Tier: Stream

What this is

The fastest tier in the cadence ladder. Cadence: milliseconds to seconds. Payload: bytes to KB. Used for in-flight transactional events, real-time alerts, SSE feeds, hardware-event subscriptions, and webhook receivers from external POSes.

Purpose

When a downstream consumer needs to react inside the user's perception window — cashier sees alert in under a second, ops dashboard updates in real time, AI agent makes a decision before the cart finalizes — the cadence is stream. Anything slower fails the use case.

Structure

Property	Value
Cadence	ms–sec
Payload size	bytes–KB
Protocol fit	SSE, WebSocket, push webhook, MSMQ-style queue
Failure metric	"no heartbeat in N seconds"
Health input	last-heartbeat timestamp
Recovery primitive	replay from queue (idempotency key required)
Alert pattern	heartbeat-lost
Cache strategy	none — real-time
Required SLA fields	`freshness_budget` (seconds)
Forbidden SLA fields	`schedule`, `window_*`, `ttl`
Required recovery	`mode: replay-from-queue`, `idempotency_key`
Watcher check interval	1 second

Health states

green:  now − last_heartbeat < freshness_budget
amber:  freshness_budget ≤ now − last_heartbeat < 2× freshness_budget
red:    now − last_heartbeat ≥ 2× freshness_budget

Examples

POST /webhooks/square — Square webhook receiver (Adapter axis)
GET /alerts/sse — live alert feed (Resource axis)
GET /inventory/sse — real-time stock-position changes (Resource axis)
MCP tool synchronous calls like compliance.create_block (Agent axis)
Hardware events from in-store devices (cash drawer open, scanner errors)

Consumers

Cashier UI (alerts must surface within human perception)
In-store AI assistants (agents that gate transactions)
Ops dashboards (live device-health grid)
Anomaly detection (Chirp consumes streams of detection-ready events)

Anti-pattern

Don't use stream tier for slow-moving data — burns cache, generates noise, no benefit. Reference tier with short TTL is almost always better for "current value" lookups.