Repository navigation
Decide where fleet reachability and GPU-seconds are measured from #511
Description
Activity
Proposing the split rather than leaving it open: state is open source, money is commercial.
modelplane_cluster_connectedandmodelplane_replicas_desiredare open source. They are the state of objects the open source API owns, and an operator running two clusters needs to know one is unreachable whether or not they pay for anything. A fleet that cannot answer "is this cluster still there" from its own API is incomplete on its own terms.modelplane_gpu_seconds_total, andmodelplane_replica_allocated_time_secondsunderneath it, is billing attribution. It answers what a tenant owes, which is the fleet view, and it wants a vantage that survives the cluster going away precisely because that is a billing edge case rather than an operational one. That belongs in Upbound Inference.If that holds, this issue shrinks to two metrics and the smaller first step in the description is enough for them: ship the
ResourceMetricsMonitordefinitions and let an operator scrape them with the Prometheus they already run. No control-plane collector, and no new component on the control plane.
Three metrics in #476's per-object-state group cannot come from a workload cluster, and #510 takes the rest.
modelplane_cluster_connectedis unreachable by construction: a collector running on a cluster you have lost cannot report that you have lost it.modelplane_replicas_desiredis the user's intent, which exists only on theModelDeployment. And the cost chain,modelplane_replica_allocated_time_secondsintomodelplane_gpu_seconds_total, wants a vantage that survives the cluster going away, because that is when the billing question gets asked.design/telemetry.mdanswers this withresource-state-metricsplus a control-plane collector. That is a sound answer to the measurement question and an expensive one to the operational question: it would be the first workload Modelplane installs on a control plane, and it brings a new XR or an install-guide component, with its own readiness, RBAC and upgrade story.TelemetryDestinationis already cluster-scoped on the control plane and its function composes nothing, so it is the natural place to hang this if we do build it.Two things to settle before anything is built.
A smaller first step exists. Shipping the
ResourceMetricsMonitordefinitions, and documenting that an operator scrapes them with the Prometheus they already run, gives the series without Modelplane owning an export path on the control plane.And these three metrics are the fleet view rather than the cluster view, which is where the open source telemetry design and the Upbound Inference one divide. Worth agreeing which side of that line they fall on before either side builds them.