Production Readiness

Production readiness checklist for deploying Spice.ai Enterprise.

This section is the canonical reference for running Spice.ai Enterprise in production. It covers the architectural, operational, and security decisions required to operate Spice at scale with the SLAs expected from an enterprise data and AI platform.

Use this checklist as the definitive sign-off list before promoting a Spice.ai Enterprise deployment to production.

At a Glance

Area	Production Target	Reference
Topology	Multi-replica, multi-AZ; SpicepodCluster for distributed query workloads.	High Availability
Storage	Local NVMe for accelerations; io2 / Premium SSD v2 for shared state.	Storage
Observability	Prometheus scraping, Grafana dashboard, log aggregation, alerts.	Observability
Security	mTLS between nodes, OIDC authentication, NetworkPolicy, image pinning.	Security
Authentication	OIDC bearer tokens or API keys; identity SQL functions.	Authentication
Upgrades	Tiered security updates; rolling, validated upgrades.	Upgrades
Backup / DR	Object-store-backed cluster state; Spicepod source of truth in Git.	Storage \u2192 Disaster recovery
Support	24/7 premium support; 99.9%+ SLA.	Spice.ai Enterprise

Production Readiness Checklist

Topology and high availability

• Run at least two scheduler replicas for HA. See High Availability.

• Run at least three executor replicas to tolerate single-node failure during shuffles.

• Use SpicepodCluster for distributed query workloads; use SpicepodSet with replicas >= 2 for stateless query routers.

• Spread replicas across availability zones with topologySpreadConstraints or pod anti-affinity.

• Configure a PodDisruptionBudget covering scheduler and executor pools.

• Front the runtime with a load balancer that performs L4 health checks against /health (port 8090).

Storage and acceleration

• Local NVMe-backed nodes (AWS i4i / m6id / c7gd / r7gd, Azure Lsv3 / Ddsv5, GCP *-lssd) for accelerations.

• Backing PVC class is io2 Block Express (AWS) or Premium SSD v2 (Azure) when shared / replica-attachable persistence is required.

• S3-compatible object store provisioned for SpicepodCluster shared state; bucket versioning and lifecycle policies are enabled.

• Object store backing runtime.scheduler.state_location supports conditional writes (ETag / If-Match) — required for multi-active correctness.

• Every accelerated dataset and view targeted at SpicepodCluster declares acceleration.partition_by. See Distributed Query → Partitioning.

• Acceleration sizing has been validated against expected dataset growth for the next 12 months. See Storage.

Observability

• Prometheus scrapes the 9090 metrics endpoint via the Spice Helm chart PodMonitor or the operator ServiceMonitor.

• Grafana dashboard is imported and connected to the Prometheus data source.

• Logs are forwarded to a centralized aggregator (CloudWatch, Loki, Datadog, Elastic).

• Alerts are configured for query error rate, refresh failures, executor crashloops, and certificate expiry.

• OpenTelemetry traces are exported when distributed tracing is in use.

Security

• Authentication is enabled (OIDC or API keys); no unauthenticated endpoints are exposed externally.

• mTLS is enabled for all SpicepodCluster deployments (default; allowInsecureConnections must remain unset).

• Container images are pinned to immutable digests (sha256:...), not floating tags.

• NetworkPolicy restricts ingress to load balancer / ingress controller IPs and egress to required upstream endpoints. See Security.

• Pod runs as the default UID 65534 with runAsNonRoot: true and a readOnlyRootFilesystem.

• Secrets are sourced from a secret store (AWS Secrets Manager, Azure Key Vault, Kubernetes Secret) rather than embedded in values.yaml.

• Operator and runtime images are scanned by the organization's image scanner (Trivy, Snyk, Aqua) before promotion.

Capacity and resource limits

• CPU and memory requests and limits are set on every pod.

• Resource sizing has been validated under projected concurrent query load. Start at cpu: 2 / memory: 8Gi and scale based on runtime.task_history and Prometheus metrics.

• Crashloop protection threshold is left at the default 10, or tuned for the deployment.

Lifecycle and upgrades

• Upgrades are validated in a non-production environment before promoting to prod.

• RollingParallel or RollingOrdered update strategy is configured on every SpicepodSet.

• Spicepod manifests are stored in Git and applied via Argo CD, Flux, or the equivalent GitOps controller.

• A documented rollback procedure exists, including the ability to roll back the operator chart and the runtime image independently.

Compliance and support

• An active Spice.ai Enterprise license or AWS Marketplace subscription is attached to the deployment.

• Tenant or workload-level audit logs are forwarded to the organization's SIEM.

• On-call rotation includes the Spice.ai Enterprise premium support contact path.

• SLA targets (99.9%+) are documented in the team's runbook with monitoring tied back to the alert system.

When to use what

Workload pattern	Recommended deployment
Single-tenant edge / sidecar accelerator	SpicepodSet with replicas: 1, local NVMe.
Multi-replica stateless query API	SpicepodSet with replicas >= 2, behind an internal LB.
File-based acceleration with persistence (DuckDB, etc.)	SpicepodSet with volume.storage_requests on io2 / Premium SSD v2.
Distributed query, multi-AZ, multi-tenant	SpicepodCluster with 2+ schedulers and 3+ executors.
Air-gapped / regulated environment	Self-hosted with private registry; pinned digests; mTLS enforced.

Next steps

1. Walk through the checklist above end to end.

2. Read the area-specific guides linked from the At a Glance table.

3. Engage Spice.ai Enterprise support for an architecture review prior to go-live.