feat: follow up of optinal, batching-of-client-side-events (#6741)

Signed-off-by: michaelfeil <63565275+michaelfeil@users.noreply.github.com>
Signed-off-by: Michael Feil <michaelfeil@users.noreply.github.com>
Signed-off-by: Michael Feil <63565275+michaelfeil@users.noreply.github.com>

lib/bindings/python/rust/llm/kv.rs

@@ -124,12 +124,14 @@ impl KvEventPublisher {
     ///         so that routers can recover events directly from this worker.
     ///     zmq_endpoint: Optional ZMQ SUB endpoint to read raw engine events from.
     ///     zmq_topic: ZMQ topic filter (default "").
-    ///     batching_timeout_us: Maximum time (in **microseconds**) to accumulate
+    ///     batching_timeout_ms: Maximum time (in **milliseconds**) to accumulate
     ///         events into a single batch before flushing.
-    ///         ``None`` uses the default window of 10000 µs (10 ms).
-    ///         ``0`` disables batching: every event is published immediately.
+    ///         ``None`` disables batching: every event is published immediately.
+    ///         ``50`` to enable batching with a 50 ms window.
+    ///         ``0`` is treated as ``None`` (also disables batching).
+    ///         Maximum allowed is 15_000 (15 seconds); larger values are capped.
     #[new]
-    #[pyo3(signature = (endpoint, worker_id=0, kv_block_size=0, dp_rank=0, enable_local_indexer=false, zmq_endpoint=None, zmq_topic=None, batching_timeout_us=None))]
+    #[pyo3(signature = (endpoint, worker_id=0, kv_block_size=0, dp_rank=0, enable_local_indexer=false, zmq_endpoint=None, zmq_topic=None, batching_timeout_ms=llm_rs::kv_router::publisher::DEFAULT_BATCHING_TIMEOUT_MS))]
     #[allow(clippy::too_many_arguments)]
     fn new(
         endpoint: Endpoint,
@@ -139,7 +141,7 @@ impl KvEventPublisher {
         enable_local_indexer: bool,
         zmq_endpoint: Option<String>,
         zmq_topic: Option<String>,
-        batching_timeout_us: Option<u64>,
+        batching_timeout_ms: Option<u64>,
     ) -> PyResult<Self> {
         let _ = worker_id;
 
@@ -161,7 +163,7 @@ impl KvEventPublisher {
             source_config,
             enable_local_indexer,
             dp_rank,
-            batching_timeout_us,
+            batching_timeout_ms,
         )
         .map_err(to_pyerr)?;
 

lib/runtime/src/metrics/prometheus_names.rs

@@ -519,12 +519,18 @@ pub mod tokio_perf {
     pub const ALIVE_TASKS: &str = "alive_tasks";
 }
 
-// KvRouter (including KvInexer) Prometheus metric names
+// KvRouter (including KvIndexer) Prometheus metric names
 pub mod kvrouter {
     /// Number of KV cache events applied to the index (including status)
     pub const KV_CACHE_EVENTS_APPLIED: &str = "kv_cache_events_applied";
 }
 
+/// KV Publisher metrics
+pub mod kv_publisher {
+    /// Total number of raw events dropped by engines before reaching publisher (detected via event_id gaps)
+    pub const ENGINES_DROPPED_EVENTS_TOTAL: &str = "kv_publisher_engines_dropped_events_total";
+}
+
 /// Additional TRT-LLM worker metrics beyond what the engine natively provides.
 ///
 /// These metrics are Python-only (registered via `prometheus_client`) and share the