LiteLLM on GKE Autopilot

LiteLLM is an open-source LLM proxy and AI gateway that provides a unified OpenAI-compatible API across 100+ providers including OpenAI, Anthropic, Google Gemini, Azure OpenAI, AWS Bedrock, and Ollama. Organizations use it to centralize AI spend tracking, manage virtual API keys, enforce rate limits, and gain full visibility over model usage. This module deploys LiteLLM on GKE Autopilot on top of the App_GKE foundation, which provisions and manages the shared Google Cloud and Kubernetes infrastructure.

This guide focuses on the cloud services LiteLLM uses and how to explore and operate them from the Google Cloud Console and the command line. For the mechanics that are common to every GKE application — Workload Identity, ingress, autoscaling, CI/CD, Cloud Armor, IAP, Binary Authorization, VPC Service Controls, backups, and the deployment lifecycle — refer to the App_GKE foundation guide rather than repeating them here.

---

1. Overview

LiteLLM runs as a Python-based proxy workload. The deployment wires together a focused set of Google Cloud services:

Capability	Google Cloud service	Notes
Compute	GKE Autopilot	Python proxy pods, 2 vCPU / 2 GiB by default, horizontally autoscaled
Database	Cloud SQL for PostgreSQL 15	Required — LiteLLM's Prisma ORM uses PostgreSQL for virtual keys and spend tracking
Object storage	Cloud Storage	Optional — no buckets created by default
Cache	Redis	Optional — reduces latency and cost for repeated identical LLM requests
Secrets	Secret Manager	Auto-generated master key and salt key; LLM provider API keys injected at runtime
Ingress	Cloud Load Balancing	External LoadBalancer, optional custom domain + managed certificate

Sensible defaults worth knowing up front:

• PostgreSQL 15 is mandatory. LiteLLM's Prisma ORM requires PostgreSQL for virtual key management and spend tracking; changing the engine breaks startup.
• A custom container image is built by Cloud Build. The image embeds an entrypoint.sh that assembles DATABASE_URL from the DB_* environment variables injected by the foundation at runtime.
• LITELLM_MASTER_KEY and LITELLM_SALT_KEY are auto-generated and stored in Secret Manager. The salt key must never be rotated after virtual keys have been issued — all existing virtual keys would become permanently invalid.
• STORE_MODEL_IN_DB = "true" is set automatically, enabling runtime model management and the Admin UI without container restarts.
• Redis is disabled by default. Enable it for multi-replica deployments to share rate-limit counters and response caches across pods.
• The startup probe targets /health/readiness, which validates database connectivity and confirms Prisma migrations have completed before traffic is routed to the pod.

---

2. Google Cloud Services & How to Explore Them

All commands assume you have run gcloud container clusters get-credentials <cluster> --region <region> --project <project> and that PROJECT, REGION, and NAMESPACE are set. The namespace and other identifiers are reported in the deployment Outputs.

A. GKE Autopilot — the LiteLLM workload

LiteLLM pods are scheduled on Autopilot, which bills for the CPU/memory the pods actually request. Horizontal Pod Autoscaling sizes the deployment between the minimum and maximum replica counts.

• Console: Kubernetes Engine → Workloads → select the LiteLLM workload to see pods, revisions, and events. Kubernetes Engine → Services & Ingress shows the external IP.
• CLI:

  kubectl get pods,svc,hpa -n "$NAMESPACE"
  kubectl logs -n "$NAMESPACE" deploy/<service-name> --tail=100
  kubectl describe hpa -n "$NAMESPACE"          # current vs target utilisation

See App_GKE for how Autopilot, scaling, and the workload type (Deployment vs StatefulSet) are managed.

B. Cloud SQL for PostgreSQL 15

LiteLLM stores all virtual keys, usage logs, cost records, and model routing rules in a managed Cloud SQL for PostgreSQL 15 instance. Pods reach it privately through the Cloud SQL Auth Proxy sidecar over a Unix socket, so no public IP is exposed. On first deploy an initialization job creates the application database and user.

• Console: SQL → select the instance for connections, backups, flags, and metrics.
• CLI:

  gcloud sql instances list --project "$PROJECT"
  gcloud sql instances describe <instance-name> --project "$PROJECT"
  # Open an interactive shell to inspect schema/data:
  gcloud sql connect <instance-name> --user=<db-user> --database=litellm_db --project "$PROJECT"

The instance name, database name, user, and the Secret Manager secret holding the password are all surfaced in the Outputs. For the connection model, automated backups, and password rotation, see App_GKE.

C. Cloud Storage

No storage buckets are created by default. Buckets can be declared via the storage_buckets variable and mounted via GCS Fuse when gcs_volumes is set — for example, to deliver a config.yaml to the container without rebuilding the image.

• Console: Cloud Storage → Buckets.
• CLI:

  gcloud storage buckets list --project "$PROJECT"
  gcloud storage ls gs://<bucket-name>/

See App_GKE for GCS Fuse and CMEK options.

D. Redis cache

Redis backs LiteLLM's optional response caching and shared rate-limit counters. When enable_redis = true, the REDIS_HOST, REDIS_PORT, and (optionally) REDIS_PASSWORD environment variables are injected into pods automatically.

• Console: Memorystore → Redis (if using a managed instance).
• CLI:

  redis-cli -h <redis-host> ping        # from a host with network access
  redis-cli -h <redis-host> info keyspace

E. Secret Manager

LITELLM_MASTER_KEY (the primary admin API key, prefixed sk-) and LITELLM_SALT_KEY (used to hash virtual keys) are generated automatically and stored in Secret Manager. LLM provider API keys (e.g. OPENAI_API_KEY) are injected by referencing pre-existing secrets via secret_environment_variables.

• Console: Security → Secret Manager.
• CLI:

  gcloud secrets list --project "$PROJECT"
  # Retrieve the master key:
  gcloud secrets versions access latest --secret=<master-key-secret> --project "$PROJECT"

The database password secret name is in the Outputs. See App_GKE for the Secret Store CSI integration and rotation.

F. Networking & ingress

By default the workload is exposed through an external Cloud Load Balancing IP. A custom domain with a Google-managed certificate can be enabled, and a static IP can be reserved so the address survives redeploys.

• Console: Network services → Load balancing; VPC network → IP addresses.
• CLI:

  kubectl get ingress,svc -n "$NAMESPACE"
  gcloud compute addresses list --project "$PROJECT"

See App_GKE for custom domains, Cloud CDN, and static IP details.

G. Cloud Logging & Monitoring

Pod stdout/stderr flow to Cloud Logging; GKE and Cloud SQL metrics flow to Cloud Monitoring. Optional uptime checks and alert policies are available.

• Console: Logging → Logs Explorer; Monitoring → Dashboards / Alerting.
• CLI:

  gcloud logging read 'resource.type="k8s_container" AND resource.labels.namespace_name="'"$NAMESPACE"'"' \
    --project "$PROJECT" --limit 50

---

3. LiteLLM Application Behaviour

• First-deploy database setup. An initialization job creates the LiteLLM database and user and grants privileges before the application starts. It connects to Cloud SQL through the Auth Proxy and is idempotent.

• Prisma migrations on start. LiteLLM runs its Prisma ORM migrations on each pod start, so upgrading the application version applies schema changes automatically. The startup probe waits until /health/readiness returns 200, confirming migrations are complete before routing traffic to the pod.

• Admin UI. The LiteLLM Admin UI is available at /ui on the service URL. Authenticate with the LITELLM_MASTER_KEY (retrieve it from Secret Manager). From the UI you can add models, create virtual keys, set budgets, and view usage dashboards — all without restarting the container.

• Adding LLM provider keys. Provider API keys are not managed by this module. Supply them at deploy time via secret_environment_variables (mapping each env var to a pre-existing Secret Manager secret), or add them after deployment via the Admin UI or the /model/new API endpoint using the master key.

• Virtual key management. Use the /key/generate API with the master key to issue per-team or per-user virtual keys with rate limits and spend budgets. These keys are stored in PostgreSQL and salted with LITELLM_SALT_KEY.

  # Retrieve the master key then create a virtual key:
  MASTER_KEY=$(gcloud secrets versions access latest --secret=<master-key-secret> --project "$PROJECT")
  curl -X POST "https://<service-url>/key/generate" \
    -H "Authorization: Bearer $MASTER_KEY" \
    -H "Content-Type: application/json" \
    -d '{"key_alias": "team-a", "max_budget": 10.0}'

• Health endpoints. /health/readiness validates database connectivity and Prisma migration completion; /health/liveliness confirms the proxy process is running. These are used as the startup and liveness probes respectively.

---

4. Configuration Variables

Variables are grouped exactly as they appear on the deployment platform. Only settings specific to or notable for LiteLLM are listed; every other input is inherited from App_GKE with its standard behaviour and defaults.

Group 1 — Project & Identity

Variable	Default	Description
project_id	(required)	Target Google Cloud project.
region	us-central1	Region for the workload and regional resources.

Group 2 — Deployment Environment

Variable	Default	Description
tenant_deployment_id	demo	Short suffix that makes resource names unique per environment.
support_users	[]	Emails granted project access and monitoring alerts.
resource_labels	{}	Labels applied to all resources for cost/ownership tracking.

Group 3 — Application Identity

Variable	Default	Description
application_name	litellm	Base name for resources. Do not change after first deploy.
display_name	LiteLLM AI Gateway	Friendly name shown in the Console.
description	(set)	Workload description annotation.
application_version	main-stable	LiteLLM image version tag; pin to a specific release for production stability.

Group 4 — Runtime & Scaling

Variable	Default	Description
deploy_application	true	Set false to provision infrastructure only.
cpu_limit	2000m	CPU per pod; 2 vCPU recommended.
memory_limit	2Gi	Memory per pod; increase to 4Gi for high-throughput deployments.
min_instance_count	1	Minimum replicas. Keep ≥ 1 to avoid cold starts on the API gateway.
max_instance_count	3	Maximum replicas (autoscaler ceiling).
container_port	4000	LiteLLM's native port.
timeout_seconds	600	Request timeout; increase for long-running LLM inference calls.
enable_cloudsql_volume	true	Cloud SQL Auth Proxy sidecar for Unix socket connections.
enable_vertical_pod_autoscaling	false	Let Autopilot tune resource requests automatically (disables HPA).

Group 5 — Environment Variables & Secrets

Variable	Default	Description
environment_variables	{ LITELLM_LOG="INFO", NUM_WORKERS="1" }	Extra non-secret settings. Core LiteLLM vars are set automatically.
secret_environment_variables	{}	Map of env var → Secret Manager secret name. Use to inject LLM provider API keys.

Group 6 — GKE Backend & Cluster

Variable	Default	Description
service_type	LoadBalancer	How the Service is exposed.
session_affinity	ClientIP	Sticky routing; useful when a client re-uses an open connection to the same pod.
workload_type	null	Auto-resolves to StatefulSet when per-pod storage is enabled.
network_tags	["nfsserver"]	Node/pod tags for firewall rules.

Group 7 — StatefulSet

Variable	Default	Description
stateful_pvc_enabled	null	Enable PVC templates; automatically selects StatefulSet workload type.
stateful_pvc_size	10Gi	Storage size for each PVC.
stateful_pvc_mount_path	/data	Container path where the PVC is mounted.
stateful_pvc_storage_class	standard-rwo	Kubernetes StorageClass for PVCs.

Group 8 — Resource Quota

Variable	Default	Description
enable_resource_quota	false	Declared for convention parity; not referenced by this module's deployment (no ResourceQuota is created).
quota_memory_requests / quota_memory_limits	""	Not referenced by this module. If wired via another module, must use binary units (4Gi, 8192Mi) — bare integers are read as bytes and block scheduling.

Group 9 — Reliability Policies

Variable	Default	Description
enable_pod_disruption_budget	true	Protect availability during node upgrades.
pdb_min_available	1	Raise min_instance_count above 1 if you need eviction headroom.
enable_topology_spread	false	Spread pods across zones.

Group 10 — Observability & Health

Variable	Default	Description
startup_probe	/health/readiness	HTTP probe; validates DB connectivity and Prisma migrations before routing traffic.
liveness_probe	/health/liveliness	HTTP probe; confirms the proxy process is running.
uptime_check_config	disabled	Optional Cloud Monitoring uptime check.
alert_policies	[]	Optional metric alert policies.

Group 11 — Jobs & Scheduled Tasks

Variable	Default	Description
initialization_jobs	[]	Leave empty to use the built-in database setup job from LiteLLM_Common.
cron_jobs	[]	Scheduled Kubernetes CronJobs for maintenance or housekeeping tasks.

Group 12 — CI/CD & GitHub Integration

Standard App_GKE Cloud Build / Cloud Deploy integration — see App_GKE. Key inputs: enable_cicd_trigger, github_repository_url, github_token, enable_cloud_deploy.

Group 13 — Filesystem (NFS)

Variable	Default	Description
enable_nfs	false	NFS is not required for LiteLLM; enable only for shared config file delivery.
nfs_mount_path	/mnt/nfs	Mount path inside the container.

Group 14 — Cloud Storage & Artifact Registry

Variable	Default	Description
create_cloud_storage	true	Provision buckets declared in storage_buckets.
storage_buckets	[]	No buckets created by default.
gcs_volumes	[]	GCS Fuse volume mounts for config file delivery.
manage_storage_kms_iam / enable_artifact_registry_cmek	false	CMEK options.

Group 16 — Database Backend

Variable	Default	Description
database_type	POSTGRES	Fixed — do not change; LiteLLM requires PostgreSQL 15.
db_name	litellm_db	Database name. Immutable after first deploy.
db_user	litellm_user	Application user. Immutable after first deploy.
database_password_length	32	Generated password length (16–64).
enable_auto_password_rotation	false	Zero-downtime DB password rotation.

Group 17 — Backup & Maintenance

Variable	Default	Description
backup_schedule	0 2 * * *	Automated backup cron (UTC).
backup_retention_days	7	Retention; raise to 30–90 for production.
enable_backup_import / backup_source / backup_uri	restore options	Restore from a backup on deploy.

Group 18 — Custom SQL Scripts

enable_custom_sql_scripts, custom_sql_scripts_bucket, custom_sql_scripts_path, custom_sql_scripts_use_root — run SQL from a GCS bucket after provisioning. See App_GKE.

Group 19 — Custom Domain, Static IP & Networking

Variable	Default	Description
enable_custom_domain	true	Provision Gateway API routing for custom hostnames + managed certificate.
application_domains	[]	Hostnames to serve.
reserve_static_ip	true	Stable external IP across redeploys.

Group 20 — Identity-Aware Proxy (IAP)

Variable	Default	Description
enable_iap	false	Require Google sign-in in front of LiteLLM.
iap_authorized_users / iap_authorized_groups	[]	Who may access.
iap_oauth_client_id / iap_oauth_client_secret	""	Required when IAP is enabled (sensitive).

Group 21 — Redis Cache

Variable	Default	Description
enable_redis	false	Enable Redis for response caching and shared rate-limit counters.
redis_host	""	Redis endpoint; required when enable_redis = true.
redis_port	6379	Redis port.
redis_auth	""	Optional Redis auth password (sensitive).

Group 21 — Cloud Armor

Variable	Default	Description
enable_cloud_armor	false	Attach a Cloud Armor (WAF) policy to the Ingress backend.
admin_ip_ranges	[]	CIDRs allowed privileged access.
cloud_armor_policy_name	default-waf-policy	Policy name.

Group 22 — VPC Service Controls & Audit Logging

Variable	Default	Description
enable_vpc_sc	false	Enforce a VPC-SC perimeter (requires organization_id).
vpc_cidr_ranges / vpc_sc_dry_run	(set)	Access level CIDRs / dry-run mode.
enable_audit_logging	false	Detailed Cloud Audit Logs.

---

5. Outputs

These values are returned on a successful deployment and are the quickest way to locate and explore the running resources.

Output	Description
service_name	Kubernetes Service name.
namespace	Namespace the workload runs in.
service_cluster_ip	In-cluster ClusterIP.
stage_service_cluster_ips	Map of ClusterIPs for stage-specific services.
service_external_ip	External LoadBalancer IP (when a static IP is reserved).
service_url	URL to reach LiteLLM.
database_instance_name	Cloud SQL instance name.
database_name	Application database name.
database_user	Application database user.
database_password_secret	Secret Manager secret holding the DB password.
database_host / database_port	DB endpoint (127.0.0.1 via the Auth Proxy) / port.
storage_buckets	Created Cloud Storage buckets.
network_name / network_exists / regions	VPC network, presence, available regions.
container_image / container_registry	Deployed image and Artifact Registry repo.
monitoring_enabled / monitoring_notification_channels	Monitoring status and channels.
initialization_jobs / db_import_job	Names of the setup and (optional) import jobs.
deployment_id / tenant_id / resource_prefix	Naming identifiers.
project_id / project_number	Project identifiers.
cicd_enabled / cicd_configuration	CI/CD status and details (repo, trigger, registry).
github_repository_url / github_repository_owner / github_repository_name	Connected repository.
artifact_registry_repository / cloudbuild_trigger_name / cloudbuild_trigger_id	Registry and build trigger.
kubernetes_ready	Whether the cluster/workload is ready.
vpc_sc_enabled / vpc_sc_perimeter_name / vpc_sc_dry_run_mode	VPC-SC status.
audit_logging_enabled / artifact_registry_cmek_enabled	Audit logging and CMEK status.

---

6. Configuration Pitfalls & Sensible Defaults

Risk: Critical (data loss / outage / security) — High (service degraded) — Medium (cost or partial degradation) — Low (minor).

Setting	Sensible value	Risk	Consequence if wrong
database_type	POSTGRES / POSTGRES_15	Critical	LiteLLM requires PostgreSQL; changing the engine breaks the Prisma ORM and prevents startup.
enable_cloudsql_volume	true	Critical	The Auth Proxy sidecar is required for database connectivity; disabling it causes Prisma to fail at startup.
LITELLM_SALT_KEY	auto-generated, never rotated	Critical	Rotating the salt key invalidates every previously issued virtual key; all API consumers lose access immediately.
db_name / db_user	set once	Critical	Immutable after first deploy; renaming recreates the DB/user and destroys all virtual keys and spend data.
enable_backup_import	false unless restoring	Critical	Enabling without a valid backup_uri fails the import job.
quota_memory_requests / _limits	binary units	Critical	Bare integers are bytes and block all scheduling.
ingress_settings / service_type	restrict for production	Critical	A public LoadBalancer exposes the master key endpoint; use ClusterIP with an authenticated Gateway for internal deployments.
iap_oauth_client_id / iap_oauth_client_secret	provide when IAP enabled	Critical	Missing values prevent the IAP gateway from initialising and make the service unreachable.
LITELLM_MASTER_KEY	auto-generated	High	Treat as a credential; rotating it breaks all existing integrations holding the key until they are updated.
enable_redis	true for multi-replica	High	Without Redis, rate-limit counters are per-pod and not shared; quotas are not enforced across replicas.
redis_host	set when Redis enabled	High	An empty host with enable_redis = true causes connection errors on every request.
min_instance_count	1	High	Cold starts add 30–60 s latency and queue all dependent services.
timeout_seconds	600	High	Large language model inference can take minutes; too-short timeout causes 504 errors on slow models.
application_version	pin for production	Medium	LiteLLM releases frequently; unpinned versions may change the Prisma schema or break virtual key formats.
enable_vertical_pod_autoscaling	false unless using VPA	Medium	Enabling VPA disables HPA; choose one or the other.
NUM_WORKERS	1 (raise for throughput)	Medium	A single worker serialises all requests; increase to 2–4 and scale cpu_limit proportionally for high-traffic gateways.
backup_schedule	0 2 * * *	High	Without backups, accidental deletion destroys all virtual keys and usage history with no recovery path.

---

For the foundation behaviour referenced throughout — IAM and Workload Identity, autoscaling, ingress and certificates, CI/CD, Cloud Armor, IAP, Binary Authorization, VPC-SC, backups, and image mirroring — see App_GKE. LiteLLM-specific application configuration shared with the Cloud Run variant is described in LiteLLM_Common.