On Demand Cluster

Prerequisites

1. CLI installed and logged in (see setup skill).
2. A Vector Lake instance in the target project/region (see cluster skill -- cluster create-vectorlake).

Commands Reference

Create an On-Demand Cluster

create is hand-written and not generated from the JSON model. It calls POST /v2/clusters/createOnDemandCluster and accepts:

zilliz on-demand-cluster create \
  --project-id <project-id> \
  --region-id <region-id> \
  --cu-size <integer>            # required, >= 8
  --cluster-name <string> \      # required, max 64 chars, letters/digits/space/_/-/CJK
  [--session-ttl <duration>] \   # auto-suspend TTL: 30m, 1h, 90s (min 60s, default 60s)
  [--max-query-node-cu <n>] \
  [--max-query-node-replicas <n>]

Session TTL controls how long an idle on-demand cluster stays running before it is automatically suspended. Format is <number><s|m|h>, with a floor of 60 seconds. Resuming a suspended on-demand cluster happens transparently on the next query.

Examples:

# Create an 8-CU on-demand cluster with the default 60s idle TTL
zilliz on-demand-cluster create \
  --project-id proj-xxxx \
  --region-id aws-us-east-1 \
  --cu-size 8 \
  --cluster-name "qc-prod-1"

# Keep alive 30 minutes between queries, cap query node CU and replicas
zilliz on-demand-cluster create \
  --project-id proj-xxxx \
  --region-id aws-us-east-1 \
  --cu-size 16 \
  --cluster-name "qc-batch-1" \
  --session-ttl 30m \
  --max-query-node-cu 16 \
  --max-query-node-replicas 4

List On Demand Clusters

zilliz on-demand-cluster list --project-id <project-id> --region-id <cloud-region>

Describe an On Demand Cluster

zilliz on-demand-cluster describe --cluster-id <on-demand-cluster-id>

Delete an On Demand Cluster

zilliz on-demand-cluster delete --cluster-id <on-demand-cluster-id-to-delete>

Guidance

• On-demand clusters belong to a Vector Lake instance and only exist inside a <projectId> + <regionId> pair. If neither list nor describe returns anything, first confirm that the project has a Vector Lake (zilliz cluster create-vectorlake).
• list requires both --project-id and --region-id -- there is no all-projects listing.
• Status flows through CREATING -> RUNNING -> SUSPENDING -> SUSPENDED and back to RESUMING -> RUNNING on the next query. Treat SUSPENDING/SUSPENDED as healthy idle state, not an error.
• --session-ttl is the idle auto-suspend timer. The minimum is 60s; values smaller than that are rejected client-side before the API call. The current default is 60s, so on-demand clusters are aggressive about suspending -- raise it (e.g. 30m) when running interactive workloads.
• --cu-size must be >= 8 and --cluster-name is at most 64 characters, allowing letters, digits, space, _, -, and Chinese characters.
• delete is irreversible and asynchronous -- always confirm with the user before invoking. The cluster row moves to DELETING and disappears once cleanup finishes.
• These commands are distinct from the regular cluster skill: a regular cluster owns its own compute, while on-demand clusters share storage with their parent Vector Lake. Do not confuse zilliz on-demand-cluster delete with zilliz cluster delete.