Where your workloads run.
The seven canonical execution surfaces need somewhere to execute, scale, drain, and expose traffic. Runtime Fabric is that substrate: horizontal scale, autoscale, load-balanced ingress, per-surface realization, vertical CPU and memory policy, storage, bandwidth, accelerators, and edge budgets. AWS ECS first-shipped; Cloud Run, Kubernetes, Nomad, Fly.io, edge substrates, and self-hosted inference clusters are peer RuntimeSubstrate connectors.
Shared, per-project, per-surface. All canonical.
Shared pool
Platform-scope binding. Project surfaces delegate to PlatformOperatorKind.AuthoredWorkerPool. One service can host many tenants while isolation stays in scope, grants, and runtime policy.
Per-project pool
Project-scope binding. The connector realizes isolated workload services for a project environment. Use it when noisy-neighbor concerns, compliance, or cost attribution demand isolation.
Per-surface services
API ingress, core handlers, event consumers, scheduled jobs, webhooks, realtime gateway, edge, and management surfaces can realize as separate services or named scale groups.
Managed ingress
RuntimeNetworkExposureSpec declares public, private, or mesh ingress plus TLS, protocols, aliases, and load-balancing intent. Provider handles stay connector-private.
Health and drain
Live probes, ready probes, desired-count drain, rollout gates, circuit breakers, endpoint health, and reconcile actions flow through the same runtime realization record.
Runtime measures
Desired count, running count, CPU, memory, request rate, latency, errors, queue depth, and custom PCG measures feed scaling and explainability through canonical dimensions.
Self-hosted inference
vLLM, Ollama, GPU pools, and on-prem inference clusters land here as operator-managed RuntimeSubstrate workloads with scaling and vertical resource policy. There is no separate reserved kind.
Capability-aware spec
Workloads declare CPU, memory, GPU, network, scaling, partitioning, load balancing, and protocol needs. Connector emits the substrate-native shape. Mismatches fail at compile.
Peer substrates
Cloud Run, Kubernetes, Nomad, Fly.io, edge containers, and local hosts implement the same canonical surface. Same workloads, different substrate bindings.
No separate scaling connector. One workload authority.
Scaling belongs to the same realization compiler that knows the project topology, environment, surface kind, binding, governance envelope, capability grants, endpoints, health probes, rollout state, and PCG measures.
Manual and autoscale modes
Fixed, manual, and autoscale modes share one policy shape. Autoscale strategies declare metrics, thresholds, steps, schedules, cooldowns, and hysteresis.
Vertical resource policy
CPU millicores, memory MiB, ephemeral storage, persistent storage, IOPS, network bandwidth, resource class, accelerators, and edge-isolate budgets are canonical project policy.
Runtime origin routing
Distribution can route CDN behavior to load-balanced runtime origins while Runtime Fabric keeps authority over scale, health, drain, and endpoint realization.
Services or named groups
Capability advertised
Policy first, binding second
One RuntimeSubstrate authority
Run your handlers where the workload fits.
Shared pool for cost. Per-project pool for isolation. Per-surface topology for independent scale. ECS today; Cloud Run, K8s, Nomad, Fly.io, and edge as peer connectors. Same canonical seven execution surfaces, different substrates.