Skip to navigation

Data plane chart release notes

This page applies to the Enterprise plan, where you run the data plane in your own VPC. It lists the changes in each version of the data plane Helm chart that matter to you, newest first, with the steps to take when you upgrade. Read every version between the one you run and the one you’re moving to. For the upgrade procedure, see Upgrade the data plane.

Chart versions are separate from gateway image versions. You set the gateway image with image.tag or image.digest.

Current chart

These changes are in the current chart but aren’t yet listed under a version number.

Added

  • Workload attestation. Set attestation.enabled: true to enroll gateway pods with their own projected Kubernetes service account token instead of a bootstrap token. The chart mounts the token and sets the matching environment variables. Register a matching trust anchor in the console first. See Connect the data plane to the control plane.
  • Control plane CA certificate in values. Set syncCA.pem to the control plane CA certificate, and the chart mounts it at /etc/fabric/sync/ca.crt. You can still create the fabric-dp-sync-ca Secret instead. See The control plane CA certificate.

Version 0.6.0

Upgrade notes

  • Production mode is now on by default. FABRIC_GW_PRODUCTION_MODE defaults to "true". A data plane without FABRIC_GW_SECRET_KEY now refuses to start, instead of running with its secret store turned off, and credential brokering enforces fail-closed behavior. Before you upgrade, make sure the key is in the app Secret. Set production mode to "false" only for throwaway local testing. This change is breaking for installs that relied on the old behavior.
  • Environment-specific defaults are now placeholders. The image repository, tenant ID, control plane address, MySQL host, and external URLs ship as placeholders, the private-host allow list for the gateway’s own authentication calls (FABRIC_GW_SSRF_ALLOW_PRIVATE_HOSTS) is empty, and the OpenTelemetry resource attributes are generic. If your values file relied on any of these defaults, set them explicitly before you upgrade.

Added

  • Distributed secret store. Set distributed.enabled: true to keep the secret store in memory across the gateway pods, with an encrypted snapshot on each pod’s volume, instead of in MySQL. The chart then runs the gateway as a StatefulSet, adds a headless Service and a NetworkPolicy for the store’s ports, and needs FABRIC_GW_DISTRIBUTED_GOSSIP_KEY and FABRIC_GW_DISTRIBUTED_AUTH_PASSWORD in the app Secret. It’s off by default. See Scale and run a highly available data plane.
  • OPA policy sidecar. Set opa.enabled: true to run OPA in the gateway pod. It’s off by default. The access-control, validator, and transformer policy plugins evaluate against it and have no in-process fallback. Without it, they fail according to the plugin’s on_error setting, which defaults to fail_closed.
  • LLM egress listener. Set llmEgress.enabled: true to open the Transponder callback port, 8090 by default. It’s off by default.
  • Collector sidecar switch. Set otelCollector.enabled: false to remove the OpenTelemetry Collector sidecar and its configuration, when the gateway sends telemetry straight to your own collector or backend. See Export telemetry with OpenTelemetry.
  • Documented settings for a coordination Redis, shared rate limiting, and related options, in the chart’s values file.

Changed

  • The distributed store needs no MySQL. With distributed.enabled, the schema migration Job is skipped. Usage metering always travels to the control plane on the data plane’s own connection, not through MySQL.

Version 0.5.0

Upgrade notes

  • The embedder now has a PodDisruptionBudget, on by default with minAvailable: 1. With the default single embedder replica, it blocks kubectl drain on the embedder’s node. Before your next node maintenance, set embedder.podDisruptionBudget.maxUnavailable: 1, or run more than one embedder replica. See Run the tool-search embedder.
  • The embedder now runs hardened: non-root (UID 1000), all capabilities dropped, no privilege escalation, and a read-only root filesystem, with /data and /tmp writable. The default image works as is. If you run a custom embedder image that writes elsewhere, mount that path as a volume or relax embedder.securityContext.readOnlyRootFilesystem.

Added

  • Embedder settings under embedder.*: the model (embedder.model.id), image (with digest pinning and pull Secrets), replica count, resources, and runtime arguments. Every setting defaults to the earlier behavior.
  • Persistent model cache. embedder.persistence.enabled keeps the model on a PVC, so it isn’t downloaded again on every restart.
  • Offline mode. embedder.offline.enabled stops the embedder from contacting huggingface.co, for air-gapped clusters with a pre-loaded model.
  • Embedder autoscaling and network policy. embedder.autoscaling.enabled adds a HorizontalPodAutoscaler, and embedder.networkPolicy.enabled restricts the embedder to the gateway pods. Both are off by default.

The embedder Service name and ports didn’t change.

Version 0.4.0

Added

  • Sidecar metrics by default. The collector sidecar now scrapes its own metrics and the embedder’s metrics, and ships them to your telemetry backend with the gateway’s. No Prometheus Operator is needed.
  • Prometheus Operator support. metrics.serviceMonitor.enabled renders a PodMonitor and a ServiceMonitor, and metrics.annotations.enabled adds prometheus.io/scrape pod annotations. Both are off by default. See Monitor the data plane.

Upgrade notes

No action needed. The collector configuration changes, so gateway pods roll on upgrade.

Version 0.3.0

Added

  • Collector metrics. The collector sidecar serves its own metrics on port 8888, including export failures.
  • Values validation. The chart ships a values schema, so some mistyped values fail when Helm renders the chart, for example in helm lint or helm template.

Upgrade notes

No action needed. The collector configuration changes, so gateway pods roll on upgrade.

Next steps