diff --git a/README.md b/README.md
index d8f0f0d..4148c1b 100644
--- a/README.md
+++ b/README.md
@@ -108,6 +108,23 @@ Current behavior (Helm values in [`helm.tf`](terraform/helm.tf)):
 
 Stale-while-revalidate: fresh hits return immediately; between `cacheTtl` and `redis.ttl` the firewall revalidates with Socket and falls back to the stale verdict if the API (or breaker) fails. After `redis.ttl` the key expires and the next request must fetch fresh (or fail-open if still down).
 
+## Availability tuning
+
+A client-visible `503` from this stack is almost never the upstream registry — it is the ALB having no healthy backend to send to, or a backend that went away mid-request. The settings below exist to close the three ways that happens. Note that CI is where this surfaces first: a lockfile resolve is a burst of parallel artifact fetches, so it hits saturation and pod churn that ordinary developer installs never reach.
+
+| `statusDetails` | Failure mode | Setting | Where |
+|-----------------|--------------|---------|-------|
+| `backend_connection_closed_before_data_sent_to_client` | **Keepalive race.** The LB holds idle connections to the backend for a fixed, non-configurable 600s. The chart's default `keepaliveTimeout: 65` makes nginx close them first, and a request that races the close becomes a 503. Independent of traffic and deploys, which is what makes it look random | `keepaliveTimeout = 620`, per Google's documented fix | [`helm.tf`](terraform/helm.tf) |
+| `backend_connection_closed_before_data_sent_to_client` | **Pod churn.** A pod leaving the NEG (rollout, HPA scale-down, node autoscale/repair/upgrade) has its in-flight requests cut | `connectionDraining.drainingTimeoutSec = 60` and `terminationGracePeriodSeconds = 210` (Google's `Tgrace > TpreStop + Tdrain + Tbuffer`) | [`tls.tf`](terraform/tls.tf), [`helm.tf`](terraform/helm.tf) |
+| `backend_connection_closed_before_data_sent_to_client` | **Pod OOM-killed or evicted**, dropping every request it was serving | Memory request == limit at `3Gi` (never above its request, so not ranked for eviction). The previous `768Mi` limit is the value the chart calls out as OOM-killing on large package indexes | [`helm.tf`](terraform/helm.tf) |
+| `failed_to_pick_backend` | **Health check ejection.** Burst saturates the pods, `/health` slows, the LB ejects the backend — and with few replicas that shifts load onto the rest and cascades | `nginx.workerProcesses = 2` matched to the 1.5-CPU limit (the chart's default 4 oversubscribes the cgroup quota), `unhealthyThreshold = 3` / `healthyThreshold = 1`, HPA target 60%, and `node_max_count` headroom so scale-up isn't capped | [`helm.tf`](terraform/helm.tf), [`tls.tf`](terraform/tls.tf), [`terraform.tfvars`](terraform/terraform.tfvars) |
+
+> **Known residue:** the chart exposes no `preStop` hook, so `TpreStop` is 0 and nothing holds a pod open for the ~120s an endpoint can take to leave the NEG. In-flight requests are covered by draining, but requests the LB dispatches during that window still land on a shutting-down pod. Expect `backend_connection_closed_before_data_sent_to_client` to shrink substantially rather than reach zero; closing the gap needs `lifecycle.preStop` support upstream.
+
+The firewall's own resilience settings (fail-open, circuit breaker, Redis stale cache) do **not** cover any of these — they govern what the firewall decides when the *Socket API* is degraded, not whether the load balancer has a pod to talk to.
+
+Pod resources stay below the chart's `4 CPU / 8Gi` default, which is sized for metadata filtering (disabled here) and would not schedule on `e2-standard-2` nodes. Requests are sized so one pod fills most of a node, which also keeps replicas on separate nodes. Only `node_min_count` is billed at idle; `node_max_count` is burst headroom.
+
 ## Components
 
 | Layer | Resource | Purpose |
@@ -120,6 +137,7 @@ Stale-while-revalidate: fresh hits return immediately; between `cacheTtl` and `r
 | **App** | Helm `socket-firewall` | Package firewall with path-based routing, HPA, pod anti-affinity, PDB; Redis-backed verdict cache with fail-open on breaker trip |
 | **Cache** | Memorystore Redis (AUTH + TLS) | Shared stale-while-revalidate cache across replicas so a tripped Socket API circuit breaker serves last known-good verdicts |
 | **Exposure** | GKE Gateway (`gke-l7-global-external-managed`) | External HTTPS when `firewall_domain` is set; `LoadBalancer` Service fallback when domain is unset |
+| **Backend tuning** | GCPBackendPolicy | 60s connection draining so pod churn doesn't reset in-flight requests, 300s backend timeout for large artifacts, LB request logging for 5xx attribution |
 | **Secrets** | Secret Manager (CMEK) → K8s secret | `SOCKET_SECURITY_API_TOKEN` for Socket.dev |
 | **TLS** | Certificate Manager + GKE Gateway + SSL policy | Google-managed cert; HTTPS terminates at the LB (TLS 1.2 minimum) |
 | **Policy** | Kubernetes NetworkPolicies | Default-deny ingress in the firewall namespace (`enable_network_policies = true`) |
@@ -174,6 +192,7 @@ When `firewall_domain` is set, Terraform always provisions GCP-managed TLS:
 4. A **global SSL policy** (`RESTRICTED`, TLS 1.2 minimum) attached via **GCPGatewayPolicy**
 5. An **HTTPRoute** — forwards traffic to the firewall pods (backend `port 80`, plain HTTP)
 6. A **HealthCheckPolicy** — points the load-balancer health check at `/health`. Without it the LB defaults to probing `/`, which the firewall does not answer with `200`, so every backend is marked unhealthy (`no healthy upstream`)
+7. A **GCPBackendPolicy** — connection draining, backend timeout, and request logging on the backend service (see [Availability tuning](#availability-tuning))
 
 The Gateway terminates the public, browser-trusted certificate and forwards to the pods over **plain HTTP on port 80** (cluster-internal `ClusterIP`, never externally reachable). The firewall image nevertheless always configures an HTTPS listener (`:8443`) that requires a certificate to load, so the chart's cert-generator init container produces a **self-signed certificate** purely so nginx will boot — it is not on the Gateway's data path.
 
@@ -192,7 +211,7 @@ The Gateway terminates the public, browser-trusted certificate and forwards to t
 | [`secrets.tf`](terraform/secrets.tf) | CMEK-encrypted Secret Manager secret for the Socket API token |
 | [`helm.tf`](terraform/helm.tf) | Namespace, K8s secret, Helm release (fail-open + Redis stale cache + circuit breaker; pod `securityContext`/`fsGroup` so nginx can read the generated cert key) |
 | [`redis.tf`](terraform/redis.tf) | Memorystore Redis (PSA, AUTH, TLS CA) + K8s secrets for the firewall verdict cache |
-| [`tls.tf`](terraform/tls.tf) | Certificate Manager, SSL policy, GKE Gateway, GCPGatewayPolicy, HTTPRoute, HealthCheckPolicy |
+| [`tls.tf`](terraform/tls.tf) | Certificate Manager, SSL policy, GKE Gateway, GCPGatewayPolicy, HTTPRoute, HealthCheckPolicy, GCPBackendPolicy |
 | [`variables.tf`](terraform/variables.tf) | Input variable declarations (chart/image versions are required, no default) |
 | [`terraform.tfvars`](terraform/terraform.tfvars) | Concrete pinned values Terraform auto-loads (project, SA emails, node counts, `firewall_domain`, chart/image versions) |
 | [`outputs.tf`](terraform/outputs.tf) | Cluster credentials, gateway IP, DNS auth record, health URL |
@@ -377,8 +396,35 @@ gcloud logging read \
 | Pods `CrashLoopBackOff`, logs show `cannot load certificate key ... Permission denied` | Cert-generator init runs as UID 1000 but the image runs as UID 1001, so nginx can't read the `0600` key | `podSecurityContext.fsGroup` + cert-generator `runAsUser` aligned to the image UID (set in `helm.tf`) |
 | `503 no healthy upstream` while pods are `Ready` | LB health check probes `/` (its default), which the firewall doesn't answer `200` | `HealthCheckPolicy` pointing the LB health check at `/health` (in `tls.tf`) |
 | CI package downloads 503; logs show `lua ssl certificate verify error: (21: unable to verify the first certificate)` to Redis `:6378` | Image ≤ 2.0.10 writes the Memorystore CA bundle into `/etc/nginx/ssl`, which the chart mounts read-only. The write fails, startup continues on OS roots, and every Redis TLS handshake fails. Shared cache is dead; bursty CI then 503s. | Pin `firewall_image_tag` ≥ `2.1.1` (writes the bundle to writable `/app/ca-bundle.pem`). Confirm Binary Authorization allows the new digest before apply. |
+| Intermittent 503 with no pod-level error | Backend went away mid-request or the LB had no healthy backend — pod churn, OOM kill/eviction, or burst saturation | See [Availability tuning](#availability-tuning); attribute the specific cause with `statusDetails` in the LB logs (below) before changing anything |
 | Gateway "load balancer" not visible | The data-plane LB is a Gateway-managed **global external ALB** (`gkegw1-…`), not a `LoadBalancer` Service | `kubectl get gateway -A` for the address; `gcloud compute forwarding-rules list --global` |
 
+### Attributing a 5xx
+
+The GCPBackendPolicy in [`tls.tf`](terraform/tls.tf) enables LB request logging, so every response carries a `statusDetails` naming the cause. Group the 5xx responses by it:
+
+```bash
+gcloud logging read \
+  'resource.type="http_load_balancer"
+   httpRequest.status>=500' \
+  --project <project_id> --limit 200 --freshness=6h \
+  --format='value(jsonPayload.statusDetails)' | sort | uniq -c | sort -rn
+```
+
+| `statusDetails` | Meaning | Where to look |
+|-----------------|---------|---------------|
+| `backend_connection_closed_before_data_sent_to_client` | Backend closed the connection first — keepalive race, pod termination, or an OOM kill | `keepaliveTimeout`, connection draining, grace period, restart counts |
+| `failed_to_pick_backend` | No backend passing the health check at that moment | Health check thresholds, pod readiness, burst saturation |
+| `failed_to_connect_to_backend` | Backend reachable but refusing connections — saturated, or listener already closed while still in the NEG | CPU limits and worker count, the preStop gap above |
+| `response_sent_by_backend` | The firewall itself returned the 5xx | Container logs (Cloud Logging query above) |
+
+Confirm or rule out OOM kills and restarts directly:
+
+```bash
+kubectl get pods -n socket-firewall \
+  -o custom-columns='NAME:.metadata.name,RESTARTS:.status.containerStatuses[0].restartCount,LAST:.status.containerStatuses[0].lastState.terminated.reason'
+```
+
 ### Verifying the data plane
 
 ```bash
diff --git a/terraform/helm.tf b/terraform/helm.tf
index f212feb..232c4ad 100644
--- a/terraform/helm.tf
+++ b/terraform/helm.tf
@@ -11,10 +11,38 @@ locals {
       pullPolicy = "IfNotPresent"
     }
 
+    # CPU-driven HPA. Target 60% rather than the chart's 70% so a CI burst
+    # starts adding replicas before the running pods are already saturated —
+    # scale-up has to wait on metrics, pod start, and (past node_min_count) a
+    # node. maxReplicas is bounded by node_max_count, since one pod fills most
+    # of a node at the requests below; asking for more would leave pods Pending.
     autoscaling = {
-      enabled = true
+      enabled                        = true
+      minReplicas                    = var.replica_count
+      maxReplicas                    = var.max_replica_count
+      targetCPUUtilizationPercentage = 60
     }
 
+    # The load balancer holds idle keepalive connections to the backend for a
+    # fixed 600s that GCP does not let you configure. The chart's default of 65s
+    # means nginx closes an idle connection the LB still considers usable; when a
+    # request races that close the client gets a 503
+    # (backend_connection_closed_before_data_sent_to_client — the top entry in
+    # this deployment's LB logs). Google's documented fix is a backend keepalive
+    # above 600s, recommended 620.
+    keepaliveTimeout = 620
+
+    # Google's guidance for the same error on GKE NEG backends:
+    # Tgrace > TpreStop + Tdrain + Tbuffer, at least 210s. Tdrain is the 60s
+    # connection draining in tls.tf.
+    #
+    # TpreStop is 0 here because the chart exposes no preStop hook, so nothing
+    # holds the pod open for the ~120s the endpoint can take to leave the NEG.
+    # In-flight requests are covered, but requests the LB sends in that window
+    # still land on a shutting-down pod. That residue is the reason to expect
+    # this error to shrink rather than vanish.
+    terminationGracePeriodSeconds = 210
+
     # The socket-registry-firewall image runs as UID 1001, but the
     # chart's cert-generator init container defaults to runAsUser 1000 and writes
     # /etc/nginx/ssl/privkey.pem mode 0600 (owner-only). The UID mismatch makes
@@ -110,17 +138,36 @@ locals {
       generateSelfSigned = true
     }
 
+    # The old 768Mi limit was the value the chart explicitly calls out as
+    # OOM-killing on large package indexes, and 512Mi requests put the pod in
+    # Burstable QoS, so it was also an eviction candidate under node pressure.
+    # An OOM kill or eviction drops every in-flight request on that pod, which
+    # reaches the client as a 503. Memory request == limit so the pod never sits
+    # above its request and is not ranked for eviction.
+    #
+    # Still below the chart's 4 CPU / 8Gi default: that assumes metadata
+    # filtering (disabled here) and would not schedule on the current nodes.
+    # These requests fit one pod per e2-standard-2 node alongside system pods,
+    # which also keeps replicas on separate nodes.
     resources = {
       requests = {
-        cpu    = "500m"
-        memory = "512Mi"
+        cpu    = "750m"
+        memory = "3Gi"
       }
       limits = {
-        cpu    = "1"
-        memory = "768Mi"
+        cpu    = "1500m"
+        memory = "3Gi"
       }
     }
 
+    # Default is 4 workers. With a 1.5-CPU limit that oversubscribes the cgroup
+    # quota and the workers throttle each other, which shows up first as slow
+    # /health responses — enough of them and the LB drops the backend and
+    # returns 503 while the pod is still Ready.
+    nginx = {
+      workerProcesses = 2
+    }
+
     # Spread replicas across nodes so a single node loss doesn't take down
     # the firewall. Soft (preferred) so scheduling still succeeds on one node.
     affinity = {
diff --git a/terraform/terraform.tfvars b/terraform/terraform.tfvars
index 6b416dd..d4657e1 100644
--- a/terraform/terraform.tfvars
+++ b/terraform/terraform.tfvars
@@ -8,10 +8,11 @@ terraformer_plan = "socket-firewall-tf-plan@sac-prod-sa.iam.gserviceaccount.com"
 cluster_name      = "socket-firewall"
 node_machine_type = "e2-standard-2"
 node_min_count    = 2
-node_max_count    = 3
+node_max_count    = 6
 
 firewall_domain = "sfw.security.sentry.io."
 
 replica_count      = 2
+max_replica_count  = 6
 helm_chart_version = "0.11.2"
 firewall_image_tag = "2.1.1"
diff --git a/terraform/tls.tf b/terraform/tls.tf
index 33c69a6..09bfc8e 100644
--- a/terraform/tls.tf
+++ b/terraform/tls.tf
@@ -207,6 +207,16 @@ resource "kubectl_manifest" "socket_firewall_health_check" {
     }
     spec = {
       default = {
+        # Defaults (unhealthyThreshold 2, healthyThreshold 2, 5s interval) eject
+        # a backend after ~10s of slow /health responses and then need two
+        # successes to bring it back. With only a couple of replicas, ejecting
+        # one doubles load on the rest and can cascade — which is what
+        # failed_to_pick_backend in the LB logs looks like. Tolerate one more
+        # consecutive failure, and return the backend on the first success.
+        checkIntervalSec   = 5
+        timeoutSec         = 5
+        unhealthyThreshold = 3
+        healthyThreshold   = 1
         config = {
           type = "HTTP"
           httpHealthCheck = {
@@ -225,3 +235,52 @@ resource "kubectl_manifest" "socket_firewall_health_check" {
 
   depends_on = [helm_release.socket_firewall]
 }
+
+# Backend-service behaviour for the Gateway-managed ALB. Without this the
+# backend runs on GCP defaults: connection draining disabled (0s) and a 30s
+# request timeout.
+#
+# drainingTimeoutSec: with draining off, a pod leaving the NEG (rollout, HPA
+# scale-down, node autoscale/repair/upgrade) has its in-flight connections cut
+# immediately, which reaches the client as a sporadic 503. The chart has no
+# preStop hook, so LB-side draining is the drain available;
+# terminationGracePeriodSeconds in helm.tf is set to outlive this window.
+#
+# timeoutSec: the firewall streams package artifacts, and large wheels
+# (torch, nvidia-*) can exceed the 30s default on a slow upstream. 300s matches
+# the firewall's own proxy.read_timeout default.
+#
+# logging: LB request logs carry jsonPayload.statusDetails, which names the
+# reason for a 5xx (failed_to_pick_backend, backend_connection_closed_*, ...).
+# Without it a 503 seen by a client is not attributable to a cause. Add
+# sampleRate (1-1000000) to cut ingest volume once the cause is known.
+resource "kubectl_manifest" "socket_firewall_backend_policy" {
+  count = local.use_gcp_managed_tls ? 1 : 0
+
+  yaml_body = yamlencode({
+    apiVersion = "networking.gke.io/v1"
+    kind       = "GCPBackendPolicy"
+    metadata = {
+      name      = "${var.cluster_name}-backend-policy"
+      namespace = var.firewall_namespace
+    }
+    spec = {
+      default = {
+        timeoutSec = 300
+        connectionDraining = {
+          drainingTimeoutSec = 60
+        }
+        logging = {
+          enabled = true
+        }
+      }
+      targetRef = {
+        group = ""
+        kind  = "Service"
+        name  = helm_release.socket_firewall.name
+      }
+    }
+  })
+
+  depends_on = [helm_release.socket_firewall]
+}
diff --git a/terraform/variables.tf b/terraform/variables.tf
index 624db43..2be7433 100644
--- a/terraform/variables.tf
+++ b/terraform/variables.tf
@@ -84,9 +84,9 @@ variable "node_min_count" {
 }
 
 variable "node_max_count" {
-  description = "Maximum number of nodes (autoscaling)"
+  description = "Maximum number of nodes (autoscaling). Only the floor (node_min_count) is billed at idle, so this is headroom for CI bursts rather than steady-state cost."
   type        = number
-  default     = 3
+  default     = 6
 }
 
 variable "kubernetes_version" {
@@ -126,11 +126,17 @@ variable "firewall_image_tag" {
 }
 
 variable "replica_count" {
-  description = "Number of firewall pod replicas (ignored when HPA is enabled; used as a baseline for the chart)"
+  description = "Baseline number of firewall pod replicas (used as the HPA floor)"
   type        = number
   default     = 2
 }
 
+variable "max_replica_count" {
+  description = "HPA ceiling for firewall pod replicas. Keep at or below node_max_count — one pod fills most of a node, so a higher ceiling only produces Pending pods."
+  type        = number
+  default     = 6
+}
+
 variable "enable_network_policies" {
   description = "Apply default-deny-ingress NetworkPolicies in the firewall namespace (requires Calico enforcement on the cluster)"
   type        = bool