Skip to main content

Install on Kubernetes

Deploy with Helm and your platform services.

Upgrade an instance

Plan, deploy, verify and recover.

Deployment components

Choose services and prepare your image allowlist.

Configure Formbricks

Connect identity, storage and integrations.

Choose your setup

You run and recover a self-hosted installation yourself. Enterprise features need a license; support commitments are set in your contract.

System requirements

Size the host for the whole Formbricks stack: the app, Hub API and worker, Cube, PostgreSQL, Redis/Valkey and SpiceDB. Kubernetes also uses a SpiceDB operator for chart-managed authorization. See deployment components for the image inventory and managed-service options. Chart defaults request 1.9 vCPU and 2.9 GiB across those services — see charts/formbricks/values.yaml, plus charts/spicedb-operator/values.yaml for the operator. On Kubernetes, add whatever your provider reserves per node on top, typically around 0.5 vCPU and 1 GB. The single-server rows are derived from those totals rather than load-tested.
HTTPS Required: Ensure your server runs with a valid HTTPS certificate.
4 GB is a floor, not a comfortable target. It leaves no headroom for a large response import, a survey bundle rebuild, or any of the optional AI services.

Single server or Kubernetes

The single-server rows cover the Docker and one-click installs. Both run every service on one host, so a reboot or an upgrade means downtime — usually an acceptable trade at low load, and not on its own a reason to reach for Kubernetes. Choose Kubernetes when you need to survive a node failure or carry an uptime SLA. For Kubernetes, start at 3 nodes of at least 4 vCPU and 8 GB, autoscaling to 6, with PostgreSQL, Redis and object storage as managed services outside the cluster. Capacity is deliberately provider-neutral: a 2 vCPU node cannot hold the app’s 1 vCPU request alongside the platform reserve, some providers exclude burstable instance families from system node pools, and a free control-plane tier carries no uptime SLA. Most chart workloads start with one replica; SpiceDB starts with two. Configure availability for the whole stack. Object storage must speak the S3 API, so a service without one, such as Azure Blob Storage, needs an S3-compatible gateway in front of it. The Kubernetes deployment guide has the replica, autoscaling and managed-service settings.

AI services are much heavier

Self-hosted AI is off by default and is a different order of magnitude: Hub embeddings alone request 4 vCPU and 8 Gi for the foreground runtime, and 8 vCPU for the worker pool. Prefer an external endpoint unless you need them constantly.
AI features are an Enterprise feature. Configure your Enterprise license before enabling them.

Sizing for a specific load

A burst of responses is a queue question rather than a request-rate one. Each response enqueues several response-pipeline.process jobs — responseCreated, one responseUpdated per progress save, then responseFinished — so the count scales with how often answers are saved, and consumers default to one job in flight per app replica. Size Redis for the deepest backlog you expect: the chart pins maxmemory-policy noeviction, so a full Redis refuses writes rather than dropping jobs. See job runner. A large contact list is storage rather than throughput. Contact attributes are one row per attribute per contact, so 50,000 contacts with 10 attributes each is 500,000 rows before a single response is stored. We publish no responses-per-second figure per replica: it is not load-tested, and a 40-question survey with file uploads is not the same write as a one-question NPS. Watch pod CPU and queue depth separately through your first campaign — they move independently.