---
name: runtime-shape
description: Pick a production runtime shape (request-response, streaming, queue, event, cron, durable) and wire observability.
version: 1.0.0
phase: 14
lesson: 29
tags: [production, runtime, queue, event, durable, observability]
---

Given a task class (expected duration, step count, trigger type, latency budget), pick the runtime shape.

Decision:

1. < 30s, user waits -> **request-response**.
2. Progressive UX or voice -> **streaming**.
3. Minutes to hours, user doesn't wait -> **queue-based**.
4. Reactive to external events -> **event-driven**.
5. Periodic housekeeping -> **cron**.
6. Any of the above where restart cost is high -> add **durable execution**.

Produce:

1. The shape scaffold in your stack.
2. Observability: OTel GenAI spans (Lesson 23), backend wired (Lesson 24).
3. For queue: DLQ + retry policy + queue depth metric.
4. For event: explicit subscriber registry + replay path.
5. For cron: lock file or distributed lock to prevent overlapping runs.
6. For durable: checkpointer backend + resume semantics.

Hard rejects:

- Synchronous HTTP for a 5-minute task. Users hang up; workers pile up.
- Queue-based without DLQ. Failed jobs vanish.
- Background work without trace export. Failures invisible until users complain.
- "No durable state, we'll just retry." Long horizons must checkpoint.

Refusal rules:

- If the product has SLA + replay requirements, refuse swarm topology + non-durable runtime.
- If the task is compliance-bound, refuse event-driven without audit trail.
- If the user wants cron + no lock, refuse. Overlapping cron runs are duplicate work at best, data corruption at worst.

Output: runtime scaffold + observability hooks + README with SLA, retry policy, checkpointer choice. End with "what to read next" pointing to Lesson 23 (OTel), Lesson 24 (observability), or Lesson 17 (Managed Agents for hosted long-running).