Skip to navigation

Requirements and sizing

This page applies to the Enterprise plan, where you run the data plane in your own VPC. Before you install, use it to check what the data plane needs from your cluster, how much capacity to reserve, and which external systems you provide. For the network paths, see Network and egress.

Kubernetes requirements

  • A conformant Kubernetes cluster, version 1.23 or later. With the optional features off, the chart renders only core resources (Deployment, Service, ServiceAccount, ConfigMap, PodDisruptionBudget, and Job) and needs no CRDs.
  • kubectl and Helm 3 or later on the machine you install from, with access to the cluster.
  • A namespace for the data plane. These pages use pm-fabric-dp, but you can choose another name.
  • Pod security that allows the chart’s defaults. The gateway and embedder containers run as non-root with the RuntimeDefault seccomp profile, drop all Linux capabilities, and use a read-only root filesystem. The collector sidecar and the schema migration Job don’t set all of these, so the data plane doesn’t meet the Kubernetes restricted Pod Security Standard. Use the baseline level for the data plane’s namespace, for example kubectl label namespace pm-fabric-dp pod-security.kubernetes.io/enforce=baseline. No component needs privileged access.

Resource sizing

The chart’s default requests and limits are below. Plan node capacity against the limits.

ComponentReplicasRequests (CPU / memory)Limits (CPU / memory)
Gateway container2300m / 512Mi500m / 1024Mi
Collector sidecar, per gateway pod2100m / 256Mi500m / 512Mi
Embedder1500m / 1Gi2 / 2Gi
Schema migration Job (runs briefly)1100m / 128Mi250m / 256Mi
OPA sidecar, per gateway pod (only with opa.enabled)250m / 64Mi250m / 128Mi

With the default replica counts, a steady-state install requests about 1.3 vCPU and 2.5 GiB of memory, and its limits add up to 4 vCPU and 5 GiB. The migration Job runs only during installs and upgrades, so it isn’t part of steady-state capacity.

With the distributed secret store, the chart runs three gateway pods by default (distributed.replicaCount), each with a 1 GiB volume. Multiply the per-pod rows by three instead of two.

These figures are starting points, not a capacity model for your traffic. To change them, set container.requests and container.limits for the gateway and embedder.resources for the embedder. For replicas and autoscaling, see Scale and high availability. The collector’s requests and limits are in the otel-collector entry under additionalContainers. Helm replaces lists instead of merging them, so to change them, copy the whole entry from the chart’s values.yaml into your values file and edit only resources.

External dependencies you provide

  • A container registry. Mirror the gateway image Postman gives you into a registry your cluster can pull from. The chart also pulls these public images by default: flyway/flyway:10 for the migration Job, ghcr.io/huggingface/text-embeddings-inference:cpu-1.9 for the embedder, otel/opentelemetry-collector-contrib:0.149.0 for the collector sidecar, and openpolicyagent/opa:1.20.2 for the OPA sidecar. Mirror them too if your cluster can’t reach those registries. The embedder also downloads its model from huggingface.co at startup unless you run it offline.

    If your registry needs credentials, create an image pull Secret and reference it for each component:

    • The gateway pod, including its sidecars: image.imagePullSecrets.
    • The embedder: embedder.image.imagePullSecrets.
    • The migration Job: it has no pull Secret setting, so add the Secret to the namespace’s default ServiceAccount. See Install with Helm.
  • A secret store for the data plane. Choose one:

    • MySQL (the chart’s default). A MySQL instance reachable from the cluster, a database for the data plane (the chart defaults to fabric_gateway), and a user with CREATE, ALTER, INDEX, SELECT, INSERT, UPDATE, and DELETE grants. Give the data plane a database of its own. The chart creates the tables, not the database or the user. See Database migrations.
    • The distributed store. The gateway pods share credential state among themselves, with no database. It needs a StorageClass for one volume per pod and a NetworkPolicy controller (see the note under Optional-feature cluster requirements). See Scale and high availability.
  • The control plane CA certificate. You supply it in your values or as a Secret. See Secrets and keys.

  • The app Secret. A Kubernetes Secret with your encryption key and passwords, created before you install. See Secrets and keys.

  • A telemetry backend, if you want one. The gateway emits standard OTLP over gRPC. You can send it to ClickHouse (the chart’s default exporter), to any OTLP-compatible backend, or nowhere. See OpenTelemetry.

  • Ingress for your clients. The chart doesn’t create an Ingress. You decide how clients reach the gateway and where TLS ends. See Expose the gateway.

Optional-feature cluster requirements

None of these are needed for a default install. Each matters only if you turn on the matching value.

ValueCluster requirement
autoscaling.enabled, embedder.autoscaling.enabledmetrics-server
keda.enabledThe KEDA operator
metrics.serviceMonitor.enabledThe Prometheus Operator
metrics.annotations.enabledAn annotation-based Prometheus scraper
istio.enabledIstio
reloader.enabledStakater Reloader
embedder.networkPolicy.enabledA NetworkPolicy controller
embedder.persistence.enabledA default StorageClass. With more than one embedder replica, or an embedder autoscaler that can scale past one, set embedder.persistence.accessModes to ReadWriteMany and use a StorageClass that supports it, for example EFS on AWS. You can also set embedder.persistence.existingClaim to a claim you created. Otherwise the install fails.
distributed.enabledA StorageClass for per-pod volumes and a NetworkPolicy controller
attestation.enabledProjected service account tokens, with an OIDC issuer the control plane can verify.
iam.enabledCrossplane with the AWS provider. AWS only.

keda.enabled, metrics.serviceMonitor.enabled, and iam.enabled create custom resources, so the install fails if your cluster doesn’t have their CRDs. The other values install without errors but have no effect until the matching component is running. For example, an autoscaler without metrics-server never scales, and a NetworkPolicy without a controller isn’t enforced.

With distributed.enabled, the chart’s NetworkPolicy allows only the store’s ports, 3320 and 3322, into gateway pods. On a cluster that enforces NetworkPolicy, it also blocks client traffic to port 8080. Add a rule for 8080 to your values file:

distributed:
networkPolicy:
extraIngress:
- ports:
- protocol: TCP
port: 8080

Next steps