tai chart deploys the same image on Kubernetes. Each release publishes it
as an OCI artifact next to the image, cosign-signed, so you install straight from
the registry.
Install
serve and backend as separate Deployments, each with a
tai metrics native sidecar, and by default installs quickstart Postgres and
Redis StatefulSets for a self-contained trial. A schemaInit hook runs
tai db apply around install and upgrade.
Pin
--version to a chart release. The chart version and the app image version
are cut from the same tag, so chart X.Y.Z targets image X.Y.Z by default
(override with image.tag).Point at your own datastores
For production, disable the quickstart StatefulSets and point every dependency at managed Postgres and Redis:serve.replicaCount scales the API; running serve.workers above one also needs
serve.statelessHttp=true, because the stateful HTTP transport refuses a
multi-worker boot. The chart requires Kubernetes 1.29+ (the metrics native
sidecar), and integrates with — but never installs — an Ingress controller,
cert-manager, the prometheus-operator, and a StorageClass.
Marketplace installs on Kubernetes need the plugin-prefix directory on a shared
persistent volume so every
serve and backend pod sees the same installed
code, plus the manifest on persistent storage. See
Run on Kubernetes for the persistence model and
Plugin persistence
in the chart README.Upgrade
tai db apply hook and rolls the Deployments. With an
external Postgres the hook runs pre-upgrade (schema first, strict ordering);
with the quickstart Postgres it runs post-upgrade. The DDL is additive, and the
data lives in your datastores, not the pods.
The values reference
Every value — probes, autoscaling, ingress, the Postgres-backed feature toggles, secret handling, and the ServiceMonitor — is documented in the chart’s owncharts/tai/README.md.
Run on Kubernetes covers the platform-side setup.
See also
- Verify artifacts — verify the chart signature.
- Deploy — the serving modes the chart wires.