,
-                        worker_ids
-                    );
+                    .await
+                    {
+                        tracing::error!("Failed to start NATS Core subscriber: {e}");
+                    }
                 } else {
+                    // Not all workers have local_indexer - use JetStream
                     tracing::info!(
-                        "KV Router startup: No KV events recovered from workers {:?}",
-                        worker_ids
+                        "Not all workers have local_indexer enabled, using JetStream subscription"
                     );
+
+                    if let Err(e) = start_kv_router_background(
+                        component_clone.clone(),
+                        consumer_id,
+                        kv_indexer_clone.event_sender(),
+                        kv_indexer_clone.remove_worker_sender(),
+                        kv_router_config
+                            .router_snapshot_threshold
+                            .map(|_| kv_indexer_clone.get_workers_sender()),
+                        kv_router_config
+                            .router_snapshot_threshold
+                            .map(|_| kv_indexer_clone.snapshot_event_sender()),
+                        cancellation_token_clone.clone(),
+                        kv_router_config.router_snapshot_threshold,
+                        kv_router_config.router_reset_states,
+                    )
+                    .await
+                    {
+                        tracing::error!("Failed to start JetStream subscriber: {e}");
                     }
                 }
+            });
         }
 
         tracing::info!("KV Routing initialized");

lib/llm/src/kv_router/publisher.rs

@@ -21,7 +21,7 @@ use dynamo_runtime::traits::{
 };
 use dynamo_runtime::{
     component::{Component, Namespace},
-    transports::nats::{NatsQueue, QUEUE_NAME, Slug},
+    transports::nats::{NatsQueue, Slug},
 };
 use futures::StreamExt;
 
@@ -29,8 +29,8 @@ use crate::kv_router::{
     KV_EVENT_SUBJECT, KV_METRICS_SUBJECT, WORKER_KV_INDEXER_BUFFER_SIZE,
     WORKER_KV_INDEXER_QUERY_SUBJECT,
     indexer::{
-        KvIndexerInterface, KvIndexerMetrics, LocalKvIndexer, RouterEvent, WorkerKvQueryRequest,
-        WorkerKvQueryResponse, compute_block_hash_for_seq,
+        KvIndexerMetrics, LocalKvIndexer, RouterEvent, WorkerKvQueryRequest,
+        compute_block_hash_for_seq,
     },
     protocols::*,
 };
@@ -179,24 +179,43 @@ impl KvEventPublisher {
                 ))
         });
 
-        let stream_name = Slug::slugify(&format!("{}.{}", component.subject(), KV_EVENT_SUBJECT))
+        // Connect the NatsQueue before passing it to the event processor
+        let cancellation_token_clone = cancellation_token.clone();
+        let local_indexer_clone = local_indexer.clone();
+
+        if enable_local_indexer {
+            // When local indexer is enabled, use NATS Core (Component) for publishing.
+            // This is simpler and doesn't require JetStream durability since recovery
+            // is handled via the local indexer's event buffer.
+            tracing::info!("Using NATS Core for KV event publishing (local_indexer mode)");
+            let component_clone = component.clone();
+            component.drt().runtime().secondary().spawn(async move {
+                start_event_processor(
+                    component_clone,
+                    worker_id,
+                    cancellation_token_clone,
+                    rx,
+                    local_indexer_clone,
+                )
+                .await
+            });
+        } else {
+            // When local indexer is disabled, use JetStream (NatsQueue) for durability.
+            let stream_name =
+                Slug::slugify(&format!("{}.{}", component.subject(), KV_EVENT_SUBJECT))
                     .to_string()
                     .replace("_", "-");
             let nats_server = std::env::var(env_nats::NATS_SERVER)
                 .unwrap_or_else(|_| "nats://localhost:4222".to_string());
-        // Create NatsQueue without consumer since we're only publishing
             let mut nats_queue = NatsQueue::new_without_consumer(
                 stream_name,
                 nats_server,
                 std::time::Duration::from_secs(60), // 1 minute timeout
             );
 
-        // Connect the NatsQueue before passing it to the event processor
-        let cancellation_token_clone = cancellation_token.clone();
-        let local_indexer_clone = local_indexer.clone();
             component.drt().runtime().secondary().spawn(async move {
                 if let Err(e) = nats_queue.connect().await {
-                tracing::error!("Failed to connect NatsQueue: {}", e);
+                    tracing::error!("Failed to connect NatsQueue: {e}");
                     return;
                 }
                 start_event_processor(
@@ -208,6 +227,7 @@ impl KvEventPublisher {
                 )
                 .await
             });
+        }
 
         Ok(Self {
             kv_block_size,
@@ -278,7 +298,9 @@ async fn start_event_processor<P: EventPublisher + Send + Sync + 'static>(
                 }
 
                 // Then publish to NATS for global distribution
-                if let Err(e) = publisher.publish(QUEUE_NAME, &router_event).await {
+                // Use KV_EVENT_SUBJECT so both JetStream and NATS Core subscribers
+                // can receive events on the expected subject.
+                if let Err(e) = publisher.publish(KV_EVENT_SUBJECT, &router_event).await {
                     tracing::error!("Failed to publish event to NATS: {}", e);
                 }
 
@@ -332,31 +354,12 @@ async fn start_worker_kv_query_service(
                     }
                 };
 
-                // TODO extract request event id range. For now, just debug print
                 tracing::debug!("Received WorkerKvQueryRequest: {:?}", request);
 
-                // Resolve which events to return based on optional start/end ids
-                let events = match (request.start_event_id, request.end_event_id) {
-                    (None, None) => {
-                        match local_indexer.dump_events().await {
-                            Ok(events) => events,
-                            Err(err) => {
-                                tracing::error!(
-                                    error = %err,
-                                    worker_id,
-                                    "Failed to dump events for WorkerKvQueryRequest; returning buffered events instead"
-                                );
-                                local_indexer.get_all_events_in_buffer()
-                            }
-                        }
-                    }
-                    _ => {
-                            local_indexer.get_events_in_id_range(request.start_event_id, request.end_event_id).await
-                    }
-                };
-
-                // Build WorkerKvQueryResponse
-                let response = WorkerKvQueryResponse { events };
+                // Query events based on optional start/end ids
+                let response = local_indexer
+                    .get_events_in_id_range(request.start_event_id, request.end_event_id)
+                    .await;
 
                 // Send reply back (if reply subject exists)
                 if let Some(reply_subject) = msg.reply {
@@ -1281,7 +1284,7 @@ mod tests_startup_helpers {
         let published = published.lock().unwrap();
         assert_eq!(published.len(), 1);
         let (subject, _) = &published[0];
-        assert_eq!(subject, QUEUE_NAME);
+        assert_eq!(subject, KV_EVENT_SUBJECT);
     }
 
     //--------------------------------------------------------------------
@@ -1339,7 +1342,7 @@ mod tests_startup_helpers {
             let published_events = published.lock().unwrap();
             assert_eq!(published_events.len(), 1);
             let (subject, _) = &published_events[0];
-            assert_eq!(subject, QUEUE_NAME);
+            assert_eq!(subject, KV_EVENT_SUBJECT);
         } // drop lock
 
         // Verify event was applied to local indexer
@@ -1726,7 +1729,7 @@ mod tests_startup_helpers {
             assert_eq!(published.len(), 1, "Worker should have published 1 event");
             (published[0].0.clone(), published[0].1.clone())
         }; // drop worker_published before await
-        assert_eq!(subject, QUEUE_NAME);
+        assert_eq!(subject, KV_EVENT_SUBJECT);
 
         let router_event: RouterEvent = rmp_serde::from_slice(&bytes).unwrap();
         router_indexer
@@ -1828,9 +1831,14 @@ mod tests_startup_helpers {
         );
 
         // Step 4b: Query worker's local indexer for events after last_known_id
-        let missed_events = local_indexer_1
+        let response = local_indexer_1
             .get_events_in_id_range(Some(last_known_id + 1), None)
             .await;
+        let missed_events = match response {
+            crate::kv_router::indexer::WorkerKvQueryResponse::Events(e) => e,
+            crate::kv_router::indexer::WorkerKvQueryResponse::TreeDump(e) => e,
+            other => panic!("Unexpected response: {:?}", other),
+        };
         assert_eq!(
             missed_events.len(),
             1,
@@ -2293,11 +2301,16 @@ mod test_integration_publisher_with_kvindexer {
             let response = kv_router
                 .query_worker_local_kv(worker_id, None, None)
                 .await?;
-            if response.events.is_empty() {
+            let events = match response {
+                crate::kv_router::indexer::WorkerKvQueryResponse::Events(e) => e,
+                crate::kv_router::indexer::WorkerKvQueryResponse::TreeDump(e) => e,
+                _ => vec![],
+            };
+            if events.is_empty() {
                 continue;
             }
 
-            let event_count = response.events.len();
+            let event_count = events.len();
             tracing::info!(
                 worker_id,
                 events = event_count,
@@ -2323,8 +2336,13 @@ mod test_integration_publisher_with_kvindexer {
             let response = kv_router
                 .query_worker_local_kv(worker_id, None, None)
                 .await?;
+            let events = match response {
+                crate::kv_router::indexer::WorkerKvQueryResponse::Events(e) => e,
+                crate::kv_router::indexer::WorkerKvQueryResponse::TreeDump(e) => e,
+                _ => vec![],
+            };
             assert!(
-                response.events.is_empty(),
+                events.is_empty(),
                 "Worker {worker_id} should not report buffered KV events; best worker {best_worker_id} reported {best_worker_event_count}"
             );
         }
@@ -2437,6 +2455,24 @@ mod test_integration_publisher_with_kvindexer {
             .endpoint("generate");
         let pre_client = pre_backend_endpoint.client().await?;
 
+        // Wait for the client to discover both workers
+        let discovery_timeout = Duration::from_secs(5);
+        let discovery_start = std::time::Instant::now();
+        loop {
+            let instances = pre_client.instance_source.as_ref().borrow().clone();
+            if instances.len() >= 2 {
+                tracing::info!("Discovered {} workers", instances.len());
+                break;
+            }
+            if discovery_start.elapsed() > discovery_timeout {
+                anyhow::bail!(
+                    "Timed out waiting for worker discovery: expected 2, found {}",
+                    instances.len()
+                );
+            }
+            tokio::time::sleep(Duration::from_millis(50)).await;
+        }
+
         // Create a PushRouter to send requests directly to a specific worker
         let pre_push_router =
             PushRouter::<PreprocessedRequest, Annotated<LLMEngineOutput>>::from_client_with_threshold(
@@ -2495,13 +2531,32 @@ mod test_integration_publisher_with_kvindexer {
             .await?,
         );
 
-        // At this point kvrouter's indexer should already have the
-        // events stored in the workers, due to the catch-up built into KvRouter::new.
+        // The KvRouter now starts its subscriber asynchronously in a background task
+        // that waits for runtime_configs. Poll until events appear or timeout.
         // Each request generates 2 events: input block (parent_hash: None) + output block (parent_hash: Some)
         // With 2 workers, that's 4 events total.
-        let global_kv_events = kv_router.indexer.dump_events().await?;
-        tracing::debug!("Global KV events: {:?}", global_kv_events);
-        assert_eq!(global_kv_events.len(), 4); // 2 workers × 2 events per request (input + output)
+        let expected_events = 4;
+        let max_wait = Duration::from_secs(10);
+        let poll_interval = Duration::from_millis(100);
+        let start = std::time::Instant::now();
+
+        let global_kv_events = loop {
+            let events = kv_router.indexer.dump_events().await?;
+            tracing::debug!("Global KV events ({}): {:?}", events.len(), events);
+            if events.len() >= expected_events {
+                break events;
+            }
+            if start.elapsed() > max_wait {
+                anyhow::bail!(
+                    "Timed out waiting for KV events: expected {}, got {}",
+                    expected_events,
+                    events.len()
+                );
+            }
+            tokio::time::sleep(poll_interval).await;
+        };
+
+        assert_eq!(global_kv_events.len(), expected_events); // 2 workers × 2 events per request (input + output)
 
         // === Cleanup ===
         for handle in server_handles {

lib/llm/src/kv_router/subscriber.rs

@@ -9,7 +9,7 @@ use dynamo_runtime::{
     config::environment_names::nats as env_nats,
     discovery::{DiscoveryEvent, DiscoveryQuery},
     prelude::*,
-    traits::events::EventPublisher,
+    traits::events::{EventPublisher, EventSubscriber},
     transports::nats::{NatsQueue, Slug},
 };
 use futures::StreamExt;
@@ -19,7 +19,7 @@ use tokio_util::sync::CancellationToken;
 
 use crate::kv_router::{
     KV_EVENT_SUBJECT, RADIX_STATE_BUCKET, RADIX_STATE_FILE,
-    indexer::{DumpRequest, GetWorkersRequest, RouterEvent},
+    indexer::{DumpRequest, GetWorkersRequest, RouterEvent, WorkerKvQueryResponse},
     protocols::WorkerId,
     router_discovery_query,
     worker_query::WorkerQueryClient,
@@ -140,8 +140,8 @@ pub async fn recover_from_worker(
         );
     } else {
         tracing::warn!(
-            "Worker {} does not have local indexer enabled, skipping recovery",
-            worker_id
+            worker_id,
+            "Worker does not have local indexer enabled, skipping recovery"
         );
         return Ok(0);
     }
@@ -151,13 +151,46 @@ pub async fn recover_from_worker(
         .query_worker(worker_id, start_event_id, end_event_id)
         .await?;
 
-    let events_count = response.events.len();
+    // Handle response variants
+    let events = match response {
+        WorkerKvQueryResponse::Events(events) => {
+            tracing::debug!(worker_id, count = events.len(), "Got buffered events");
+            events
+        }
+        WorkerKvQueryResponse::TreeDump(events) => {
+            tracing::info!(
+                worker_id,
+                count = events.len(),
+                "Got tree dump (range too old or unspecified)"
+            );
+            events
+        }
+        WorkerKvQueryResponse::TooNew {
+            requested_start,
+            requested_end,
+            newest_available,
+        } => {
+            tracing::warn!(
+                worker_id,
+                ?requested_start,
+                ?requested_end,
+                newest_available,
+                "Requested range is newer than available data"
+            );
+            return Ok(0);
+        }
+        WorkerKvQueryResponse::InvalidRange { start_id, end_id } => {
+            anyhow::bail!("Invalid range: end_id ({end_id}) < start_id ({start_id})");
+        }
+    };
+
+    let events_count = events.len();
 
     if events_count == 0 {
         tracing::debug!(
             worker_id,
             start_event_id = ?start_event_id,
-            "No missed events to recover from worker"
+            "No events to recover from worker"
         );
         return Ok(0);
     }
@@ -166,19 +199,14 @@ pub async fn recover_from_worker(
         worker_id,
         start_event_id = ?start_event_id,
         events_count,
-        "Recovered {} missed events from worker",
-        events_count
+        "Recovered {events_count} events from worker"
     );
 
     // Apply recovered events to the indexer
-    for event in response.events {
+    for event in events {
         if let Err(e) = event_tx.send(event).await {
-            tracing::error!(
-                worker_id,
-                error = %e,
-                "Failed to send recovered event to indexer"
-            );
-            anyhow::bail!("Failed to send recovered event: {}", e);
+            tracing::error!(worker_id, error = %e, "Failed to send recovered event to indexer");
+            anyhow::bail!("Failed to send recovered event: {e}");
         }
     }
 
@@ -606,6 +634,187 @@ pub async fn start_kv_router_background(
     Ok(())
 }
 
+/// Start a simplified background task for event consumption using NATS Core.
+///
+/// This is used when local indexer mode is enabled. Unlike `start_kv_router_background`,
+/// this function:
+/// - Uses NATS Core pub/sub instead of JetStream
+/// - Does not support snapshots, purging, or durable consumers
+/// - On worker Added: dumps worker's local indexer into router
+/// - On worker Removed: removes worker from router indexer
+///
+/// This is appropriate when workers have local indexers enabled.
+pub async fn start_kv_router_background_nats_core(
+    component: Component,
+    kv_events_tx: mpsc::Sender<RouterEvent>,
+    remove_worker_tx: mpsc::Sender<WorkerId>,
+    cancellation_token: CancellationToken,
+    worker_query_client: WorkerQueryClient,
+) -> Result<()> {
+    // Subscribe to KV events using NATS Core
+    let mut subscriber = component.subscribe(KV_EVENT_SUBJECT).await?;
+
+    tracing::info!(
+        "KV Router using NATS Core subscription on subject: {}.{} (local_indexer mode)",
+        component.subject(),
+        KV_EVENT_SUBJECT
+    );
+
+    // Get the generate endpoint and watch for instance events (add/remove)
+    let discovery_client = component.drt().discovery();
+    let generate_discovery_key = DiscoveryQuery::Endpoint {
+        namespace: component.namespace().name().to_string(),
+        component: component.name().to_string(),
+        endpoint: "generate".to_string(),
+    };
+    let mut instance_event_stream = discovery_client
+        .list_and_watch(generate_discovery_key, Some(cancellation_token.clone()))
+        .await?;
+
+    tokio::spawn(async move {
+        // Track last received event ID per worker for gap detection
+        let mut last_event_ids: HashMap<WorkerId, u64> = HashMap::new();
+
+        loop {
+            tokio::select! {
+                biased;
+
+                _ = cancellation_token.cancelled() => {
+                    tracing::debug!("KV Router NATS Core background task received cancellation signal");
+                    break;
+                }
+
+                // Handle generate endpoint instance add/remove events
+                Some(discovery_event_result) = instance_event_stream.next() => {
+                    let Ok(discovery_event) = discovery_event_result else {
+                        continue;
+                    };
+
+                    match discovery_event {
+                        DiscoveryEvent::Added(_instance) => {
+                            // Extract worker_id from the instance
+                            let worker_id = _instance.instance_id();
+
+                            tracing::info!(
+                                worker_id = worker_id,
+                                "DISCOVERY: Worker added, dumping local indexer into router"
+                            );
+
+                            // Query worker's local indexer and dump all events
+                            match recover_from_worker(
+                                &worker_query_client,
+                                worker_id,
+                                None, // Start from beginning
+                                None, // Get all events
+                                &kv_events_tx,
+                            )
+                            .await
+                            {
+                                Ok(count) => {
+                                    tracing::info!(
+                                        worker_id = worker_id,
+                                        events_recovered = count,
+                                        "Successfully dumped worker's local indexer"
+                                    );
+                                }
+                                Err(e) => {
+                                    tracing::warn!(
+                                        worker_id = worker_id,
+                                        error = %e,
+                                        "Failed to dump worker's local indexer (may not have local indexer enabled)"
+                                    );
+                                }
+                            }
+                        }
+                        DiscoveryEvent::Removed(worker_id) => {
+                            tracing::warn!(
+                                worker_id = worker_id,
+                                "DISCOVERY: Worker removed, removing from router indexer"
+                            );
+
+                            if let Err(e) = remove_worker_tx.send(worker_id).await {
+                                tracing::warn!("Failed to send worker removal for worker {worker_id}: {e}");
+                            }
+                        }
+                    }
+                }
+
+                // Handle event consumption from NATS Core subscription
+                Some(msg) = subscriber.next() => {
+                    let event: RouterEvent = match serde_json::from_slice(&msg.payload) {
+                        Ok(event) => event,
+                        Err(e) => {
+                            tracing::warn!("Failed to deserialize RouterEvent from NATS Core: {e:?}");
+                            continue;
+                        }
+                    };
+
+                    let worker_id = event.worker_id;
+                    let event_id = event.event.event_id;
+
+                    // Gap detection: check if event ID is monotonically increasing per worker
+                    // Note: event_id <= last_id is duplicate/out-of-order, apply anyway (idempotent)
+                    if let Some(&last_id) = last_event_ids.get(&worker_id)
+                        && event_id > last_id + 1
+                    {
+                        // Gap detected - recover missing events before processing current
+                        let gap_start = last_id + 1;
+                        let gap_end = event_id - 1;
+                        tracing::warn!(
+                            worker_id,
+                            gap_start,
+                            gap_end,
+                            gap_size = gap_end - gap_start + 1,
+                            "Event ID gap detected, recovering events [{gap_start}, {gap_end}]"
+                        );
+
+                        // Note: While recovering, new events may queue in the NATS subscriber's
+                        // internal buffer. We don't explicitly buffer them here for simplicity.
+                        // The subscriber will process them in order after recovery completes.
+                        if let Err(e) = recover_from_worker(
+                            &worker_query_client,
+                            worker_id,
+                            Some(gap_start),
+                            Some(gap_end),
+                            &kv_events_tx,
+                        ).await {
+                            tracing::error!(
+                                worker_id,
+                                gap_start,
+                                gap_end,
+                                error = %e,
+                                "Failed to recover gap events; proceeding with current event anyway"
+                            );
+                            // Note: If recovery fails, we still apply the current event.
+                            // The tree will have a gap, but it's better than dropping the event.
+                        }
+                    }
+                    // First event from this worker is always valid - we accept whatever ID it has.
+                    // This handles initial startup and worker restarts without requiring event 0.
+
+                    // Update last seen event ID (use max to handle out-of-order)
+                    last_event_ids
+                        .entry(worker_id)
+                        .and_modify(|id| *id = (*id).max(event_id))
+                        .or_insert(event_id);
+
+                    // Forward the RouterEvent to the indexer
+                    if let Err(e) = kv_events_tx.send(event).await {
+                        tracing::warn!(
+                            "failed to send kv event to indexer; shutting down: {e:?}"
+                        );
+                        break;
+                    }
+                }
+            }
+        }
+
+        tracing::debug!("KV Router NATS Core background task exiting");
+    });
+
+    Ok(())
+}
+
 /// Cleanup orphaned NATS consumers that no longer have corresponding router entries
 async fn cleanup_orphaned_consumers(
     nats_queue: &mut NatsQueue,

lib/llm/src/kv_router/worker_query.rs

@@ -57,19 +57,16 @@ impl WorkerQueryClient {
         // Check if worker has local indexer enabled
         if !self.has_local_indexer(worker_id) {
             anyhow::bail!(
-                "Worker {} does not have local indexer enabled (enable_local_indexer=false or not set in MDC user_data)",
-                worker_id
+                "Worker {worker_id} does not have local indexer enabled (enable_local_indexer=false or not set in MDC user_data)"
             );
         }
 
         // Match worker's subscribe format
-        let subject_str = format!("{}.{}", WORKER_KV_INDEXER_QUERY_SUBJECT, worker_id); // see publisher.rs/start_worker_kv_query_service()
-        let subject = format!("{}.{}", self.component.subject(), subject_str);
+        let subject_str = format!("{}.{worker_id}", WORKER_KV_INDEXER_QUERY_SUBJECT); // see publisher.rs/start_worker_kv_query_service()
+        let subject = format!("{}.{subject_str}", self.component.subject());
 
         tracing::debug!(
-            "Router sending query request to worker {} on NATS subject: {}",
-            worker_id,
-            subject
+            "Router sending query request to worker {worker_id} on NATS subject: {subject}"
         );
 
         // Create and serialize request
@@ -89,10 +86,7 @@ impl WorkerQueryClient {
             .kv_router_nats_request(subject.clone(), request_bytes.into(), timeout)
             .await
             .with_context(|| {
-                format!(
-                    "Failed to send request to worker {} on subject {}",
-                    worker_id, subject
-                )
+                format!("Failed to send request to worker {worker_id} on subject {subject}")
             })?;
 
         // Deserialize response

tests/router/test_router_e2e_with_mockers.py

@@ -173,6 +173,8 @@ def _build_mocker_command(
         command.extend(["--watermark", str(mocker_args["watermark"])])
     if "dp_size" in mocker_args:
         command.extend(["--data-parallel-size", str(mocker_args["dp_size"])])
+    if mocker_args.get("enable_local_indexer"):
+        command.append("--enable-local-indexer")
 
     return command
 
@@ -559,22 +561,32 @@ def test_query_instance_id_returns_worker_and_tokens(
 
 
 @pytest.mark.parallel
-def test_router_decisions(request, runtime_services_session, predownload_tokenizers):
-    """Validate KV cache prefix reuse and dp_rank routing by sending progressive requests with overlapping prefixes."""
+@pytest.mark.parametrize("use_nats_core", [False, True], ids=["jetstream", "nats_core"])
+def test_router_decisions(
+    request, runtime_services_session, predownload_tokenizers, use_nats_core
+):
+    """Validate KV cache prefix reuse and dp_rank routing by sending progressive requests with overlapping prefixes.
+
+    Parameterized to test both JetStream (default) and NATS Core (local indexer) modes.
+    """
 
     # runtime_services starts etcd and nats
-    logger.info("Starting test router prefix reuse and KV events synchronization")
+    mode = "NATS Core (local indexer)" if use_nats_core else "JetStream"
+    logger.info(
+        f"Starting test router prefix reuse and KV events synchronization ({mode})"
+    )
 
     # Create mocker args dictionary with dp_size=4
     mocker_args = {
         "speedup_ratio": SPEEDUP_RATIO,
         "block_size": BLOCK_SIZE,
         "dp_size": 4,
+        "enable_local_indexer": use_nats_core,
     }
 
     try:
         logger.info(
-            "Starting 2 mocker instances with dp_size=4 each (8 total dp ranks)"
+            f"Starting 2 mocker instances with dp_size=4 each (8 total dp ranks), {mode}"
         )
         mockers = MockerProcess(request, mocker_args=mocker_args, num_mockers=2)
         logger.info(f"All mockers using endpoint: {mockers.endpoint}")