| name | architecture-serverless |
| description | Serverless architecture workflow for event-driven and bursty workloads using managed compute and platform services. Use when elasticity and reduced platform operations justify managed-service constraints; do not use when workload shape requires long-lived stateful control loops. |
Architecture Serverless
Overview
Use this skill to design serverless systems that are cost-aware, failure-tolerant, and operationally predictable.
Scope Boundaries
- Traffic is variable or bursty and elastic scaling is valuable.
- Team capacity for infrastructure operations is limited.
- Workloads are naturally event-driven or request/response with bounded execution.
Core Judgments
- Compute model: function granularity, execution time bounds, and concurrency profile.
- State model: where session/process state lives outside function runtime.
- Integration model: event source guarantees, retry semantics, and idempotency.
- Cost model: invocation frequency, cold-start impact, data transfer/storage economics.
Practitioner Heuristics
- Keep functions narrow by business action, not by tiny technical helpers.
- Externalize state deliberately (DB/cache/object store) and design for retries.
- Control concurrency at event source and downstream dependency limits.
- Treat timeouts as business behavior decisions, not only platform defaults.
Workflow
- Profile workload shape and latency/cost sensitivity.
- Design event and request paths with explicit idempotency rules.
- Choose managed services for state, orchestration, and messaging.
- Define concurrency limits and backpressure behavior.
- Model cost and failure behavior under peak and degraded conditions.
- Record escape hatches for workloads that outgrow serverless constraints.
Common Failure Modes
- Hidden coupling through shared environment variables and broad IAM policies.
- Unbounded retries causing duplicate side effects.
- Cost surprises from chatty event chains and storage/egress growth.
Failure Conditions
- Stop when critical workflows cannot tolerate at-least-once delivery semantics.
- Stop when latency SLOs are incompatible with runtime startup behavior.
- Escalate when concurrency or cost risk cannot be bounded.
