Repository navigation
Conversation
| - it: connects an enabled RAG pod to the same-release Server service | ||
| template: rag-deployment.yaml | ||
| set: | ||
| rag.enabled: true |
There was a problem hiding this comment.
Please load the RAG overlay for enabled test fixtures.
The enabled cases set only rag.enabled, but the root chart values do not supply the overlay's RAG defaults. Validation requires an image repository, so these cases cannot reach their success assertions. The provider Secret example likewise omits the overlay and fails validation. Load values-rag.yaml or supply complete fixtures, and update the example to include it.
This is a static finding; Helm was not run and this is not a proven cause of the current CI failures.
| {{- end }} | ||
| ports: | ||
| - name: http | ||
| containerPort: {{ include "hugegraph.rag.port" . }} |
There was a problem hiding this comment.
Please align the configured port with the application listener.
rag.port changes this declared container port, the Service target and the named probe port, but this Deployment does not change the application arguments. The official hugegraph-ai source Dockerfile at 1cfad720 starts the application with --port 8001. A non-default rag.port therefore mismatches an image built from that source. Pass the listener argument through to the application, or separate the Service port from a fixed listener port.
This is a static finding for that inspected source version; the mutable latest image and Kubernetes behavior were not inspected or run, and this is not a proven cause of current CI failures.
9816d5f to
c00c00f
Compare
Keep generated runtime configuration ephemeral so rotated Kubernetes Secrets remain authoritative after restarts. Retain prompts and indexes on the PVC, enforce the application's single-active replica contract, reject unsafe security overrides and unknown RAG fields, and strengthen Helm regressions and documentation.
Keep the editable HugeGraph AI .env, prompts, indexes and resources on the retained PVC while reconciling current Kubernetes env and envFrom keys before startup so rotated Secrets remain authoritative. Inherit the global persistence switch, support platform-assigned non-root IDs, avoid archive-mode writes to CSI roots, exclude documentation images from the packaged chart, and document the minimal same-release configuration and external editing workflow.
Fix the retained PVC claim helper name, remove an extra readiness get invocation, and keep the non-persistent prompt path aligned with the HugeGraph AI image layout.
Reject extra RAG volumes and mounts that collide with the chart-managed persistent data volume, add regression coverage, and document the special GRAPH_PWD envFrom rule plus explicit GitOps rollout triggers for rotated external credentials.
Reject rag.env names containing '=' or non-printable bytes during Helm rendering, matching the Kubernetes relaxed EnvVar name contract, and add regression coverage for invalid names.
bitflicker64
left a comment
bitflicker64
left a comment
There was a problem hiding this comment.
Blocking: yes. Summary: the RAG Pod keeps Kubernetes service links, so it crash-loops at import when a qdrant or milvus Service shares its namespace; the docs also name the RAG Deployment and PVC wrongly for the documented release name.
| {{- $serviceAccount := get $rag "serviceAccount" | default dict }} | ||
| {{- $serviceAccountCreate := true }} | ||
| {{- if hasKey $serviceAccount "create" }}{{- $serviceAccountCreate = get $serviceAccount "create" }}{{- end }} | ||
| automountServiceAccountToken: {{ get $serviceAccount "automountServiceAccountToken" | default false }} |
There was a problem hiding this comment.
Important: the RAG Pod keeps Kubernetes service links, so a qdrant or milvus Service in the same namespace injects QDRANT_PORT=tcp://<ip>:6333 or MILVUS_PORT=tcp://<ip>:19530, and IndexConfig raises ValueError at import on int(os.environ.get("QDRANT_PORT", "6333")). Set enableServiceLinks: false in the RAG Pod spec.
|
|
||
| ```bash | ||
| RELEASE=hugegraph | ||
| kubectl -n hugegraph rollout restart deployment/${RELEASE}-hugegraph-rag |
There was a problem hiding this comment.
Minor: with the documented release name hugegraph, hugegraph.fullname resolves to hugegraph, so the objects are hugegraph-rag and hugegraph-rag-data, and this command and those at lines 162, 269 and 291 target hugegraph-hugegraph-rag, which does not exist. Use the real names, or select the Deployment by app.kubernetes.io/instance=${RELEASE},app.kubernetes.io/component=rag.
Move configuration precedence and resource ownership to the RAG application, preserve full Server URLs, support external mounts, and fix RAG regression discovery across Helm versions. Requires the updated HugeGraph AI latest/1.8 image from AI #96. Helm 1+1+1 runtime acceptance, credential rotation and persistence checks were completed with a local source-overlay image.
|
The earlier review comments below still apply at 7b69834; this head does not fix them.
|
Before → After
Before: HugeGraph RAG uses a separate Kubernetes chart, creating configuration and credential drift from the main HugeGraph deployment.
After: add RAG as an opt-in component of the canonical Helm chart, with release-scoped Server and Secret wiring, single-node and cluster presets, and optional networking and persistence controls.
Validation boundary
The source PR reports local source-contract and configuration checks; rendered tests were skipped locally because Helm was unavailable. Live deployment acceptance has not been established.
Source: hugegraph#365