Chroma on GKE Autopilot — Lab Guide

📖 Configuration Guide

Overview

Estimated time: 45–90 minutes

Chroma is an AI-native open-source vector database purpose-built for embeddings and similarity search. It powers RAG pipelines, semantic search, and LangChain/LlamaIndex workflows. This lab takes you through the full operational lifecycle of the Chroma on GKE Autopilot module on Google Cloud: deploy it, access and verify it, run it day-to-day, observe it, diagnose common problems, and tear it down.

The lab focuses on operating the GKE module and the Google Cloud platform, not on Chroma product features. For the complete list of provisioned services and every configuration input (organised by group), see the Configuration Guide — this lab deliberately does not duplicate that detail so it stays accurate over time.

Objectives

By the end of this lab you will be able to:

• Deploy the module from the RAD platform and locate the resources it provisions.
• Connect to the GKE cluster and access the running workload.
• Perform day-2 operations — inspect, scale, update, and manage secrets and storage.
• Observe the workload with Cloud Logging and Cloud Monitoring.
• Diagnose and resolve the most common deployment and runtime issues.
• Tear the deployment down cleanly.

Prerequisites

• Services_GCP deployed in the target project (provides the VPC, GKE Autopilot cluster, Artifact Registry, and shared service accounts this module depends on).
• A Google Cloud project with billing enabled.
• gcloud CLI and kubectl installed; gcloud auth login and gcloud auth application-default login completed.
• Project Owner (or equivalent) IAM on the project.
• RAD platform access with permission to deploy modules into the project.

Set these shell variables once; every task below reuses them:

export PROJECT="<your-gcp-project-id>"
export REGION="us-central1"           # the region you deploy into

---

Task 1 — Deploy the module [Automated]

1. Click Deploy in the RAD platform top navigation, open Chroma (GKE) from the Platform Modules list to start configuration, set project_id, and review the inputs. Configure only what you need — the Configuration Guide documents every input by group, with defaults. Review the estimated cost (if credits are enabled) and click Deploy, which opens the deployment status page with real-time logs.

2. The platform deploys the workload into the GKE Autopilot cluster, provisions a PersistentVolumeClaim (when stateful_pvc_enabled = true) or a GCS FUSE-backed Cloud Storage bucket as Chroma's persistence backend, builds the container image, and optionally creates a Secret Manager auth token. Chroma requires no database and no initialisation job. First deploys take roughly 10–20 minutes (image build dominates).

3. Connect to the cluster and discover the namespace with name-agnostic filters:

   CLUSTER=$(gcloud container clusters list --project="$PROJECT" --format="value(name)" --limit=1)
   gcloud container clusters get-credentials "$CLUSTER" --region="$REGION" --project="$PROJECT"
   
   NS=$(kubectl get ns -o name | grep chroma | head -1 | cut -d/ -f2)
   echo "Cluster: $CLUSTER   Namespace: $NS"
   kubectl get all -n "$NS"

---

Task 2 — Access & verify [Manual]

1. Confirm the workload is running and locate the service:

   kubectl get pods,svc -n "$NS"

   The service defaults to ClusterIP (internal cluster access only). If service_type = "LoadBalancer" was set, retrieve the external IP:

   EXTERNAL_IP=$(kubectl get svc -n "$NS" \
     -o jsonpath='{.items[?(@.spec.type=="LoadBalancer")].status.loadBalancer.ingress[0].ip}')
   echo "External IP: $EXTERNAL_IP"

2. Verify the heartbeat from within the cluster (or via the LoadBalancer IP if externally exposed). Chroma exposes a single health endpoint on port 8000:

   # From outside the cluster via LoadBalancer
   curl -s "http://${EXTERNAL_IP}:8000/api/v2/heartbeat"   # expect {"nanosecond heartbeat": <timestamp>}
   
   # From inside the cluster via port-forward
   kubectl port-forward svc/"$(kubectl get svc -n "$NS" -o jsonpath='{.items[0].metadata.name}')" \
     8000:8000 -n "$NS" &
   curl -s "http://localhost:8000/api/v2/heartbeat"

3. If enable_auth_token = true was set at deploy time, retrieve the auth token from Secret Manager before making further API calls:

   AUTH_SECRET=$(gcloud secrets list --project="$PROJECT" \
     --filter="name~chroma AND name~auth-token" --format="value(name)" --limit=1)
   gcloud secrets versions access latest --secret="$AUTH_SECRET" --project="$PROJECT"

   Pass the token as Authorization: Bearer <token> on every API request.

---

Task 3 — Operate & keep it running (Day-2) [Manual]

1. Inspect the workload — pods, HPA, and (when PVC-backed) the persistent volumes:

   kubectl get deploy,statefulset,pods,hpa,pvc -n "$NS"
   kubectl describe deploy -n "$NS" 2>/dev/null || kubectl describe statefulset -n "$NS"

2. Scale by changing the min/max instance inputs and clicking Update on the deployment details page — the module owns the workload spec, so scaling is a configuration change, not a manual kubectl scale (a manual edit would be reverted on the next apply). Keep max_instance_count = 1: multiple Chroma pods sharing the same PVC or GCS path have no distributed write lock and will corrupt collections.

3. Update the application version by changing the version input via Update on the deployment details page; a new image builds and a rolling update replaces the pods.

4. Manage secrets and storage:

   kubectl get secrets -n "$NS"
   gcloud secrets list --project="$PROJECT" --filter="name~chroma"
   kubectl get pvc -n "$NS"          # PVC status when stateful_pvc_enabled = true

5. List any scheduled jobs:

   kubectl get jobs,cronjobs -n "$NS"

---

Task 4 — Observe: Logging & Monitoring [Manual]

1. Logs — from kubectl or the Logs Explorer:

   kubectl logs -n "$NS" "$(kubectl get pod -n "$NS" -o jsonpath='{.items[0].metadata.name}')" --tail=50

   Logs Explorer filter: resource.type="k8s_container" AND resource.labels.namespace_name="<namespace>".

2. Monitoring — open the GKE / Kubernetes dashboards and review pod CPU and memory utilisation, restart counts, and HPA behaviour. The module also provisions an uptime check (when enabled) targeting /api/v2/heartbeat; review Monitoring → Uptime checks and Alerting → Policies.

---

Task 5 — Troubleshoot & debug [Manual]

Durable techniques for the failure modes you are most likely to hit. These are platform-level diagnostics and do not change with Chroma releases.

• Pod not Ready / CrashLoopBackOff: inspect events and logs. Chroma loads HNSW indexes from its storage backend on start — allow time for the /api/v2/heartbeat probe to pass.

  kubectl describe pod -n "$NS" <pod>          # Events section shows probe/mount errors
  kubectl logs -n "$NS" <pod> --previous       # logs from the crashed container

• PVC not Bound: confirm stateful_pvc_enabled = true is set and the storage class exists in the cluster (kubectl get storageclass).
• GCS FUSE mount errors (when PVC is not used): verify the GCS bucket exists and the workload's service account has storage.objectAdmin on the bucket.
• Auth token errors (401): confirm enable_auth_token = true was set, the secret exists, and the Authorization: Bearer <token> header is included in every request.
• Pending pod / no external IP: check kubectl describe pod events for resource or quota issues, and confirm the LoadBalancer Service has an assigned IP.
• Image pull errors: confirm the image exists in Artifact Registry and the node service account can pull it.

See the Configuration Guide's Configuration Pitfalls section for setting-specific gotchas.

---

Task 6 — Tear down [Automated]

On the Deployments page, open the deployment and click the Trash icon (Delete). Delete runs terraform destroy and is irreversible (the deployment record is retained for history). If a deployment is stuck and the RAD platform can no longer manage it (for example after manual changes that conflict with the Terraform state), use Purge instead — it removes the deployment from RAD's records without destroying the cloud resources (it makes RAD forget the project). This removes everything the module created — the Kubernetes workload and namespace, PVC and all stored collections, GCS data bucket, Secret Manager secrets, and Artifact Registry images. Resources owned by Services_GCP (the VPC, GKE cluster, shared registry) are managed separately and are not removed here.

---

Summary

Task	Type	Outcome
1 — Deploy	Automated	Module deploys the GKE workload, PVC or GCS data bucket, and optional auth token
2 — Access & verify	Manual	Connect to the cluster; heartbeat check passes; auth token retrieved if enabled
3 — Operate	Manual	Inspect workload, scale, update version, manage secrets and storage
4 — Observe	Manual	Query Cloud Logging; review Cloud Monitoring metrics and uptime check
5 — Troubleshoot	Manual	Diagnose pod, PVC/GCS mount, auth token, scheduling, and image-pull issues
6 — Tear down	Automated	Delete (Trash) removes all module resources