Serverless

Serverless is the cloud-native compute model where the platform runs your code in response to events and scales (and bills) per-invocation, with no servers for you to provision or keep warm — AWS Lambda is the canonical example. It is the other major production substrate beside container orchestration (Kubernetes): where K8s gives you a cluster you operate, serverless gives you functions the provider operates, moving the operational burden from capacity to efficiency and governance.

Why it sits in this spoke

The spoke owns the practice of operating cloud-native systems in production, and serverless is a first-class operating model with its own SRE problems — they just shift:

• No capacity ops, but real coordination ops. You don’t tune nodes, but you do manage concurrency limits, throttles, scheduling, cost, and multi-account governance at fleet scale (scaling-to-1m-lambda).
• Isolation by account, not namespace. The serverless blast-radius unit is the AWS account (one-per-tenant) rather than a Kubernetes namespace — same seams-not-components concern, different boundary.
• Scheduling hazards. Event/cron-driven functions invite synchronized-load failures (the self-DDoS pattern) that jitter fixes — an SRE concern with no direct container analogue.

The sandbox variant

Sandboxes (Modal, AWS’s Firecracker microVMs) are the model pushed further: the per-invocation unit becomes an isolated environment for untrusted, usually agent-written code, created by the million and discarded in seconds (modal-1m-sandboxes). The serverless bargain holds — the provider operates the runtime, you manage concurrency and cost — but the isolation boundary moves again, from the account down to the single execution, and the operating problem moves with it: from governing a fleet of functions to placing a burst without a coordination bottleneck.

Contrast with Kubernetes

• Kubernetes — you operate the cluster (nodes, scheduler, ~20–30 CNCF tools); maximal control, maximal integration tax.
• Serverless — the provider operates the runtime; you trade control for far less infra ops, and the cost moves to per-invocation efficiency and fleet governance across many accounts.

The deployment/observability/SLO disciplines the spoke documents apply to both; the levers differ.

Related

scaling-to-1m-lambda · kubernetes · platform-ops · site-reliability-engineering · internal-developer-platform · compute-sandbox · container-scheduling · modal-1m-sandboxes