Changelog

Notable changes in each release of the Apollo GraphOS Operator


1.7.0

Features

  • Track mutable OCI tags for Supergraph and Subgraph schemas

    The Operator now resolves OCI references to a manifest digest on every poll (expirations.schema), loads the schema by that digest, and records it in the resource status. Routers are given a digest-pinned reference, so they roll when a tag moves without needing to hot-reload. Subgraph statuses gain an optional digest field.

Bug fixes

  • Stop reporting changes the API server doesn't store

    The Operator decided whether a Deployment or Service had changed by comparing it with a client-side merge. The API server stores some values differently from how they are written (a 1024Mi memory limit is stored as 1Gi, 1000m CPU as 1), so such a spec never compared equal and every reconcile counted as a change. Supergraphs and McpServers with these values stayed Ready=False with reason HasChanges, and each reconcile wrote to the API server for nothing.

    A change is now reported only when the write moves the resource's resourceVersion, so equivalent values no longer count.

  • Fix ${env.VAR} expansion and inconsistent keys in the operator config

    The operator never enabled environment variable expansion, so ${env.VAR} references in the config file were not expanded despite being documented. They now are, and an unset variable fails the load.

    Config keys are now consistently camelCase, matching the docs: controllers.common.defaultNamespace, requeueDelays, studioRateLimit, supergraphSchema.idleRequeueInterval, reconcileTimeout and telemetry.exporters.*.otlp.grpc.domainName. The previous snake_case spellings are still accepted but deprecated: the operator logs a warning for each one at startup and on config reload, naming the key to use instead. The http section is defined by apollo-http-client and remains snake_case for now.

    studioRateLimit.interval, supergraphSchema.composition.debounce, composition.interval and idleRequeueInterval now accept duration strings such as 5s, as documented.

  • Accept managed Kubernetes version strings in the Helm chart

    The chart's kubeVersion constraint is now >=1.34.0-0, so clusters that report a pre-release style version such as v1.36.5-eks-abc123 or v1.36.5-gke.100 are no longer rejected by Helm's semver check.

  • Keep the order of env vars, ports, volume mounts and volumes in generated pod templates

    Supergraph and McpServer pod templates were built from hash maps, so the order of these lists was random on every reconcile. This broke $(VAR) references in user-supplied env vars, since Kubernetes only expands variables defined earlier in the list. User-supplied entries now keep the order they were given in, and the Operator's own entries follow them, replacing any user entry with the same name, port or mount path in place.

    Existing Deployments are not rolled by this change. A Deployment already running with an unhelpful order picks up the new one the next time its pod template changes.

  • Stop spec.labels changes from breaking a Supergraph's or McpServer's Deployment

    spec.labels were part of the Deployment's selector, which can't be changed once the Deployment exists. Adding a label made every update to the Deployment fail, including unrelated ones such as routerConfig or replicas, while the Service was still updated to select the new label and so matched no pods.

    Selectors now only use the label the Operator itself sets (apollographql.com/supergraph or apollographql.com/mcp-server), which never changes. spec.labels are still applied to every resource and to the pods. An existing Deployment or Rollout keeps the selector it was created with, and its pod template keeps the labels that selector needs.

  • Report failed writes to a Supergraph's or McpServer's Deployment, Service and ConfigMap

    A failed write used to be dropped without a trace: nothing in the logs, and the resource stayed Ready with reason DeploymentCompleted while its Deployment was never updated. The reconcile now fails, which logs the error and retries, and a Warning event with reason UpdateFailed is published on the resource so kubectl describe shows it. A write that merely loses a race with another writer is retried shortly without being reported as an error.

1.6.2

Bug fixes

  • Log Kubernetes API requests that are still pending after 15 seconds, with their method, path and query, so a stalled call is visible. Also label the Subgraph schema status log with subgraph rather than supergraph.

  • Use the standard kube-rs HTTP client for Kubernetes API requests instead of the shared Apollo HTTP client. Kubernetes API calls no longer emit the http.client.* metrics.

1.6.1

Bug fixes

  • Allow non-object top-level values in McpServer and Supergraph config

    The McpServer spec.config and Supergraph spec.routerConfig CRD schemas required every top-level value to be an object, so the API server rejected valid keys such as the MCP Server's instructions (string) and forward_headers (array), or the Router's forbid_mutations (boolean). The CRDs now only require the config itself to be an object. Upgrade the CRDs (Helm doesn't upgrade CRDs automatically) to pick up the fix.

1.6.0

Features

  • Mint McpServer API keys with the OBSERVER role

    The MCP Server only reads the graph's schema, so the Operator now mints its graph API key with the OBSERVER role instead of GRAPH_ADMIN. Supergraph API keys are unchanged. Existing McpServer keys aren't rotated: delete an McpServer's API key Secret to have the Operator mint a new OBSERVER key.

  • Allow schema.source to be configured on McpServer

    The Operator no longer forces schema.source to uplink in an McpServer's spec.config. It still defaults to uplink, but you can now set graphos (fetch the schema from the GraphOS Platform API, which also supports non-federated graphs) or local. schema.source must be unset or uplink when the McpServer references a Supergraph resource, so that the MCP Server stays on the same schema source as the Supergraph's routers; any other value fails the McpServer's reconciliation.

Bug fixes

  • Surface ECR credential helper errors in the operator

    The ECR credential helper only logs to a file and reports every failure as "credentials not found in native keychain", which hid the real cause. The image now points that log at the helper's stderr, so the operator's error includes why a token could not be retrieved.

1.5.0

Features

  • Add McpServer CRD, which supports referencing an existing Supergraph CR or a Supergraph stored in Studio.

    Adding a new CRD, which represents a deployment of the Apollo MCP Server. The CRD can reference either an existing Supergraph CR or any Supergraph that is already in Studio. This also means the Operator can be used to deploy only MCPServers if required!

    For more details see documentation here

  • Update minimum supported Kubernetes version to 1.34

    The operator now requires Kubernetes 1.34 or later, following the N-3 support policy.

  • Document and harden TLS/proxy configuration for outbound Apollo traffic

    The http config block's tls and proxy settings (a custom CA bundle, mutual TLS, and a forward proxy for Apollo Studio/OCI traffic) are now documented here. This lets the operator run in environments with TLS-inspecting egress proxies or corporate forward proxies.

    Config values also support ${env.VAR_NAME} interpolation, so a CA bundle, client certificate, or proxy URL can be sourced from an environment variable (for example one populated from a mounted Secret) instead of inlined as a literal value.

  • Add imagePullSecrets support to the Operator Helm chart

    The Operator's Deployment now accepts a top-level imagePullSecrets value, so the Operator image can be pulled from a private registry that requires authentication (for example, a JFrog proxy-mirror). Provide a list of references to existing Kubernetes docker-registry Secrets:

    YAML
    1imagePullSecrets:
    2  - name: my-registry-secret
  • Add kubectl get printer columns to all CRDs

    Supergraph, SupergraphSchema, SupergraphSet, Subgraph, and McpServer previously only showed NAME/AGE under kubectl get. Each now surfaces its key status at a glance (readiness condition, or suspend for SupergraphSet), with secondary conditions and replica counts available via -o wide.

Bug fixes

  • Resolve an issues H1 and H2 connection pools were not evicting dead connections on error

1.4.0

Features

  • Update minimum supported Kubernetes version to 1.33

    The operator now requires Kubernetes 1.33 or later, following the N-3 support policy.

Bug fixes

  • Add a default startup probe to the router container

    Without a startup probe, Kubernetes evaluates the readiness probe immediately on container start. Because it fires on a fixed periodSeconds: 10 schedule, a router that becomes ready at t+11s won't be marked ready until the next tick at t+20s — a 10-second step-function penalty on every deployment.

    The operator now injects a startup probe (/health?live, periodSeconds: 2, failureThreshold: 30) when no user-supplied startup probe is present. This gives the router a 60-second startup window at 2-second resolution, eliminating the step-function effect. Users who supply their own startupProbe in spec.deployment.podTemplate.routerContainer are unaffected.

  • Fix SupergraphSchema controller stalling on hung HTTP connections

    Outgoing HTTP to Apollo Studio is now backed by a shared, configurable transport with connect timeouts, TCP keepalive, HTTP/2 keepalive, and per-request OTel instrumentation. Connections that go silently unresponsive — for example when a NAT or network middlebox drops the mapping mid-flight — are now detected and torn down instead of stalling reconciliation indefinitely.

    As a belt-and-braces guard, each SupergraphSchema reconciliation now also has an explicit wall-clock budget. A reconcile that exceeds it is aborted and requeued — the controller can never be held open indefinitely by a stuck call. The bound is tunable:

    YAML
    1controllers:
    2  supergraph_schema:
    3    reconcile_timeout: 10m

    One related diagnostic: if a controller stops running for any reason the operator now logs an ERROR and emits an apollo.operator.controller.stopped counter.

    The http: config block remains available to tune the transport (all fields optional); existing deployments need no config changes.

  • Fix RUSTSEC-2026-0190

  • Fix RUSTSEC-2026-0145, RUSTSEC-2026-0112 and RUSTSEC-2026-0113

  • Fix RUSTSEC-2026-0104

  • Fix router pods entering CrashLoopBackOff with ACCESS_DENIED on first deployment

    A newly-provisioned graph API key could take a few seconds to propagate from Apollo Studio to Uplink. The operator now waits for the Secret to be created before starting the router Deployment, preventing the router from starting before its key is valid.

  • Fix OCI schema delivery for Supergraph CRD and make valid_oci_deployment test fully end-to-end

    Fix the MIME type the operator uses when fetching OCI schema artifacts: accept both application/apollo.schema (current canonical type) and the legacy application/vnd.apollographql.schema so artifacts pushed with either type are retrieved correctly

  • Fix Supergraph schema.oci.graphRef being ignored and router credentials not being mounted

    When schema.oci.graphRef is set, the operator now correctly reads the field (previously silently ignored due to a deserialisation bug) and mounts APOLLO_KEY and APOLLO_UPLINK_ENDPOINTS on the router container, matching the behaviour of schema.studio and schema.resource. This gives a router running a schema from your own OCI registry a GraphOS identity so that insights and other features that require a graphRef will work correctly.

  • Fix spec.routerConfig leaf values being silently overridden by operator defaults

    User-provided leaf values in spec.routerConfig (e.g. supergraph.path) were being overwritten by the operator's base configuration due to incorrect merge order. The merge now starts from base config and applies user config on top, so user values take precedence while base defaults still fill in omitted keys.

1.3.0

Breaking changes

  • Supergraph and SupergraphSet CRDs updated to v1alpha4 to achieve parity with the Router Helm chart — existing manifests must be migrated before upgrading

    spec.podTemplate has moved to spec.deployment.podTemplate, and spec.podTemplate.annotations / spec.podTemplate.labels have been promoted to top-level spec.annotations and spec.labels fields that apply to all managed resources. Resource-specific annotations can now be set independently under spec.deployment.annotations (for the Deployment) and spec.networking.annotations (for the Service). SupergraphSet follows the same structure under spec.supergraphTemplate.

    New fields added in this version include:

    • Service & Deployment annotations

    • NodeSelector

    • ImagePullSecrets

    • RollingUpdateParameters: maxUnavailable and maxSurge

    For other resources not managed directly by the Operator, see the migration guide.

Bug fixes

  • Bump dependencies to resolve CVE-2026-32766

  • Bump dependencies to resolve GHSA-pwjx-qhcg-rvj4, GHSA-9f94-5g5w-gf6r and GHSA-394x-vwmw-crm3

  • Bump dependencies to resolve CVE-2026-31812

  • Supergraphs with Argo Rollout strategy now report a clear failure when Argo is not installed

    If you configure a Supergraph with deploymentStrategy: rollout but the Argo Rollouts API is not installed in your cluster, the operator previously created a plain Kubernetes Deployment silently in its place. The Supergraph appeared to be running, but was not using the strategy you specified.

    The operator now sets Progressing: False with reason DeploymentFailed on the Supergraph, making the misconfiguration immediately visible via kubectl describe supergraph. No Deployment is created until the configuration is corrected or Argo Rollouts is installed.

  • Fix credentials expiring when using OCI registries with short-lived tokens (e.g. ECR, GAR)

  • Bump dependencies to resolve RUSTSEC-2026-0097

  • Fix leftover Rollout resource not being removed after migration to Kubernetes Deployment strategy

    After a successful migration, the Argo Rollout resource was not reliably cleaned up, leaving a stale resource in the cluster. It will now be removed once the Deployment has fully rolled out.

  • Fix SupergraphSchema controller becoming unresponsive after extended inactivity with no compositions

1.2.0

Breaking changes

  • Expose full router container configuration via routerContainer spec

    The Supergraph and SupergraphSet CRDs now expose a routerContainer field under spec.podTemplate for fine-grained control over the router container, including lifecycle hooks, liveness/readiness/startup probes, additional ports, and volume mounts. spec.podTemplate.additionalVolumes is also now available for mounting volumes into containers in the router pods. As a result, the Supergraph and SupergraphSet CRDs have been bumped from v1alpha2 to v1alpha3.

    Several fields have moved from spec.podTemplate into the new spec.podTemplate.routerContainer sub-object. Existing Supergraph and SupergraphSet resources using these fields must be updated:

    Old fieldNew field
    spec.podTemplate.envspec.podTemplate.routerContainer.additionalEnv
    spec.podTemplate.envFromspec.podTemplate.routerContainer.additionalEnvFrom
    spec.podTemplate.resourcesspec.podTemplate.routerContainer.resources
    spec.podTemplate.securityContextspec.podTemplate.routerContainer.securityContext

    See our documentation for more details as to exact fields exposed

Features

  • Add installCRDs and rbac.create flags to Helm chart values

    Set these to false when installing a second instance of the operator in the same cluster. installCRDs: false skips CRD creation so Helm does not conflict with the CRDs owned by the first installation, and rbac.create: false skips ClusterRole creation so the existing ClusterRoles are reused with new namespace-scoped RoleBindings.

Bug fixes

  • Bump dependencies to resolve GHSA-65p9-r9h6-22vj, GHSA-vw5v-4f2q-w9xf, and GHSA-hfpc-8r3f-gw53.

  • Improve Argo Rollouts detection and compatibility with namespaced mode

    Argo Rollouts availability detection and the Rollout resource watcher have been updated to work correctly when the operator is running in namespaced mode alongside a namespace-scoped Argo Rollouts install.

1.1.1

Bug fixes

  • Fix 404 errors when Argo Rollouts is not installed