LMCache MP mode

.github/workflows/default-install-smoke.yml

@@ -12,24 +12,16 @@
 #   - cacheindex/cluster-default.status.observedServer is populated (proves the
 #     controller's CacheIndex poller is talking to the server's /snapshot)
 #   - gRPC LookupRoute returns reason_code=NO_HINT (fail-open default)
-#   - paired sample (config/samples/cachebackend-with-engine.yaml) wires the
-#     CacheBackend ↔ engine-pod binding: status.matchedEnginePods=1, the
-#     injected-by annotation is stamped on the engine pod, and an
-#     InjectedByCacheBackend Event lands on the persisted pod's UID; then
-#     scaling the engine to 0 drives status.matchedEnginePods=0 via the
-#     reconciler's RequeueAfter cadence (no CR-side change).
-#   - External CacheBackend end-to-end: type=External renders no
-#     Deployment/Service, status.endpoint mirrors spec.endpoint, Ready=True,
-#     a matching engine pod is admitted with LMCACHE_REMOTE_URL injected
-#     from the operator-supplied endpoint, and admission rejects the
-#     known-bad shapes (non-lm:// scheme, empty host, non-External + endpoint)
+#   - server /readyz and /metrics are reachable through the installed Service
+#   - current samples pass live server-side admission, and the generic
+#     CachePolicy/CacheTenant/PromptTemplate/PDTopology APIs remain usable
+#   - the served CacheBackend CRD contains only the MP schema
+#   - typed PodLocal MP admission injects the vLLM connector and native sidecar
+#   - managed Redis renders independently and publishes remote-storage status
+#   - an idempotent bundle re-apply preserves typed CacheBackend objects
 #
-# Lightweight (two distroless ~30 MB images + cert-manager + a busybox
-# stand-in for the engine container + a pause-image pod, no real engine
-# pull), so it runs on every PR. Sister-canaries
-# (c2-reconciler-canary, c6-engine-wiring-canary, cpu-substrate-canary)
-# cover real engine pods + cross-pod cache reuse with multi-GB images;
-# they stay schedule-only.
+# Lightweight: no real inference engine, LMCache server, GPU, or model is
+# started, so it runs on every PR.
 name: default-install-smoke
 
 on:
@@ -49,11 +41,9 @@ concurrency:
 jobs:
   install-smoke:
     runs-on: ubuntu-latest
-    # Script targets ~6 min end-to-end (incl. the Calico CNI install the smoke
-    # needs to enforce NetworkPolicy). The timeout is a circuit breaker for a
-    # wedged kind cluster / image pull, not the expected runtime; it sits above
-    # the worst-case Calico readiness budget (~510s) plus the smoke so a wedged
-    # CNI still leaves room for the exit trap to collect diagnostics.
+    # The timeout is a circuit breaker for a wedged kind cluster, image build,
+    # cert-manager rollout, or admission check; the exit trap still has time to
+    # collect diagnostics.
     timeout-minutes: 20
     steps:
       - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5  # v4

Makefile

@@ -538,8 +538,8 @@ install-hooks: ## Install git hooks (vendor-neutral naming guard) via core.hooks
 
 .PHONY: verify-naming
 verify-naming: ## Fail if core-identity files reference OCI/Oracle (see CONTRIBUTING.md).
-	@bad=$$(grep -rniEI '\boci\b|oci\.com|oraclecloud|\boracle\b' \
-		api proto gen pkg config/crd config/rbac config/default config/manager config/observability config/samples config/server config/webhook config/certmanager config/overlays docs/observability internal PROJECT go.mod 2>/dev/null || true); \
+	@bad=$$(grep -rniEI '(^|[^[:alnum:]_])(oci|oracle)([^[:alnum:]_]|$$)|ocir\.io|oci\.com|oraclecloud' \
+		api proto gen pkg config/crd config/rbac config/default config/manager config/observability config/samples config/server config/webhook config/certmanager config/overlays docs/design docs/observability docs/reference-stack internal PROJECT go.mod 2>/dev/null || true); \
 	if [ -n "$$bad" ]; then \
 		echo "✗ OCI/Oracle reference in core-identity files (banned per CONTRIBUTING.md):"; \
 		echo "$$bad" | sed 's/^/    /'; \

README.md

@@ -7,7 +7,7 @@ A **vendor-neutral, Kubernetes-native cache-policy control plane for LLM inferen
 inference-cache makes routing **cache-aware**: it tracks which replica already holds a
 prompt's prefix warm and returns that as a routing *hint*, so a gateway can reuse
 KV/prefill instead of recomputing it — cutting time-to-first-token and cost. It
-**orchestrates** existing KV-cache technology (LMCache, Mooncake); it is **not** a new
+**orchestrates** existing KV-cache technology (LMCache); it is **not** a new
 distributed cache and **not** the data-plane gateway. Guiding principle — **"we decide
 routing; the gateway follows"**: all routing intelligence lives in the server, and the
 gateway simply tokenizes → calls `LookupRoute` → routes to the returned replica →
@@ -27,16 +27,17 @@ The API separates three choices: `spec.runtime` selects the inference runtime,
 provider. Supporting another combination is an adapter addition; the core gRPC
 contract stays stable.
 
-- **vLLM + LMCache** supports host-only caching, a managed or external
-  `LMCacheServer`, and managed or external `Mooncake` storage.
-- **SGLang + LMCache** supports host-only caching or a managed/external Redis
-  remote store.
+- **vLLM + LMCache** supports typed PodLocal MP with host-only caching or an
+  optional managed/external Redis L3.
+- **SGLang + LMCache** supports the same typed PodLocal MP and Redis profiles;
+  the engine launch surface remains SGLang-specific.
 - **SGLang + SGLangHiCache** uses SGLang's native host tier and does not accept
   a remote-storage binding.
 
 See [`config/samples/`](config/samples/) for canonical manifests and
 [`docs/design/cachebackend-api.md`](docs/design/cachebackend-api.md) for the
-compatibility rules retained for older v1alpha1 resources.
+current contract and clearly labeled legacy-compatibility sections retained
+until Phase 7 of the migration.
 
 ## What's Inference Cache?
 
@@ -205,8 +206,7 @@ kubectl get cacheindex cluster-default -o yaml
 Both binaries expose Prometheus metrics on their pod's `:8080/metrics`
 (prefixed `inferencecache_*`) — the server binary's series cover the
 in-memory index and gRPC handlers; the controller binary's series cover
-the reconcilers (e.g. `inferencecache_backend_probe_result_total`,
-`inferencecache_backend_server_restart_cascades_total`). A default
+the reconcilers (for example, `inferencecache_backend_probe_result_total`). A default
 alert bundle for the operational silent-failure patterns this code has
 hit in production ships under
 [`config/observability/`](config/observability/) and is **not** included
@@ -219,13 +219,13 @@ For prometheus-operator / kube-prometheus installs:
 kubectl apply -k config/observability
 ```
 
-This ships THREE resources: a `ServiceMonitor` (so Prometheus scrapes
-`inference-cache-server:8080/metrics`), a `PodMonitor` (so Prometheus
-scrapes the controller pod's `:8080/metrics` — required for the
-controller-side alerts like `ServerProbeFail` to have a series to
-evaluate), and the `PrometheusRule` carrying the alerts.
+This ships FOUR resources: a `ServiceMonitor` for
+`inference-cache-server:8080/metrics`, one `PodMonitor` for the controller
+pod's `:8080/metrics`, one cross-namespace `PodMonitor` for successfully
+injected PodLocal LMCache sidecars on their named `lmcache-http` port, and
+the `PrometheusRule` carrying the alerts.
 
-> **Caveat — Prometheus Operator selectors.** All three CRs carry
+> **Caveat — Prometheus Operator selectors.** All four CRs carry
 > example labels (`prometheus: k8s`, plus `role: alert-rules` on the
 > PrometheusRule) that match the upstream kube-prometheus stack
 > (default `Prometheus` named `k8s`). The `kube-prometheus-stack`
@@ -250,19 +250,19 @@ expressions) and only fire when the conditions are met.
 > **The fifth alert needs a vLLM scrape this bundle does NOT ship.**
 > [`LMCacheT2NoHits`](docs/observability/alerts.md#lmcachet2nohits) reads
 > `vllm:external_prefix_cache_*` from vLLM engine pods directly. The
-> shipped `ServiceMonitor` covers only `inference-cache-server`. To make
+> shipped scrape configs do not collect vLLM's own metrics. To make
 > that alert effective, add a separate `PodMonitor` for your vLLM
 > Deployment (or `kubernetes_sd_configs: pod` for vanilla Prometheus)
 > so engine `/metrics` is scraped with both `namespace` and `pod` labels
 > attached. See alerts.md "How to enable" for the requirement.
 
 For vanilla Prometheus, ConfigMap mounts, or Helm `prometheus.serverFiles`,
 use the flat [`alerting-rules.yaml`](config/observability/alerting-rules.yaml).
-**You must also configure scraping yourself, for BOTH the server AND
-the controller pod.** The server's `:8080` exposes the index, lookup,
+**You must also configure scraping yourself for the server, the controller
+pod, and every injected PodLocal LMCache sidecar.** The server's `:8080` exposes the index, lookup,
 and auth series; the controller pod's `:8080` exposes the per-stage
-probe-result counter (`inferencecache_backend_probe_result_total`)
-and the cache-server restart-cascade counter — the controller-side
+probe-result counter (`inferencecache_backend_probe_result_total`); each LMCache sidecar exposes
+its own `lmcache_mp_*` series on `:8080/metrics` — the controller-side
 alerts (`ServerProbeFail` today) load against the controller's
 series, so a server-only scrape leaves them inert.
 

api/v1alpha1/cachebackend_effective_test.go

@@ -13,7 +13,7 @@ func TestEffectiveRemoteStorageUsesOnlyExplicitDeclaration(t *testing.T) {
 	}
 
 	want := &CacheBackendRemoteStorageSpec{
-		Provider:  CacheBackendRemoteStorageProviderMooncake,
+		Provider:  CacheBackendRemoteStorageProviderRedis,
 		Ownership: CacheBackendRemoteStorageOwnershipManaged,
 	}
 	spec.RemoteStorage = want