Install the chart
Every release publishes the chart tooci://ghcr.io/galaxy-io/charts/alchemy, with server and host images of the same version under ghcr.io/galaxy-io/alchemy. Without --version, Helm installs the newest release.
Alchemy needs three values:
Without
ENCRYPTION_KEY the server still starts, with a warning, and refuses any function that uses secrets. The key must decode to 16, 24 or 32 bytes.
The recommended setup keeps them in a Secret you manage, named by existingSecret:
existingSecret set, the chart creates no Secret of its own. The server reads every key in it. Hosts read only NATS_URL. When authentication is on, put its credentials in the same Secret.
Without existingSecret, the chart creates the Secret itself from the chart values above, named alchemy-runtime for a release named alchemy.
To keep tenant secrets in AWS Secrets Manager or GCP Secret Manager instead, set secrets.type. The server then uses the SDK’s default credentials, such as its IRSA role or workload identity, and needs no ENCRYPTION_KEY. See Configuration.
Test clusters
For a local or test cluster, the chart vendors the Bitnami PostgreSQL and NATS charts:max_payload: 64MB, which covers the 8 MiB messages Alchemy sends, with JetStream file storage on a 10Gi volume for the queue and records and 1Gi of memory storage for function logs. The password must match the one in the DSN.
What the chart creates
The server runs as a non-root user with a read-only root filesystem and no capabilities. It needs no KVM and can run on any node. Hosts run privileged, with the node’s
/dev/kvm and an emptyDir for the kernel, the root filesystem cache and microVM jails. host.data.sizeLimit caps that volume.
Server
Every server applies database migrations before it listens, one at a time under a PostgreSQL lock. There is no separate migration Job. The startup probe allows two minutes for a slow first install. The server answers/healthz with 204 once it listens. The startup, readiness and liveness probes all use it.
To expose the API and UI, enable the Ingress and set its hosts, class and TLS:
server.invoke.deadline (5 minutes by default), and the UI follows live logs over open streams. Set the load balancer’s idle timeout above that deadline.
Hosts
Hosts are a Deployment, not a DaemonSet, so pending hosts are what make a node autoscaler add KVM nodes. Anti-affinity keeps one host per node, and each host is named after its node. Rolling updates replacehost.maxUnavailable hosts at a time with no surge, since a surged host would need a node of its own. On shutdown a host stops taking work, releases its warm instances and gives in-flight invocations up to 40 seconds to finish before it cancels them. host.terminationGracePeriodSeconds (60) covers that.
Size host.resources for the microVMs, not the host process. See KVM nodes.
Cloud identity
The server and hosts reach the buckets, and hosts reach ECR, Artifact Registry and GCR, through their ServiceAccounts’ cloud identity. Annotate each one:iam.gke.io/gcp-service-account instead. See Storage for which identity needs which access.
Upgrade
Upgrade in place with the same command and a new--version:
existingSecret is not checksummed. After you change it, restart both Deployments:
Verify a release
The release workflows attach build provenance attestations to the chart and the images. Verify the chart before you install it:Values
Every value is documented incharts/alchemy/values.yaml. The ones most deployments set:
Configuration maps every value to the environment variable it sets.